Advances in Applied Energy: Ning Zhao, Fengqi You

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Advances in Applied Energy 2 (2021) 100019

Contents lists available at ScienceDirect

Advances in Applied Energy


journal homepage: www.elsevier.com/locate/adapen

New York State’s 100% renewable electricity transition planning under


uncertainty using a data-driven multistage adaptive robust optimization
approach with machine-learning
Ning Zhao a, Fengqi You a,b,c,∗
a
Systems Engineering, Cornell University, Ithaca, NY 14853, USA
b
Robert Frederick Smith School of Chemical and Biomolecular Engineering, Cornell University, Ithaca, NY 14853, USA
c
Atkinson Center for a Sustainable Future, Cornell University, Ithaca, NY 14853, USA

a r t i c l e i n f o a b s t r a c t

Keywords: Power system decarbonization is critical for combating climate change, and handling systems uncertainties is
Decarbonization essential for designing robust renewable transition pathways. In this study, a bottom-up data-driven multistage
Renewable electricity transition adaptive robust optimization (MARO) framework is proposed to address the power systems’ renewable transition
Optimization under uncertainty
under uncertainty. To illustrate the applicability of the proposed framework, a case study for New York State is
Multistage adaptive robust optimization
presented. Machine learning techniques, including a variational algorithm for Dirichlet process mixture model,
Big data
principal component analysis, and kernel density estimation, are applied for constructing data-driven uncertainty
sets, which are integrated into the proposed MARO framework to systematically handle uncertainty. The results
show that the total renewable electricity transition costs under uncertainty are 21%-42% higher than deter-
ministic planning, and the costs under the data-driven uncertainty sets are 2%-17% lower than the conventional
uncertainty sets. By 2035, on-land wind and offshore wind would be the major power source for the deterministic
planning case and robust optimization cases, respectively.

1. Introduction gest that the future highly decarbonized electricity systems would have
a high penetration of variable renewable energy [15], such as solar,
The Paris Agreement aims to limit the increase in global average tem- wind, and wave [16]. To manage the intermittency and curtailment is-
perature to well below 2 °C above the preindustrial levels and to pursue sues due to large proportions of variable renewable energy [17], there
efforts to limit it to 1.5 °C [1]. To achieve this target requires global are multiple approaches, including electricity storage [18], distributed
cooperation [2], and the systems-level electricity transition from cur- energy systems [19], demand-side response [20], increasing technolog-
rent carbon-intensive and low-efficiency electric power systems [3] to ical diversity [21], and increasing spatial diversity [22]. Considering
future highly decarbonized and energy-efficient electricity sector [4] is that uncertainties may have great impacts on power systems’ renewable
a priority topic for counties around the world [5]. The United States transition [23], it is important to develop an optimization framework
is moving ambitiously on carbon neutral electricity, and the country that handles the potential power systems uncertainties as well as the
is planning to achieve a carbon pollution-free electric power sector by sequences of their realizations. Such an optimization framework would
2035, according to the transition plan of president Joseph R. Biden Jr. benefit investigating the feasibility of climate targets and providing ro-
[6]. In addition, the New York State passed the Climate Leadership and bust and economically efficient electricity transition pathways under
Community Protection Act (CLCPA) in 2019, which sets climate goals of systems uncertainties.
achieving net-zero greenhouse gas emissions for the electricity sector, To the best of our knowledge, there is no existing literature on
as well as energy goals on minimum capacity requirements for multi- the power systems’ renewable transition planning that handles the se-
ple renewable generation technologies and electricity storage [7]. The quences of uncertainty realizations using robust optimization frame-
decarbonization transition of the electric power sector [8] has been in- works. Existing electricity transition planning literature that apply ro-
vestigated from technology [9], economy [10], strategy [11], planning bust optimization to account for systems uncertainties include the de-
[12], sociology [13], and deterministic optimization [14] perspectives ployment of certain types of renewable generation technologies [24],
in the literature. Most decarbonization studies on power systems sug- static robust optimization for strategic energy planning [25], and static
robust optimization on overcapacity issues [26]. Note that the static


Corresponding author at: Systems Engineering, Cornell University, Ithaca, NY 14853, USA.
E-mail address: fengqi.you@cornell.edu (F. You).

https://doi.org/10.1016/j.adapen.2021.100019
Received 21 January 2021; Received in revised form 11 March 2021; Accepted 13 March 2021
Available online 17 March 2021
2666-7924/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Nomenclature CEEi,t,y the annual cost of electricity generation technology i


with age of t in year y for unit generation capacity
Sets/indices CIMPs,y cost associated with unit amount of imported electricity
I set of all technologies, indexed by i, including electric- from source s in year y
ity generation technologies (Ie ) and electricity storage EEi,t,y annual greenhouse gas emissions regarding unit amount
technologies (Is ) of electricity generated from technology i in year y with
Ie set of electricity generation technologies, a subset of I, an age of t
including bituminous coal (BIT), fuel oil (FO), methane ELMTy annual emissions limit for in year y
from biogas (MTE), refuse of solid waste (REF), util- Fi,y annual electricity generation amounts from unit gener-
ity solar PV (PVU), distributed solar PV (PVD), nu- ation capacity using technology i in year y
clear power (UR), hydropower (WAT), on-land wind IMPMs,y maximum electricity importation from source s in year
(WON), offshore wind (WOFF), natural gas combined y
cycle (NGCC), natural gas combustion turbine (NGGT), INIEi,t electricity generation from technology i with an age of
natural gas steam turbine (NGST), natural gas combined t in year YINI
cycle with carbon capture and storage (NCCS) INIBi,t storage capacity of technology i with an age of t in year
Iei set of intermittent renewable electricity generation tech- YINI
nologies, a subset of Ie , including utility solar PV (PVU), LIFEi lifetime for technology i
distributed solar PV (PVD), on-land wind (WON), and MABi,y market availability of storage technology i in year y
offshore wind (WOFF) MAGi,y market availability of generation technology i in year y
Ier set of renewable electricity generation technologies, a PROBi,y proposed additions of electricity storage capacity using
subset of Ie , including methane from biogas (MTE), technology i in year y
hydropower (WAT), PV (SUNU), distributed solar PV PRODi,t,y proposed decommission of electricity generation corre-
(SUND), on-land wind (WON), and offshore wind sponding to technology i in year y with an age of t
(WOFF) PROGi,y proposed additions of power generation capacity with
Is set of electricity storage technologies, a subset of I, in- technology i in year y
cluding lithium-ion battery (LIB) and flywheel (FW) RATE discount rate
K set of uncertain parameters REQMi percentage of marginal addition of capacity for re-
L set of hyperplanes that formulate the uncertainty sets newable generation technology i that marks the lower
N set of time periods, indexed by n, from the initial year bound of electricity storage deployment capacity
(YINI) to the end of planning horizon (YM) RPSGy renewable portfolio standards on renewable generation
S source of imported electricity, indexed by s for the electric power sector in year y, determined by
T set of ages for all technologies, indexed by t, and t = 0 the government
indicates that the technology has just been deployed in TARi,y target capacity level for technology i in year y
this year TARSy target capacity level for electricity storage in year y
Y set of time periods, indexed by y, from the initial year YINI initial year for the planning horizon
(YINI) to the end of planning horizon (YM) YMAX planning time horizon in years

Uncertain parameters
uk=1 ,n uncertain electricity generation demands in time period
robust optimization approach generally adopts the assumption of pre-
n
defined uncertainty sets and does not consider the sequences of uncer-
uk=2 ,n uncertain market availability of wind turbines in time
tainty realizations [27], so it could not fully utilize the uncertainty data
period n
[28] and tends to generate over conservative solutions compared to mul-
uk=3 ,n uncertain market availability of solar PVs in time period
tistage adaptive robust optimization (MARO) [29], which is an advan-
n
tage of MARO over the conventional robust optimization approaches.
Non-Negative Continuous variables Notably, there are studies on renewable electricity transition planning
bdisi,t,y electricity storage capacity of technology i discarded in [30] that integrate Monte Carlo sampling techniques and deterministic
year y with an age of t optimization to account for the uncertainties of power systems [31]. Al-
bexti,t,y electricity storage capacity of technology i existing in though such an approach considers various potential combinations of
year y with an age of t, and t = 0 indicates new units uncertainty data, it is worth noting that applying uncertain parameters
deployed in year y obtained from Monte Carlo sampling into deterministic optimization
edisi,t,y electricity generation capacity of technology i decom- does not guarantee a globally optimal solution, because of the stochas-
missioned in year y with an age of t tic nature of the approach. Additionally, the MARO approach has been
eexti,t,y electricity generation capacity from technology i exist- applied to electric power systems operations [32], such as unit commit-
ing in year y with an age of t, and t = 0 indicates new ment problems [33] and the investigation of demand response uncer-
units deployed in year y tainties [34]. However, there are no research works that use MARO for
eimps,y annual electricity importation in year y from source s power systems’ decarbonization planning, to the best of our knowledge.
Therefore, a knowledge gap exists in addressing the renewable electric-
Parameters ity transition optimization under uncertainties from multiple time peri-
CBDi,t,y unit retirement cost of electricity storage units for tech- ods within the planning horizon. To fill this knowledge gap, this study
nology i with age of t in year y aims to propose a data-driven MARO framework for power systems’ re-
CBEi,t,y annual cost of power storage units using technology i newable transition. To achieve this goal, multiple research challenges
with age of t in year y for unit storage capacity need to be addressed in this work. The first one is to develop a novel elec-
CEDi,t,y unit decommissioning cost of power generators for tech- tricity transition MARO framework that involves systems uncertainties,
nology i with age of t in year y climate and renewable electricity generation targets, as well as systems
stability, to overcome the over-conservation of static robust optimiza-

2
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

tion approaches. The second challenge is to address the computational 2. Problem statement
intractability of the MARO problems. The third challenge is to integrate
data-driven technique advances regarding big-data and machine learn- In this work, a bottom-up data-driven MARO framework is proposed
ing to further overcome the over-conservative issues resulted from the to determine potential renewable transition pathways for the electricity
construction of uncertainty sets. sector. The planning horizon is partitioned equally into multiple time
In this work, we propose a data-driven bottom-up power systems’ periods on an annual basis. The uncertain parameters include the elec-
transition MARO framework to address the 100% decarbonization plan- tricity generation demand, the market availability of wind turbines, and
ning optimization under multiple uncertainties, including the electricity the market availability of solar PVs in each time period. Notably, the
generation demands, the market availability of wind turbines, and the sequences of uncertainty realizations are handled in the proposed re-
market availability of solar photovoltaics (PVs). Specifically, a novel newable electricity transition MARO framework, so is the sequential
bottom-up MARO framework is developed for the power systems’ re- decision-making process across the planning horizon. The decision vari-
newable transition planning. Note that the top-down approach is not ables that define the renewable electricity transition pathway include
selected for this study, because it does not focus on individual tech- the installation, operation, and deactivation for both the electricity gen-
nologies or facilities [35]. On the other hand, the bottom-up model eration and storage units, and electricity importation for each time pe-
includes each power generation facility and technology, so it can be riod within the planning horizon.
more appropriate for investigating the electricity transition process that In terms of the objective, the MARO framework aims to determine
is directly related to active technologies and facilities. The proposed the renewable electricity transition pathway that minimizes the total
framework ensures that the reliability of power systems is maintained system cost associated with electricity transition while handling all po-
throughout the transition, and the framework also involves proposed re- tential realizations of the uncertainties. In the proposed MARO frame-
newable electricity generation targets, renewable generation technology work, parameters and input data comprise the initial information about
targets, climate targets, and scheduled changes to the energy systems, power generation capacities and electricity storage capacities, climate
such as future additions of generators and electricity storage facilities. targets, renewable generation targets, scheduled system changes, a set of
Moreover, to manage the intermittency issues resulted from a power technology options for the power systems, as well as their corresponding
system with potentially high penetration of variable renewable energy, economic and carbon intensity data. Notably, the proposed framework
electricity storage and increasing technological diversity are involved in ensures that the system reliability for the electricity sector is maintained
the proposed MARO framework. In terms of processing the uncertainty throughout the transition. Specifically, the electricity generation de-
datasets, both the conventional approaches and machine-learning-based mand and supply are balanced each year, and the framework integrates
data-driven techniques are implemented. Specifically, multiple machine mandatory electricity storage requirements and increasing the techno-
learning techniques are used in this study to construct data-driven un- logical diversity to manage intermittency issues of the potentially high
certainty sets, including a variational inference algorithm for Dirich- penetration of variable renewable energy. The optimization formulation
let process mixture model (DPMM) [36], principal component analysis of the renewable electricity transition MARO framework is presented in
(PCA), and kernel density estimation (KDE) [37]. Notably, the MARO re- the next section, and the reformulated robust counterpart using affine
newable electricity transition framework is computationally intractable, decision rules is shown in the Appendix.
owing to the multi-level optimization structure of its objective function
and the semi-infinite constraints, so affine decision rules are applied to 3. Renewable electricity transition MARO framework and
reformulate the MARO model into its robust counterpart and facilitate data-driven uncertainty sets
the solution process for optimization. To illustrate the applicability of
the proposed electricity transition framework, a case study on the 100% Following the problem statement, the optimization formulation of
renewable power systems’ transition of New York State under electric- the renewable electricity transition MARO framework and the construc-
ity generation uncertainties and uncertain market availabilities of re- tion of data-driven uncertainty set using machine learning techniques
newable power generation units is presented. The potential electricity are presented in this section, aiming to address the optimization of
transition pathways under multiple cases are investigated, including a 100% renewable power systems transition under multiple types of un-
deterministic planning case that involves no uncertainties, two robust certainty. The overview of the proposed framework is shown in Fig. 1.
optimization cases under conventional uncertainty sets, and two robust The planning horizon is partitioned on an annual basis, and the re-
optimization cases under data-driven uncertainty sets using machine newable electricity transition planning covers all time periods within
learning techniques. the planning horizon. Notably, the proposed framework deals with
The major novelties of this work are summarized as follows: both the sequential decision-making process for power systems’ renew-
able transition and the sequences of uncertainty realizations. To avoid
• The first study using robust optimization for 100% renewable power
over-conservative issues of conventional uncertainty sets [38], machine
systems transition planning under uncertainty that handles the se-
learning techniques [39] are used to construct data-driven uncertainty
quences of uncertainty realizations and the sequential decision-
sets, and the techniques include a variational inference algorithm for
making process.

DPMM, PCA, and KDE. By integrating the uncertainty set information,
Application of the proposed framework for the renewable transition
power systems’ renewable transition under uncertainty is mathemati-
of power systems in New York State under uncertainty, using both
cally programmed into an electricity transition MARO framework. Since
conventional approaches and machine learning techniques to con-
cost minimization approaches are recommended for investigating the
struct uncertainty sets.

technological transition pathways and climate impacts based on elec-
On-land wind and offshore wind would be the major power sources
tricity generation and climate targets [40], the optimal renewable elec-
based on the deterministic planning and the MARO results, respec-
tricity transition pathways are obtained on a least-cost basis that min-
tively.
imizes total system cost under all potential realizations of uncertainty
The remainder of this article is organized as follows: §2 provides over the entire transition period. Note that the proposed electricity tran-
the problem statement for the power systems’ renewable transition op- sition MARO problem cannot be solved directly using off-the-shelf op-
timization under uncertainty. §3 presents the optimization formulation timization solvers. To address the intractability issue, we apply affine
of the electricity transition MARO model and the construction of data- decision rules [41] and reformulate the problem into its linear program-
driven uncertainty sets using machine learning techniques. The applica- ming robust counterpart to facilitate the solution process [42]. Lastly,
tion of the proposed renewable electricity transition MARO framework the renewable electricity transition pathways could be determined us-
for New York State is shown in §4. Conclusions are drawn in §5. ing computational optimization based on the reformulated model and

3
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Fig. 1. The overview of the proposed frame-


work for determining renewable electricity
transition pathways under uncertainty, includ-
ing the structure of the electricity transition
problem, the construction of data-driven uncer-
tainty set using machine learning techniques,
mathematical programming of power systems’
renewable transition under uncertainty as a
MARO framework, and the determination of
electricity transition pathways using computa-
tional optimization.

the information on data-driven uncertainty sets. It is worth mention- tion approach may be more suitable for renewable electricity transition
ing that the proposed MARO framework focuses on the long-term re- planning. Note that robust optimization also has limitations, including
newable transition optimization for the electricity sector, which would its limited ability to provide a unified hedging strategy under uncer-
inform and facilitate the policy-making process, in terms of the feasi- tainty and its incapability for investigating in-depth information regard-
bility, economics, and technology selection aspects of the renewable ing the uncertainties when the probability distributions and covariance
power systems’ transition under uncertainty. Short-term hourly simu- can be determined [53].
lations [43] are not included in this work, and it could be a valuable
future research extension that investigates the adequacy of the fully- 3.1. General form of the MARO framework and uncertainty sets
decarbonized power systems. Instead, considering that electricity stor-
age is a practical approach to deal with the intermittency issues [44], A general form of the proposed electricity transition MARO is shown
we estimate the energy storage need to address the intermittency is- in problem (1), where u1 , …, uYM are uncertainties revealed over all
sues from the potentially high penetration of variable renewable en- stages within the planning horizon, vector x0 denotes the “here-and-
ergy, based on a study from the U.S. Department of Energy [45], which now” decision variables that are determined prior to any realizations
is elaborated in §3.3. of uncertainty, x1 , …, xYM represent “wait-and-see” decisions that are
Note that the robust optimization approach is applied in this frame- made from stage 1 to stage YM. YM stands for the last year in the plan-
work to deal with systems uncertainties, instead of stochastic program- ning horizon. Ω(x0,…, xt-1, u1,…, ut ) indicates the feasible region of the
ming or chance constraints, while stochastic programming, robust opti- recourse decisions at stage t, which is affected by both the previous re-
mization, and chance constraints are three major approaches that handle alizations of uncertainties u1 , …, ut-1 and the past determined decisions
systems uncertainties by bringing in conservatism [46]. In terms of the x0 , …, xt-1 . Moreover, the objective of the proposed framework is to
chance constraint approach, it requires a probability density function minimize the total systems cost related to the decarbonization transition
before calculating the probabilities [47], and a key assumption is that across the planning horizon. Therefore, c0 , …, cYM represent the vectors
the probability density function is known before solving the problem of cost coefficients that are associated with all decision variables.
[48]. However, it is practically infeasible to obtain such precise func- {
tions for the energy systems uncertainties in the future. As for stochastic min 𝐜′ 0 𝐱0 + max min
𝐱0 𝐮1 ∈𝑈1 𝐱1 ∈Ω(𝐱0 ,𝐮1 )
programming, the approach has been successful in various applications { }}
owing to its feature of simultaneous optimization of design variables un- 𝐜′ 1 𝐱1 + ⋅ ⋅ ⋅ + max min 𝐜′ 𝑌 𝑀 𝐱𝑌 𝑀 (1)
der uncertainty [49], it may not be appropriate for power systems de- 𝐮𝑌 𝑀 ∈𝑈𝑌 𝑀 𝐱𝑌 𝑀 ∈Ω(𝐱0 ,⋅⋅⋅,𝐱𝑌 𝑀−1 ,𝐮1 ,⋅⋅⋅𝐮𝑌 𝑀 )
carbonization compared to robust optimization for three reasons. First,
Following the problem statement, the uncertainties for the electric-
stochastic programming requires knowledge of the probability distribu-
ity transition model consist of the electricity generation demands, the
tions of uncertainty sets in advance [50], and this can be unrealistic in
market availability of wind turbines, and the market availability of solar
practice for the electricity transition application. On the other hand, ro-
PVs. The data-driven uncertainty set for each stage could be expressed
bust optimization needs no predetermined probability distributions of
in a general form as shown in Eq. (2).
uncertainty. Second, stochastic programming can become computation-
ally intractable for large-scale applications with many possible scenar- { ∑ }
𝜃 ⋅ 𝑢 ≤ 𝜔𝑙,𝑛 , ∀𝑙 ∈ 𝐿
ios, which could be the case for power systems decarbonization, while 𝑈𝑛 = 𝐮𝑛 | 𝑘∈𝐾 𝑙,𝑘,𝑛 𝑘,𝑛 , ∀𝑛 ∈ 𝑁 (2)
𝑢𝑘,𝑛 ≥ 0, ∀𝑘 ∈ 𝐾, 𝑛 ∈ 𝑁
robust optimization could be computationally tractable for large prob-
lems [36]. Third, robust optimization solutions are guaranteed to be where N denotes the set of years within the planning horizon, L indicates
feasible for all potential realizations of systems uncertainties [51], in- the set of hyperplanes that formulate the uncertainty sets, and K stands
cluding the worst-case scenarios [52], which is essential for the stability for the set of uncertain parameters. Specifically, the electricity genera-
of power systems during the transition. Therefore, the robust optimiza- tion demands, the market availability of wind turbines, and the market

4
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

availability of solar PVs in each time period n are represented by k = 1, an annual basis, and this has been applied in research works that inves-
k = 2, and k = 3, respectively. Un represents the data-driven uncertainty tigate the energy systems transition optimization with high-penetration
set for time period n, and uk,n is the uncertainty parameter k in year n. of variable renewable energy [40]. Note that an inequality constraint is
𝜃 l,k,n and 𝜔l,n denote the coefficient and the constant for hyperplane l in used, considering that the capacity factor is a fixed parameter for each
time period n. Furthermore, the uncertainty set U is a direct product of technology and each year. If an equation is used for this constraint, then
the uncertainty sets from different stages, namely U = U1 × U2 × …UYM . the electricity generation capacities must change according to demand
Notably, the objective function of problem (1) has a multi-level op- fluctuations, which would be unrealistic in practice.
timization structure, while the constraints of the proposed framework ∑ ∑∑
𝑒𝑖𝑚𝑝𝑠,𝑦 + 𝐹𝑖,𝑦 ⋅ 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 ≥ 𝑢𝑘=1,𝑛=𝑦 , ∀𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (5)
are semi-infinite, given the formulation of uncertainty sets as shown 𝑠∈𝑆 𝑖∈𝐼𝑒 𝑡∈𝑇
in Eq. (2). Therefore, the renewable electricity transition problem un-
der uncertainty cannot be solved directly using any off-the-shelf opti- Considering that the potentially high penetration of variable renew-
mization solvers. To make the data-driven electricity transition MARO able energy may raise issues on intermittency and curtailment, electric-
computationally tractable, affine decision rules [41] are implemented ity storage is included as a balancing option to deal with these issues in
to derive the robust counterpart of the MARO model, namely a single- the proposed MARO framework. Following the results and conclusions
level optimization problem with a finite number of constraints, which in the work of Kintner-Meyer et al. [45], it is assumed that the mini-
can be solved efficiently using off-the-shelf solvers like CPLEX. The for- mum electricity storage capacities should be determined based on the
mulation of the robust counterpart derived using affine decision rules is generation capacities using variable renewable energy. Specifically, the
presented in the appendix. electricity storage capacity is required to be higher than a proportion of
the total generation capacity from variable renewable energy sources,
3.2. Objective function as presented in constraint (6).
∑∑ ∑∑
𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦 ≥ 𝑅𝐸𝑄𝑀𝑖 ⋅ 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 , ∀𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (6)
The objective function of minimizing the total electricity transition 𝑖∈𝐼𝑠 𝑡∈𝑇 𝑖∈𝐼𝑒𝑖 𝑡∈𝑇
cost as presented in general model (1) should hold for all potential real-
where REQMi indicates the percentage of marginal capacity addition
izations of uncertain parameters. Therefore, the cost minimization ob-
based on variable renewable generation technology I, set Iei represents
jective function is equivalent to Eq. (3) and constraint (4). z represents
the set of variable renewable generation technologies, including solar
the value of the objective function, namely the total electricity transi-
PVs and wind turbines.
tion cost throughout the planning horizon, while the right-hand side of
constraint (4) accounts for the discounted total costs associated with
3.4. Target constraints on renewable electricity generation and emissions
electricity generation, generator decommissioning, electricity storage,
the retirement of storage units, and electricity importation.
Multiple targets and mandates have been proposed by the govern-
min 𝑧 (3) ment in terms of power systems decarbonization, including a carbon
neutrality target for the electric power sector [6], emissions targets from
∑∑ ∑ electricity generation, targets on renewable electricity, and technology
𝑧≥ CE𝐸𝑖,𝑡,𝑦 ⋅ eex𝑡𝑖,𝑡,𝑦
𝑖∈𝐼𝑒 𝑡∈𝑇 𝑦∈𝑌 −{YINI}
deployment targets [7]. Notably, these constraints could be easily cus-
∑∑ ∑ tomized to accommodate such targets for any other regions or time pe-
+ CE𝐷𝑖,𝑡,𝑦 ⋅ edi𝑠𝑖,𝑡,𝑦 riods.
𝑖∈𝐼𝑒 𝑡∈𝑇 𝑦∈𝑌 −{YINI}
∑∑ ∑ The annual greenhouse gas emissions from electricity generation are
+ CB𝐸𝑖,𝑡,𝑦 ⋅ bex𝑡𝑖,𝑡,𝑦 presented in constraint (7), where EEi,t,y is the annual greenhouse gas
𝑖∈𝐼𝑠 𝑡∈𝑇 𝑦∈𝑌 −{YINI} emissions from unit generation capacity with an age of t in year y, and
∑∑ ∑
+ CB𝐷𝑖,𝑡,𝑦 ⋅ bdi𝑠𝑖,𝑡,𝑦 ELMTy stands for the upper limit of annual greenhouse gas emissions
𝑖∈𝐼𝑠 𝑡∈𝑇 𝑦∈𝑌 −{YINI} from the electricity sector in year y.
∑ ∑ ∑∑
+ CIM𝑃𝑠,𝑦 ⋅ eim𝑝𝑠,𝑦 , ∀𝐮 ∈ 𝑈 (4) 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 ⋅ 𝐸 𝐸𝑖,𝑡,𝑦 ≤ 𝐸𝐿𝑀 𝑇𝑦 , ∀𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (7)
𝑠∈𝑆 𝑦∈𝑌 −{YINI} 𝑖∈𝐼𝑒 𝑡∈𝑇

where CEEi,t,y denotes the annual cost of electricity generation technol- Targets on renewable electricity generation are presented in con-
ogy i with an age of t in year y for unit generation capacity, including straint (8). RPSGy denotes the renewable portfolio standards for renew-
its capital costs, variable maintenance (O&M) costs, fixed O&M costs, able generation technologies in stage y, uk=1 ,n = y is an uncertain pa-
and emissions costs owing to carbon tax policies. CBEi,t,y stands for the rameter indicating the electricity generation demands in time period y,
annual cost of power storage units using technology i with an age of and Ier represents the set of renewable power generation technologies.
t in year y for unit storage capacity. CEDi,t,y and CBDi,t,y are the unit ∑∑
decommissioning cost of power generators and the unit retirement cost 𝐹𝑖,𝑦 ⋅ 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 ≥ 𝑅𝑃 𝑆 𝐺𝑦 ⋅ 𝑢𝑘=1,𝑛=𝑦 , ∀𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (8)
of electricity storage units, respectively, for technology i with an age of 𝑖∈𝐼𝑒𝑟 𝑡∈𝑇
t in year y. CIMPs,y indicates the unit cost of imported electricity from The deployment constraints for renewable generation technologies
source s in year y. and electricity storage units are shown in constraints (9) and (10), re-
spectively. The deployment targets for renewable generation technology
3.3. Electric power generation demand and electricity storage requirement i in year y are formulated into constraint (9), where TARi,y denotes the
constraints target on total power generation capacity level for renewable power
generation technology i in year y.
The renewable electricity transition should guarantee that the elec- ∑
tric power systems provide enough electricity for the region-level de- 𝑒𝑒𝑥𝑡 𝑖,𝑡,𝑦 ≥ 𝑇 𝐴𝑅 𝑖,𝑦 , ∀𝑖 ∈ 𝐼𝑒𝑟 , 𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (9)
𝑡∈𝑇
mands, to avoid significant damage to social and economic activi-
ties. The electricity generation demands are presented by constraint The electricity storage capacity targets are formulated into constraint
(5), where the parameter Fi,y denotes the annual electricity generation (10), where TARSy stands for the target on total electricity storage ca-
amounts from unit generation capacity using technology i in year y. pacity for year y.
uk=1 ,n = y represents the electricity generation demands in stage y, as il- ∑∑
𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦 ≥ 𝑇 𝐴𝑅𝑆𝑦 , ∀𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (10)
lustrated in §3.1. The electricity generation demands are considered in 𝑖∈𝐼𝑠 𝑡∈𝑇

5
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

3.5. Transition relationship constraints market availability of technology i in stage y. On the other hand, such
constraints for wind and solar power capacities are formulated as con-
The initialization of variables regarding electricity generation and straints (21) and (22), considering that the market availabilities of wind
storage capacities is presented as constraints (11) and (12), where INIEi,t turbines and solar PVs are uncertainties in the framework. uk=2 ,n = y and
stands for the electricity generation from technology i with an age of t in uk=3 ,n = y symbolize the uncertain market availability for wind turbines
the initial year YINI, and INIBi,t represents the initial electricity storage and solar PVs, respectively.
capacity based on technology i with an age of t. ∑ ∑
𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 − 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦−1 ≤ 𝑀𝐴𝐺𝑖,𝑦 ,
𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦=𝑌 𝐼 𝑁 𝐼 = 𝐼 𝑁 𝐼 𝐸𝑖,𝑡 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝐮 ∈ 𝑈 (11) 𝑡∈𝑇 𝑡∈𝑇
∀𝑖 ∈ 𝐼𝑒 − {𝑊 𝑂𝑁, 𝑊 𝑂𝐹 𝐹 , 𝑆𝑈 𝑁𝐷, 𝑆𝑈 𝑁𝑈 }, 𝑦 ∈ 𝑌 − {𝑌 𝐼𝑁𝐼}, 𝐮 ∈ 𝑈
𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦=𝑌 𝐼 𝑁 𝐼 = 𝐼 𝑁 𝐼 𝐵𝑖,𝑡 , ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝐮 ∈ 𝑈 (12)
The transition relationships for electricity and storage capacities be- ∑ ∑ ∑ ∑
𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 − 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦−1
tween consecutive time periods are formulated as constraints (13) and 𝑖∈{𝑊 𝑂𝑁,𝑊 𝑂𝐹 𝐹 } 𝑡∈𝑇 𝑖∈{𝑊 𝑂𝑁,𝑊 𝑂𝐹 𝐹 } 𝑡∈𝑇
(14), respectively. Notably, a generation or storage unit with an age of
≤ 𝑢𝑘=2,𝑛=𝑦 , ∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝐮 ∈ 𝑈 (21)
t = 0 in stage y indicates that the unit has just been installed in year y.
The two transition relationship constraints show that the existing gener- ∑ ∑ ∑ ∑
ation or storage capacities using technology i in year y with age t should 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 − 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦−1
equal to the existing capacities using this technology with an age of t-1 𝑖∈{𝑆 𝑈 𝑁𝐷,𝑆 𝑈 𝑁𝑈 } 𝑡∈𝑇 𝑖∈{𝑆 𝑈 𝑁𝐷,𝑆 𝑈 𝑁𝑈 } 𝑡∈𝑇
in the previous year that are not retired in year y. ≤ 𝑢𝑘=3,𝑛=𝑦 , ∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝐮 ∈ 𝑈 (22)

𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 = 𝑒𝑒𝑥𝑡𝑖,𝑡−1,𝑦−1 − 𝑒𝑑𝑖𝑠𝑖,𝑡,𝑦 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 The market availability constraint for electricity storage units is
−{𝑌 𝐼 𝑁 𝐼 }, 𝑡 ≥ 1, 𝐮 ∈ 𝑈 (13) shown in constraint (23). MABi,y denotes the annual market availability
of electricity storage technology i in stage y.
∑ ∑
𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦 = 𝑏𝑒𝑥𝑡𝑖,𝑡−1,𝑦−1 − 𝑏𝑑𝑖𝑠𝑖,𝑡,𝑦 , ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦 − 𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦−1 ≤ 𝑀𝐴𝐵𝑖,𝑦 , ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼𝑁𝐼}, 𝐮 ∈ 𝑈
−{𝑌 𝐼 𝑁 𝐼 }, 𝑡 ≥ 1, 𝐮 ∈ 𝑈 (14) 𝑡∈𝑇 𝑡∈𝑇
(23)
The power generators and electricity storage facilities are assumed
to retire if their ages are larger than their lifetime, as presented in con- 3.8. Electricity importation constraints and non-negative constraints
straints (15) and (16), where LIFEi stands for the lifetime of generation
technology i or storage technology i. The amount of inter-state or international electricity importation
is physically limited by the transmission capacity, so the maximum
𝑒𝑒𝑥𝑡𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 −1,𝑦−1 = 𝑒𝑑𝑖𝑠𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 ,𝑦 , ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (15)
amount of electricity importation should be derived and included in
𝑏𝑒𝑥𝑡𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 −1,𝑦−1 = 𝑏𝑑𝑖𝑠𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 ,𝑦 , ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (16) the proposed framework, as presented in constraint (24), where IMPMs,y
symbolizes the maximum electricity importation amounts from external
3.6. Scheduling constraints source s in time period y. In this work, it is assumed that the imported
electricity is generated from carbon-neutral sources, considering that
For obtaining more accurate and realistic renewable electricity tran- hydropower from Ontario is a major electricity importation source for
sition pathways, we include the scheduled systems changes in the pro- New York State [54].
posed framework, such as the proposed additions of generators, the 𝑒𝑖𝑚𝑝𝑠,𝑦 ≤ 𝐼 𝑀 𝑃 𝑀𝑠,𝑦 , ∀𝑠 ∈ 𝑆, 𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (24)
scheduled decommissioning of power generators, and the scheduled in-
stallation of storage facilities. These scheduled changes could have great The non-negative constraints for the decision variables of the renew-
impacts on the electric power systems during the transition, considering able electricity transition MARO model are formulated into constraints
that the government or independent system operators may have pro- (25)-(29).
posed such schedules regarding the large power generation facilities. 𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (25)
The proposed generator additions and proposed storage unit addi-
tions are represented by constraints (17) and (18), respectively. PROGi,y 𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (26)
is the proposed capacity additions for power generation using technol-
ogy i in year y. PROBi,y stands for the scheduled installation of electricity 𝑒𝑑𝑖𝑠𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (27)
storage capacity with technology i in year y.
𝑏𝑑𝑖𝑠𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (28)
𝑒𝑒𝑥𝑡𝑖,𝑡=0,𝑦 ≥ 𝑃 𝑅𝑂𝐺𝑖,𝑦 , ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (17)
𝑒𝑖𝑚𝑝𝑠,𝑦 ≥ 0, ∀𝑠 ∈ 𝑆, 𝑦 ∈ 𝑌 (29)
𝑏𝑒𝑥𝑡𝑖,𝑡=0,𝑦 ≥ 𝑃 𝑅𝑂𝐵𝑖,𝑦 , ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (18)
3.9. Renewable electricity transition MARO model formulation and
The scheduled decommissioning of power generators are represented
reformulated robust counterpart
using constraint (19), where PRODi,t,y denotes the generation capacity
associated with the generators using technology i with age t in year y
The detailed optimization formulation of the electricity transition
that is scheduled for deactivation.
MARO model is presented in the previous subsections. The objective is
𝑒𝑑𝑖𝑠𝑖,𝑡,𝑦 ≥ 𝑃 𝑅𝑂𝐷𝑖,𝑡,𝑦 , ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 , 𝐮 ∈ 𝑈 (19) the minimization of total transition cost, in terms of electricity gener-
ation and storage. The objective function is subject to electricity gen-
3.7. Market availability constraints eration demand constraints, electricity storage requirement constraints,
renewable electricity generation and climate target constraints, technol-
The annual increases of generation capacities and storage capacities ogy deployment constraints, transition relationship constraints, schedul-
are limited by the market availability for each technology in the elec- ing constraints, market availability constraints, electricity importation
tricity transition MARO framework. The market availability constraint constraints, non-negative constraints, and uncertainty set formulation.
for generation technologies that are not using variable renewable en- We present the outline of the renewable electricity transition MARO
ergy is shown as constraint (20), where MAGi,y represents the annual framework as (P0).

6
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

(P0) min-max Total Electricity Transition Cost Under Uncertainty given New York State under uncertainty. Namely, the framework proposed
by (1) in the previous section is applied to New York State to investigate the
s.t. Electric power generation demand constraint: (5) most economically efficient electricity transition pathways for the elec-
Electricity storage requirement constraint: (6) tric power sector that could achieve the climate and decarbonization
Renewable electricity generation and emissions targets constraints: targets, ensure the balance of electricity supply and demands, and main-
(7)–(8) tain systems stability. Specifically, the climate and renewable electricity
Technology deployment constraints: (9)–(10) generation targets for the state are set by both the state-level legislation
Transition relationship constraint: (11)–(16) [7] and a nationwide energy transition plan [6]. As for decarbonization,
Scheduling constraint: (17)–(19) New York State is scheduling to reduce 40% of greenhouse gas emissions
Market availability constraints: (20)–(22) by 2030, and the nationwide energy transition plan aims to achieve
Electricity importation and non-negativity constraints: (24)–(29) 100% carbon neutrality for the electricity sector by 2035. Additionally,
Uncertainty set formulation: (2) the state also requires that 70% of electricity should be generated from
renewable sources by 2030. As for the technology deployment targets,
Considering that the MARO framework (P0) has a multi-level op- New York State mandates minimum capacity requirements through its
timization structure and semi-infinite constraints, it cannot be solved CLCPA for multiple types of generation technologies and electricity stor-
directly using off-the-shelf optimization solvers. To address the compu- age, such as distributed solar PV and offshore wind turbines [7]. Specif-
tational issue, affine decision rules and duality theories are implemented ically, the capacity of distributed solar PV should be higher than 6 GW
to reformulate (P0) into its single-level linear optimization robust coun- by 2025, and the installed generation capacity for offshore wind should
terpart with a finite number of constraints. The mathematical formu- be at least 9 GW by 2035. As for electricity storage, 1.5 GW and 3 GW
lation of the robust counterpart is presented in the Appendix, and the of electricity storage capacities are required for New York State by 2025
outline of the reformulated electricity transition framework is shown and 2030, respectively. These requirements serve as the lower bounds
as (P1). for the application of renewable electricity generation systems in the
(P1) min Total Electricity Transition Cost given by (3), (A12)–(A14) state. Two electricity storage technologies are considered in this appli-
s.t. Reformulated electric power generation demand constraints: cation, including lithium-ion batteries and flywheels, considering that
(A16)–(A19) they are the existing electricity storage technologies in New York State
Reformulated electricity storage requirement constraints: (A20)– [57]. The generation technologies in this application include bituminous
(A22) coal, fuel oil, methane from biogas, refuse of solid waste, utility solar
Reformulated renewable generation and emissions targets con- PV, distributed solar PV, nuclear power, hydropower, on-land wind, off-
straints: (A23)–(A29) shore wind, natural gas combined cycle, natural gas combustion turbine,
Reformulated technology deployment constraints: (A30)–(A35) natural gas steam turbine, natural gas combined cycle with carbon cap-
Reformulated transition relationship constraints: (A36)–(A47) ture and storage. Notably, the proposed framework is formulated in a
Reformulated scheduling constraints: (A48)–(A56) bottom-up structure that classifies the electricity generation units by
Reformulated market availability constraints: (A57)–(A70) their technologies and ages, so it would be convenient to include differ-
Reformulated electricity importation and non-negativity con- ent sets of technologies into framework while applying to other regions
straints: (A71)–(A88) in the world. Moreover, a carbon tax policy is included in this applica-
tion, following a previous research work [58]. Specifically, the carbon
3.10. Data-driven uncertainty set using machine learning tax in this study is US$40 per ton CO2 in 2020, and it increases by 5% per
year. The projected technological and economic parameters for the gen-
To avoid over-conservative issues of conventional uncertainty sets, eration technologies are obtained from a work of the National Renew-
machine learning techniques, including a variational inference algo- able Energy Laboratory [59]. Considering that the proposed bottom-up
rithm using DPMM, PCA, and KDE, are used to construct data-driven MARO framework considers electricity generation by technology types
uncertainty sets for the proposed renewable electricity transition MARO and generator ages, application to different planning regions may in-
framework. The resulting data-driven uncertainty sets are formulated crease the size of the optimization problem, in terms of its number of
into sets of inequalities directly from the uncertainty data. Two types of variables and the number of constraints, as more types of technologies
data-driven uncertainty sets are constructed. The first type uses a varia- could be involved, but the solution time would not change too much for
tional inference algorithm using DPMM [55]. Considering that DPMM is different scales, if the same type of uncertainty set is applied.
a Bayesian nonparametric model, its model complexity can be adjusted Three types of uncertain parameters are included in this application,
according to the uncertainty data [56]. Note that the model parameters including the annual electricity generation demands, the market avail-
are considered as random variables in Bayesian statistics. To obtain the ability for wind turbines, and the market availability for solar PVs in
posterior distribution of these random variables, the variational infer- each stage. The first type of uncertain parameter is the annual electric-
ence algorithm is used. Subsequently, the data-driven uncertainty sets ity generation demand. The uncertainty data samples for annual gener-
can be constructed using both l1 and l∞ norms based on the posterior dis- ation demands are collected and projected based on multiple load and
tribution information [36]. The second type of data-driven uncertainty capacity reports that are published by the New York Independent Sys-
set uses both PCA and KDE [37]. First, PCA is employed to obtain the tem Operator (NYISO) [57], which include future electricity generation
orthogonal principal components based on the uncertainty data, then demand projections. Considering that the future generation demands in
these data are projected onto each principal component accordingly. some of the reports are not projected to 2035, we made further pro-
Next, KDE is applied to extract the distributional information from the jections based on 10–15 years of the NYISO data. It is worth mention-
projected data for each principal component. Lastly, the principal com- ing that the electricity demand increases owing to the potential shift
ponents and the extracted distributional information are integrated to from internal combustion vehicles to electric vehicles are included in
formulate the data-driven uncertainty set. the uncertainty data, as well as the transmission and distribution losses
for electricity. Market availability for wind turbines is the second type
4. Application: New York State 100% renewable electricity of uncertainty. The data samples are collected and projected based on
transition planning under uncertainty the total wind capacities from domestic manufacturers, as reported by
the U.S. Department of Energy [60]. The third type of uncertainty is
In this section, the applicability of the proposed framework is demon- the market availability for solar PVs, and the corresponding uncertainty
strated through the application to renewable electricity transition in data samples are obtained and projected from annual solar PV module

7
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Fig. 2. Multiple types of uncertainty set formu-


lation for the year 2030. (a) Box uncertainty
set. (b) Budget uncertainty set. (c) Data-drive
uncertainty set formulated using a variational
inference algorithm for DPMM. (d) Data-driven
uncertainty formulated by PCA and KDE.

Table 1 Table 2
Range of uncertain parameters for the year 2030. Solution time of the proposed MARO framework for under different uncertainty
sets and planning horizons.
Electricity Market availability Market
generation of wind turbines availability of Uncertainty set Solution time under different planning horizon (CPUs)
demand (TWh) (GW) solar PV(GW)
2020–2025 2020–2030 2020–2035
Box [155.2, 212.6] [13.6, 22.0] [2.71, 4.51]
Budget [155.2, 238.9] [13.6, 24.9] [2.71, 5.30] Deterministic 0.1 0.2 0.3
DPMM [155.2, 212.6] [13.6, 22.0] [2.71, 4.51] Box 8 56 266
PCA & KDE [164.2, 212.2] [13.7, 21.4] [2.71, 4.51] Budget 50 340 3703
DPMM 99 642 824
PCA & KDE 64 508 1550

shipment reports published by the U.S. Energy Information Administra-


tion [61].
represented by red crosses. In addition, the renewable electricity transi-
Five cases are investigated based on the formulation of uncertainty
tion pathways are investigated based on the optimization results, from
set, including deterministic planning, box uncertainty sets, budget un-
research perspectives that include energy sources, generation technol-
certainty sets, data-driven uncertainty sets using a variational inference
ogy contributions, economics, and greenhouse gas emissions.
algorithm for DPMM, and data-driven uncertainty sets formulated using
PCA and KDE. Specifically, the deterministic planning case includes no
uncertainties, so its optimization results are not robust. The four types 4.1. Computational optimization results
of uncertainty set for 2030 are visualized in Fig. 2, and the uncertainty
ranges under all types of uncertainty set in 2030 are shown in Table 1. Both the deterministic planning and the robust optimization of the
Note that the ranges of the uncertainty parameters for the data-driven renewable electricity transition problem for New York State are coded
uncertainty sets are included in the ranges for the box uncertainty set, in GAMS 27.3 [62] on a PC with an Intel Core i7–8700 @ 3.20 GHz
because the upper and lower bounds of the uncertainties in the data- and 32.00 GB RAM, running on a Windows 10 Enterprise, 64-bit oper-
driven uncertainty sets are constrained by the maximum and minimum ating system. The deterministic planning and the reformulated MARO
values of the uncertainty data samples. Namely, the data-driven uncer- problems are solved using CPLEX 12.9.0.0 with an optimality tolerance
tainty sets are subsets of the box uncertainty set. In Fig. 2, the x axis, y of 0%. The optimal objective value and the solution time are two main
axis and z axis represent the uncertainties of annual electricity demand, indexes to evaluate an uncertainty set. The optimal objective value rep-
market availability of wind turbines, and the market availability of solar resents the minimum power systems transition cost for the proposed
PV, respectively. The box uncertainty set and the budget uncertainty set framework, so smaller objective values are preferred. The solution time
for 2030 are shown in Fig. 2(a) and Fig. 2(b). Both types of uncertainty stands for the execution time of the optimization program, and shorter
set are not data-driven uncertainty set, and all data samples are included solution time is better.
in the box and budget uncertainty sets. On the other hand, Fig. 2(c) and The solution times of the optimization programs for different plan-
Fig. 2(d) visualize the DPMM uncertainty set and the PCA & KDE un- ning horizons are presented in Table 2. The deterministic case has the
certainty set for 2030, respectively. These data-driven uncertainty sets shortest solution times of less than 1 CPU second for all planning hori-
are tuned to have the same coverage level of data samples (90%). In zons, while the robust optimization cases tend to have much longer so-
Fig. 2, The data samples included in the data-driven uncertainty sets lution times. This is because the optimization problems for deterministic
are plotted using red dots, while the samples that are not included are planning cases have significantly smaller problem sizes than the robust

8
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Table 3 types of uncertainty set. This is because the annual electricity generation
Optimal objective values under different uncertainty sets and planning horizons. demand is an uncertain parameter, and the ranges of its potential real-
Total systems cost under different planning horizon ($MM) izations are likely to be different for different types of uncertainty sets.
Specifically, the total electricity generation during the renewable elec-
2020–2025 2020–2030 2020–2035
tricity transition under uncertainty is 12%, 23%, 13%, and 12% higher
Deterministic 75,778 105,963 124,763 than the generation of the deterministic solution for box uncertainty set,
Box 96,511 121,371 156,684
budget uncertainty set, DPMM data-driven uncertainty set, and PCA &
Budget 111,819 139,638 177,326
DPMM 94,701 119,067 152,980 KDE data-driven uncertainty set, respectively. The electricity generation
PCA & KDE 92,646 117,270 150,941 under the budget uncertainty set is likely to be much higher compared
to other types of uncertainty set, and this can be illustrated by Fig. 2(c)
which presents a higher range for potential realizations of uncertain
optimization cases, considering that a number of dual variables and dual electricity generation. Note that there is no electricity importation in
problem constraints are introduced while implementing affine decision 2033 and 2034 for both the deterministic and the robust solutions, and
rules. Note that several factors can affect the solution time of the MARO electricity importation may return to a high level at the end of the plan-
problem, including the sparseness formulation, the size of the feasible ning horizon. This is because natural gas-based electricity still exists in
region, and the number of auxiliary variables. The budget uncertainty the two stages before 2035, and replacing such carbon-emitting genera-
set for planning horizon 2020–2035 leads to the largest solution time, tion technologies with electricity importation and carbon-neutral tech-
and this is likely to be resulted from its non-sparse uncertainty set for- nologies at the last two years of the planning horizon could be more
mulation and its large feasible region for the uncertain parameters com- economically favorable in both the deterministic planning case and the
pared to other types of uncertainty set, as shown in Fig. 2. In compari- robust optimization cases. In addition, we note that the electricity im-
son, the DPMM uncertainty set has a similar non-sparse formulation, but portation price could be uncertain, because of the carbon-neutral transi-
its feasible region for uncertain parameters is considerably smaller than tion of the exporting states or countries. Such uncertainty is not consid-
that of the budget uncertainty set, leading to a shorter solution time for ered in this work, and it could be an important factor for future related
a longer planning horizon. The uncertainty set constructed using PCA research works.
& KDE has more auxiliary variables that may increase the solution time In terms of renewable electricity generation, renewable sources pro-
for short planning horizons, but its sparse formulation and relatively vide around 80% of total electricity generation by the end of the plan-
small uncertainty feasible region can lead to less solution time under ning horizon. Note that the optimal electricity generation from renew-
the planning horizon 2020–2035, compared to the budget uncertainty able energy sources varies under different uncertainty sets, and this is
set. because of the different feasible regions for uncertain parameters that
The optimal objective values, namely the minimum renewable elec- are defined by the uncertainty sets. Specifically, the robust optimiza-
tricity transition cost under uncertainty, for different uncertainty sets tion results should hold for all potential realizations of the uncertain
and planning horizons are shown in Table 3. The deterministic plan- parameters, in addition to the achievement of renewable generation re-
ning solutions show lower electricity transition costs compared to the quirements. Considering that the feasible regions of the data-driven un-
robust optimization results, because the potential uncertainty realiza- certainty sets tend to be smaller than that of conventional uncertainty
tions are not considered in the deterministic cases. The total systems sets, the optimal solutions obtained from the data-driven approaches
costs of the electricity transition for the robust optimization solutions may become infeasible for some extreme uncertainty realizations under
are 21%−42% higher than deterministic planning. In terms of robust the conventional uncertainty sets. Therefore, different types of uncer-
optimization, using the data-driven DPMM and PCA & KDE uncertainty tainty set can lead to different renewable electricity generation plans.
sets could achieve less conservative solutions compared to the conven- Based on optimization results, on-land wind may become a major source
tional box and budget uncertainty sets, as shown by their relatively low of electricity for New York State under the deterministic case in 2035,
objective values. Specifically, the total renewable electricity transition while offshore wind accounts for the largest proportion of electricity
costs under the data-driven uncertainty sets are 2%−17% lower than the generation by the end of the planning horizon for all cases under un-
conventional uncertainty sets. Additionally, the data-driven MARO so- certainty. The differences in the application of different types of wind
lutions could cover 90% of the potential uncertainty realizations across power for the deterministic planning and robust optimization are re-
the planning horizon, while the solution under the deterministic scenar- sulted from both the economic and the market availability reasons. Eco-
ios could only cover 24%−57% of the potential uncertainty realizations nomically, the capital investment of on-land wind turbines is less than
during the renewable power systems’ transition. offshore wind, but on-land wind turbines would need more investment
on electric energy storage for unit electricity generation. Therefore, con-
4.2. Electricity generation during the power systems’ transition under sidering that the electricity generation demand for the robust cases are
uncertainty much higher than the deterministic case, the cost of electricity storage
would significantly increase if on-land wind becomes the major renew-
The electricity generation from different sources during the power able power source in the robust cases. In terms of market availability, the
systems’ renewable transition under the deterministic planning solution offshore wind turbines have higher capacity factors [59], so they could
and the robust optimization solutions is shown in Fig. 3. It is illustrated generate more electricity under unit installed capacity compared to on-
that the complete decarbonization target for the power systems of New land wind. This is particularly important considering that the market
York State would be achieved under all deterministic and robust op- availability of wind turbines in the robust cases could be considerably
timization cases, by Fig. 3(a)–(e). Specifically, wind, hydro, solar, and less than the deterministic case. Due to the economic and market avail-
nuclear would account for the majority of electricity generation demand ability reasons, on-land wind is selected as the main power source at
by the end of the planning horizon. Notably, the electricity generation the end of the planning horizon in the deterministic case, given that
for the deterministic solution in Fig. 3(a) is much lower compared to this case has complete information about the annual generation de-
all robust optimization solutions shown in Fig. 3(b)–(e). This is because mands and larger market availability of wind turbines. On the other
the cases under uncertainty should compensate for any potential real- hand, the robust optimization solutions would ensure that the demands
izations of the uncertain electricity generation demands, which could and targets are fulfilled for all realizations of the uncertainties through-
be higher than the generation demands under the deterministic case for out the electricity transition process, so the robust optimization results
any time period within the planning horizon. Notably, the annual elec- tend to be more conservative than deterministic planning, which could
tricity generation within the planning horizon may vary for different be essential for power systems’ transition planning. Therefore, offshore

9
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Fig. 3. Electricity generation by source and


electricity importation during the power sys-
tems’ renewable transition for the deterministic
planning solution and the robust optimization
solutions under multiple types of uncertainty
sets.

wind may be a more appealing option for the major source of electricity and they would account for 23%, 24%, 21%, 24%, and 24% of elec-
generation when uncertainties are considered. Additionally, a notice- tricity generation by 2035 for box uncertainty set, budget uncertainty
able increase of the offshore wind capacity in 2024 exists for all solu- set, DPMM data-driven uncertainty set, and PCA & KDE data-driven un-
tions, because two wind farms are scheduled to begin their operations certainty set, respectively. Note that the major reductions come from
this year. Notably, wind may become the major source of electricity for nuclear power for all the deterministic and robust solutions, because of
both deterministic and robust optimization solutions, mainly due to the the scheduled decommissioning for large nuclear plants and relatively
relatively high capacity factor (over 20%) compared to other variable short lifetime compared with hydropower plants. For instance, two nu-
renewable power sources. The optimization result showing that wind clear plants are scheduled to be deactivated in 2021, and the impacts
would become a major electricity energy source is consistent with pre- of this scheduled decommissioning can be observed from Fig. 3 for all
vious research works [63]. The on-land wind power would gradually in- cases.
crease during the planning horizon, while the offshore wind generation In terms of fossil-based electricity, the natural gas combined cycle
gradually increases during the first decade of the renewable electricity would continue to be the dominant technology in the first decade of
transition, and it may have noticeable increases in the last two years of the electricity transition process for all cases. It is worth noting that the
the planning horizon to balance the electricity demand shortages owing electricity generation from natural gas combined cycle increases slightly
to the decrease of fossil-based electricity. during the first four years of the planning horizon for both deterministic
As for solar power, distributed solar PVs account for a small per- and robust optimization solutions, indicating that natural gas plays an
centage of electricity generation by the end of the planning horizon for important role during the early stage of power systems decarbonization.
all deterministic and robust solutions, while the utility solar PVs have This is consistent with the conclusion of a report from the International
a much smaller market share compared to distributed solar PVs under Energy Agency (IEA), which emphasizes that replacing more polluting
all cases. The relatively high generation level of distributed solar PVs fossil fuels with natural gas could provide near-term benefits from a re-
is probably because of the deployment requirements from the govern- newable electricity transition perspective [2]. The electricity generated
ment that mandates the capacity of distributed solar PV to be higher from natural gas combined cycle under the deterministic case would
than 6 GW by 2025, and no such requirements exist for utility solar PVs gradually decrease from the fifth year of the planning horizon and ex-
in New York State. On the other hand, the market availability for solar tinct exactly by 2035. On the other hand, the generation from natural
PVs is much less than that of wind turbines as presented in Fig. 2, which gas combined cycle would mainly be replaced by offshore wind power
suggests that it may benefit the application of solar power by increas- one year before the end of the planning horizon for all solutions under
ing the market availability of solar PVs through increasing importation uncertainty. Notably, although the natural gas combined cycle with car-
or domestic manufacturing. Hydro (dam) and nuclear power currently bon capture and storage (CCS) is not selected in the deterministic solu-
provide around half of the annual power generation in the state [57], tion, this technology may contribute to electricity generation for robust

10
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Fig. 4. Annual power system cost by source


during the power systems’ renewable transition
for the deterministic planning solution and the
robust optimization solutions under multiple
types of uncertainty sets.

optimization cases with any type of uncertainty set. This suggests that than the peak of 2034 for all uncertainty sets, and this is consistent with
CCS may benefit natural gas-based generation technologies when devel- the installed capacities of offshore wind turbines as shown in Fig. 3(b)–
oping more conservative power systems’ renewable transition plans. (e), which indicate that the installation amount in 2033 is lower than
that of 2034 for offshore wind for all robust optimization solutions. Cost
4.3. Economic costs during power systems’ renewable transition peaks for distributed solar PV exist in 2024 and 2025 for all determin-
istic and robust optimization solutions in Fig. 4, which corresponds to
The annual power system cost by source during the renewable elec- the noticeable increases of distributed solar PV in these two years in
tricity transition under all deterministic and robust optimization solu- Fig. 3. Such capacity increases are likely owing to the deployment re-
tions with multiple uncertainty sets is shown in Fig. 4. The annual cost quirement of distributed solar PV by the New York State government,
of a power generation technology consists of the capital cost of new namely the capacity of distributed solar PVs should be 6 GW or higher
power generation units, fixed O&M cost, variable O&M cost, and de- by 2025. The cost peaks regarding the installation of variable renew-
commissioning cost. Furthermore, the costs of electricity storage and able electricity generation units are resulted from the market availabil-
power importation are also included in Fig. 4. The annual costs of the ity constraints, because the market availability of solar PVs and wind
deterministic solution tend to be lower than all robust optimization so- turbines are represented by their production in the United States, which
lutions, and this is consistent with the optimal total electricity transi- could be larger than the real case. If small-scale estimations on the mar-
tion costs that are presented in the subsection on computational opti- ket availability data are available, the capital investment on variable
mization results. For all deterministic planning and robust optimization renewable generation units would split into multiple years. Decommis-
solutions in Fig. 4, the installation and decommissioning of electricity sioning of generators is another source of cost peak in Fig. 4. For in-
generators are two major sources for the cost peaks. For instance, the stance, the cost peaks for nuclear power under deterministic planning
capacities of on-land wind increase considerably in 2030 for the deter- in 2021, 2030, and 2031 in Fig. 4(a) correspond to the nuclear capacity
ministic planning solution as presented in Fig. 3(a), and correspondingly reductions in these years presented in Fig. 3(a). Similarly, cost peaks
the economic cost of on-land wind power in 2030 is noticeably higher for nuclear power are observed in 2021, 2029, 2030, and 2035 for all
than its costs in 2031. Considering the similar generation capacities of robust optimization solutions in Fig. 4(b)–(e), which correspond to the
on-land wind in 2030 and 2031, it can be inferred that the installation generation reductions of nuclear power under all types of uncertainty set
process accounts for the peak of annual cost, and this is probably be- in Fig. 3(b)–(e). Notably, the annual electricity generation cost by the
cause the capital costs of on-land wind turbines are much higher than end of the planning horizon tends to be considerably lower than that
its fixed and variable O&M costs. For the robust optimization solutions, of the beginning. A potential reason is the relatively low operating cost
two cost peaks for offshore wind can be observed for the years 2033 of renewable generation technologies compared with the conventional
and 2034 under any type of uncertainty set. The peak in 2033 is lower fossil-based generation technologies. Furthermore, the cost of electric-

11
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Fig. 6. Annual greenhouse gas emissions from electricity generation for the de-
terministic planning solution and the robust optimization solutions under mul-
tiple types of uncertainty sets.
Fig. 5. Accumulated total renewable electricity transition cost of the determin-
istic planning solution and the robust optimization solutions under multiple
types of uncertainty sets, normalized based on the optimization results under for the deterministic case, considering that deterministic planning does
the box uncertainty set. not need to accommodate the potential realizations of uncertain annual
electricity demands. On the other hand, the total greenhouse gas emis-
sions for robust optimization solutions under the budget uncertainty set,
ity importation takes a small proportion of the total annual cost under
the DPMM uncertainty set, and the PCA & KDE uncertainty set are 6%,
both deterministic and robust optimization solutions, and the importa-
2%, and 3% higher than the emissions under the box uncertainty set, re-
tion cost tends to be consistent with the amounts of imported electricity
spectively. This is attributed to the relatively low electricity generation
shown in Fig. 3. Additionally, the cost associated with the electricity
using natural gas for the optimal electricity transition pathway under
storage units can be considerably affected by the installation amounts
the box uncertainty set compared to other types of uncertainty sets, con-
of new storage units, because of the relatively high capital cost of new
sidering that natural gas would be the dominant fossil fuel for electric-
electricity storage units compared with their O&M costs.
ity generation during power system decarbonization under all solutions.
The accumulated total transition costs through the renewable elec-
Notably, the case under the budget uncertainty set has more greenhouse
tricity transition process for all deterministic planning and robust opti-
gas emissions at the early stage of the transition, and this is mainly due
mization solutions are shown in Fig. 5, and the accumulated total costs
to its large feasible range of the uncertain electricity generation demand
are normalized based on the optimal results under the box uncertainty
compared with other types of uncertainty set, as shown in Fig. 2. The po-
set. Notably, the accumulated total costs for the budget uncertainty set
tentially large generation demand requires more fossil-based electricity
are considerably higher than those of the deterministic solution and
to compensate, considering that fossil-based power could be more eco-
the robust optimization solutions under other types of uncertainty set
nomically favorable at the early stage of energy transition. For the data-
throughout the entire planning horizon. On the other hand, the deter-
driven approaches, the DPMM uncertainty set leads to less emissions at
ministic case shows notably lower accumulated total electricity transi-
the early transition stage than the PCA & KDE uncertainty set, mainly
tion cost in any time period within the planning horizon, compared to
due to the early generation reduction of the natural gas steam turbine
the costs from all robust optimization solutions. This is because the de-
for the DPMM case, as presented in Fig. 3(d). Note that the emissions at
terministic solution would not accommodate the potential realizations
the end of the planning horizon become zero for both the deterministic
of systems uncertainties. In other words, the deterministic solution is
planning case and the cases under uncertainty, indicating the achieve-
less conservative than the robust optimization solutions. Considering
ment of the decarbonization target for New York State’s power systems.
that the objective of the renewable electricity transition planning is to
Furthermore, the annual greenhouse gas emissions would increase at the
minimize total electricity transition cost, it is expected that less conser-
early stage of the transition for all cases, which is consistent with the
vative solutions tend to have lower total systems costs. Moreover, the
increases of electricity generation from natural gas that are presented
data-driven uncertainty sets, including the DPMM uncertainty set and
in Fig. 3. Furthermore, the reduction of greenhouse gas emissions in the
the PCA & KDE uncertainty set, show consistently lower total accumu-
last 3 years is mainly attributed to the large market availability of vari-
lated systems costs during the planning horizon than the conventional
able renewable generation units, which is represented by their produc-
uncertainty sets, such as the box uncertainty set and the budget uncer-
tion amounts in the United States. If smaller and more accurate market
tainty set. This indicates that the data-driven uncertainty sets could lead
availability estimations are available, the reduction of greenhouse gas
to less conservative electricity transition pathways for any stage within
emissions would be more steady. Additionally, the differences in annual
the planning horizon while handling electric power systems’ uncertain-
greenhouse gas emissions for different types of uncertainty sets at the
ties.
early stages of the renewable electricity transition are much larger than
their differences at the late stages.
4.4. Greenhouse gas emissions

Fig. 6 presents the annual total greenhouse gas emissions in CO2 4.5. Electricity storage
equivalent (CO2 -eq) across the planning horizon for the determinis-
tic planning solution and the robust optimization solutions. The over- The electricity storage capacities in each time period within the plan-
all greenhouse gas emissions through the electricity transition process ning horizon for all deterministic and robust solutions are shown in
for the deterministic solution are 11% lower than the total emissions Fig. 7. The renewable electricity transition pathway under the budget
for the robust optimization solution with the box uncertainty set. This uncertainty set has the largest electricity capacity at the end of the plan-
is because of the relatively low annual electricity generation amounts ning horizon, because it has the largest generation capacity from vari-

12
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

conservation of conventional uncertainty sets, machine learning tech-


niques were applied to construct data-driven uncertainty sets, and such
machine learning techniques consisted of a variational inference algo-
rithm for Dirichlet process mixture model, principal component analy-
sis, and kernel density estimation.
To illustrate the applicability of the proposed framework, a case
study for the power systems’ renewable transition under uncertainty in
New York State was presented. Five cases were investigated based on
deterministic planning, robust optimization under conventional uncer-
tainty sets, and robust optimization under data-driven uncertainty sets.
The total electricity transition costs for the robust optimization solu-
tions were 21%−42% higher than deterministic planning. On the other
hand, the total renewable electricity transition costs for the data-driven
uncertainty sets were 2%−17% lower than that of the conventional un-
certainty sets, indicating that data-driven uncertainty sets constructed
using machine learning techniques could achieve less conservative solu-
tions compared to the conventional uncertainty sets. In terms of electric-
Fig. 7. Electricity storage capacity for the deterministic planning solution and ity generation, the total electricity generated for the robust optimization
the robust optimization solutions under multiple types of uncertainty sets. cases were 12%−23% higher than the deterministic planning case. On-
land wind would be a major source of electricity for New York State in
2035 under the deterministic case, while the offshore wind would gen-
able renewable energy, such as solar and wind, as presented in Fig. 3(c). erate the most electricity by the end of the planning horizon for all solu-
The deterministic solution has larger electricity storage capacities than tions under uncertainty. Distributed solar PVs would account for a small
most robust optimization solutions, and this is because the on-land wind percentage of electricity generation by 2035 for all deterministic and ro-
is a major power source for the deterministic case, while offshore wind bust cases, and utility solar photovoltaics would have a much smaller
contributes the most for electricity generation under the robust opti- market share compared to distributed solar photovoltaics. Natural gas
mization solutions. Specifically, on-land wind turbines tend to have rel- combined cycle would be the dominant fossil-based generation technol-
atively low capacity factors compared to offshore wind, so larger elec- ogy in the first decade for all cases, suggesting that natural gas would
tricity storage capacities are needed for a unit amount of electricity gen- play an important role at the early stage of the electricity transition. Fur-
eration from on-land wind turbines than offshore wind turbines. Fur- thermore, the economic analysis indicated that the installation and de-
thermore, the electricity storage capacities under the data-driven un- activation of power generators and electricity storage units would have
certainty sets are lower than the deterministic planning solution and considerable impacts on annual systems cost and should be accountable
the robust optimization solutions under conventional uncertainty sets. for the cost peaks. As for greenhouse gas emissions, the robust optimiza-
This is mainly because the cases using the data-driven approach have tion solutions would have 5%−11% more emissions compared to deter-
relatively low generation capacity using variable renewable energy, in- ministic planning. In terms of electricity storage, the storage capacities
cluding solar and wind, as presented in Fig. 3, considering that the elec- in 2035 for the deterministic case would be higher than the storage ca-
tricity storage capacities are directly related to the variable renewable pacities under all types of uncertainty sets, due to its large percentage of
generation capacities in the proposed framework. on-land wind power that had relatively low capacity factors. Addition-
In addition, the rapid increases of electricity storage capacity are ally, the rapid increases of storage capacities would tend to correspond
likely to associate with the installation peaks of generation capacities to installation peaks of generation capacities using variable renewable
using variable renewable energy for all cases. For example, the increase energy.
of storage capacities in 2030 for the deterministic planning solution cor- There are limitations for the proposed energy transition multistage
responds to the capacity increase of on-land wind power this year. adaptive robust optimization framework. For instance, electricity expor-
tation could be considered if the framework is applied to other regions
5. Conclusion that mainly export electricity to their neighbors annually. In terms of
future research works, more types of uncertainties could be included,
In this work, a bottom-up data-driven multistage adaptive robust such as the uncertain electricity importation price, which could be af-
optimization framework was proposed to provide potential renewable fected by the carbon-neutral power systems transition of the exporting
transition pathways under uncertainty for the electricity sector. The un- states or countries. It would also be helpful and important for the policy-
certainties involved in the framework included the electricity generation making processing to investigate the trade-off between economic costs
demands, the market availability of wind turbines, and the market avail- and greenhouse gas emissions during the transition, by revising the pro-
ability of solar photovoltaics. Furthermore, the proposed framework en- posed framework into a multi-objective multi-stage adaptive robust op-
sured that the system reliability for the electricity sector was maintained timization model for power systems transition. The minimum storage
during the renewable electricity transition process. Specifically, the elec- capacity determination is not precise enough in the work, and more so-
tricity generation demand and supply were balanced for all time periods, phisticated assumptions could be applied in the future. Furthermore, if
and the multistage adaptive robust optimization framework considered reliable hourly energy profiles and short-term uncertainty distribution
mandatory electricity storage requirements and increasing technolog- are available, conducting short-term hourly simulation would be an-
ical diversity as measures to manage the intermittency issue resulted other future research orientation that could assess the adequacy of the
from variable renewable energy. Additionally, the framework also ac- deeply decarbonized future power systems.
commodated climate targets, renewable electricity generation targets,
and scheduled systems changes, while handling potential uncertainties
involved in the electric power systems. To address the computationally Declaration of Competing Interest
intractable issue of the multistage renewable electricity transition adap-
tive robust optimization, affine decision rules were applied to reformu- The authors declare that they have no known competing financial
late the original framework into its robust counterpart, which could be interests or personal relationships that could have appeared to influence
solved directly using off-the-shelf optimization solvers. To avoid over- the work reported in this paper.

13
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

Acknowledgments Similarly, the affine decision rules for retired electricity storage ca-
pacity in each year are presented by constraints (A9) and (A10), where
The authors acknowledge financial support from the National Sci- bdisi,t,y denotes the retired power storage capacities using technology i
ence Foundation (NSF) under grant number CBET-1643244. with an age of t in year y. 𝜇 i,t,y,k,n and 𝜇 0 i,t,y are also “here-and-now”
decision variables that stand for the affine decision rule coefficients for
Appendix: Robust counterpart reformulation of data-driven the retirement of electricity storage units.
MARO model for renewable electricity transition under ∑ ∑
𝑏𝑑𝑖𝑠𝑖,𝑡,𝑦 = 𝜇𝑖,𝑡,𝑦,𝑘,𝑛 ⋅ 𝑢𝑘,𝑛 + 𝜇𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 , 𝑢𝑘,𝑛 ∈ 𝑈𝑘
uncertainty 𝑘∈𝐾 𝑛∈𝑁
(A9)
In this appendix, both the objective function and the constraints
shown in §3 are reformulated using affine decision rules. The affine 𝜇𝑖,𝑡,𝑦,𝑘,𝑛 = 0, ∀𝑛 > 𝑦 (A10)
functions for the decision variables are presented in (A1)-(A10).
A.2. Reformulation of objective function
A.1. Affine decision rules for decision variables in the MARO model
To reformulate the objective function (4), we plug the affine decision
rules(A1), (A3), (A5), (A7), and (A9) into (4) and obtain:
The affine functions for electricity generation demand from each
∑ ∑ ∑
technology of each age in each time period can be represented by con- ⎛ ⎧ ∑𝑖∈𝐼𝑒 ∑𝑡∈𝑇 ∑𝑦∈𝑌 −{YINI} CE𝐸𝑖,𝑡,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 ⎫ ⎞
straints (A1) and (A2). I represents the set of all technologies electricity ⎜ ⎪ + 𝑖∈𝐼 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CE𝐷𝑖,𝑡,𝑦 ⋅ 𝜂𝑖,𝑡,𝑦,𝑘,𝑛 ⎪ ⎟
⎜∑ ∑ ⎪ ∑𝑒 ∑ ⎪ ⎟
∈𝑌 −{YINI} CIM𝑃𝑠,𝑦 ⋅ 𝜌𝑠,𝑦,𝑘,𝑛 ⎬ ⋅ 𝑢𝑘,𝑛 ⎟
generation technologies (Ie ) and power storage technologies (Is ), T is ⎜ ⎨ ∑+ 𝑠∈𝑆 𝑦∑
⎜ 𝑘 𝑛 ⎪+ ∑
a set of ages for power generators and electricity storage units, and Y CB𝐸𝑖,𝑡,𝑦 ⋅ 𝛿𝑖,𝑡,𝑦,𝑘,𝑛 ⎪ ⎟
denotes the set of years within the planning horizon. It is worth noting ⎜ ⎪ ∑ 𝑖∈𝐼𝑠 ∑ 𝑡∈𝑇 ∑ 𝑦∈𝑌 −{YINI} ⎪ ⎟
⎝ ⎩+ 𝑖∈𝐼 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CB𝐷𝑖,𝑡,𝑦 ⋅ 𝜇𝑖,𝑡,𝑦,𝑘,𝑛 ⎭ ⎠
𝑠
that set Y and set N represent the same set, and different symbols are ∑ ∑ ∑
used because the value of a variable in a certain year could be adjusted ⎛𝑧 − 𝑖∈𝐼 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CE𝐸𝑖,𝑡,𝑦 ⋅ 𝛼 0 ⎞
𝑖,𝑡,𝑦
⎜ ∑ 𝑒∑ ∑ ⎟
following the uncertainty realizations from previous years. This can be ⎜ − 𝑖∈𝐼𝑒 CE 𝐷 𝑖,𝑡,𝑦 ⋅ 𝜂 0

⎜ ∑ 𝑡∈𝑇 ∑ 𝑦∈𝑌 −{YINI} 𝑖,𝑡,𝑦
⎟, ∀𝑢 ∈ 𝑈
≤ ∈𝑌 −{YINI} CIM𝑃𝑠,𝑦 ⋅ 𝜌𝑠,𝑦
− 0
further illustrated by constraint (A2), which specifies that the electricity (A11)
⎜ ∑ 𝑠∈𝑆 ∑ 𝑦∑ ⎟
⎜ − 𝑖∈𝐼𝑠 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CB𝐸𝑖,𝑡,𝑦 ⋅ 𝛿𝑖,𝑡,𝑦 ⎟
0
generation decision cannot be affected by future realizations of uncer-
⎜ −∑ ∑ ∑ ⎟
𝑦∈𝑌 −{YINI} CB𝐷𝑖,𝑡,𝑦 ⋅ 𝜇𝑖,𝑡,𝑦 ⎠
tainties. 0
⎝ 𝑖∈𝐼𝑠 𝑡∈𝑇
∑ ∑
𝑒𝑒𝑥𝑡𝑖,𝑡,𝑦 = 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 ⋅ 𝑢𝑘,𝑛 + 𝛼𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 , 𝑢𝑘,𝑛 ∈ 𝑈𝑘 It is worth noting that constraint (A11) should hold for all potential
𝑘∈𝐾 𝑛∈𝑁 realizations of uncertainty. Considering that both the uncertainty set
(A1) formulation given by Eq. (2) and the left-hand side of constraint (A11)
has a linear form for the uncertain parameter uk,n . Therefore, by taking
𝛼𝑖,𝑡,𝑦,𝑘,𝑛 = 0, ∀𝑛 > 𝑦 (A2) the dual of left-hand side maximization, subject to the formulation of
uncertainty sets, constraint (4) can be equivalently reformulated into
where eexti,t,y denotes the electricity generation from technology i ex-
constraints (A12)–(A14), where DVl,n are dual variables corresponding
isting in year y with an age of t. 𝛼 i,t,y,k,n and 𝛼 0 i,t,y are the coefficients
to the constraints of uncertainty sets from different stages that are shown
of the affine decision rule for the decision on electricity generation, and
in Eq. (2) to reformulate constraint (A11).
they are considered as “here-and-now” decision variables. ∑∑
The affine functions for electricity storage capacity from different 𝜔𝑙,𝑛 ⋅ 𝐷𝑉𝑙,𝑛
𝑙 𝑛
technologies, ages, and time periods are represented by constraints (A3) ∑ ∑ ∑ ∑ ∑ ∑
⎛𝑧 − 𝑖∈𝐼 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CE𝐸𝑖,𝑡,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦
0
− 𝑖∈𝐼𝑒 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CE𝐷𝑖,𝑡,𝑦 ⋅ 𝜂𝑖,𝑡,𝑦
0 ⎞
and (A4). bexti,t,y stands for the electricity storage capacity of technology ⎜ 𝑒

⎜ ∑ ∑ ∑ ∑ ∑ ⎟
i existing in year y with an age of t, while 𝛿 i,t,y,k,n and 𝛿 0 i,t,y are the affine ≤⎜ − 𝑠∈𝑆 𝑦∈𝑌 −{YINI} CIM𝑃𝑠,𝑦 ⋅ 𝜌0𝑠,𝑦 − 𝑖∈𝐼𝑠 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CB𝐸𝑖,𝑡,𝑦 ⋅ 𝛿𝑖,𝑡,𝑦
0

decision rule coefficients for the decision on existing electricity storage ⎜ ⎟
⎜ ∑ ∑ ∑ ⎟
⎝ − 𝑖∈𝐼𝑠 𝑡∈𝑇 𝑦∈𝑌 −{YINI} CB𝐷𝑖,𝑡,𝑦 ⋅ 𝜇𝑖,𝑡,𝑦
0

capacities.
∑ ∑ (A12)
𝑏𝑒𝑥𝑡𝑖,𝑡,𝑦 = 𝛿𝑖,𝑡,𝑦,𝑘,𝑛 ⋅ 𝑢𝑘,𝑛 + 𝛿𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 , 𝑢𝑘,𝑛 ∈ 𝑈𝑘 (A3)
𝑘∈𝐾 𝑛∈𝑁 ∑
𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉𝑙,𝑛
𝑙
𝛿𝑖,𝑡,𝑦,𝑘,𝑛 = 0, ∀𝑛 > 𝑦 (A4) ∑ ∑ ∑ ∑ ∑ ∑
⎛ 𝐶𝐸 𝐸𝑖,𝑡,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 + 𝐶𝐸 𝐷𝑖,𝑡,𝑦 ⋅ 𝜂𝑖,𝑡,𝑦,𝑘,𝑛 ⎞
⎜𝑖∈𝐼∑
𝑒 𝑡∈𝑇 𝑦∈𝑌 −{𝑌 𝐼 𝑁 𝐼 }
∑ ∑
𝑖∈𝐼𝑒 𝑡∈𝑇 𝑦∈𝑌 −{𝑌 𝐼 𝑁 𝐼 }
∑ ∑ ⎟
The annual electricity importation can be expressed using affine de- ⎜ ⎟
cision rules as shown in constraints (A5) and (A6), where eimps,y stands ≥ ⎜+ 𝑠∈𝑆 𝑦∈𝑌 −{𝑌 𝐼 𝑁 𝐼 } 𝐶𝐼 𝑀 𝑃𝑠,𝑦 ⋅ 𝜌𝑠,𝑦,𝑘,𝑛 + 𝑖∈𝐼 𝑡∈𝑇 𝑦∈𝑌 −{𝑌 𝐼 𝑁 𝐼 } 𝐶𝐵 𝐸𝑖,𝑡,𝑦 ⋅ 𝛿𝑖,𝑡,𝑦,𝑘,𝑛 ⎟, ∀𝑘, 𝑛
𝑠
⎜ ∑ ∑ ∑ ⎟
for the annual electricity importation from source s in year y. 𝜌s,y,k,n and ⎜+ 𝐶𝐵 𝐷𝑖,𝑡,𝑦 ⋅ 𝜇𝑖,𝑡,𝑦,𝑘,𝑛 ⎟
⎝ 𝑖∈𝐼𝑠 𝑡∈𝑇 𝑦∈𝑌 −{𝑌 𝐼 𝑁 𝐼 } ⎠
𝜌0 s,y are “here-and-now” decision variables, and they are the coefficients
of the affine function on annual electricity importation. (A13)
∑ ∑
𝑒𝑖𝑚𝑝𝑠,𝑦 = 𝜌𝑠,𝑦,𝑘,𝑛 ⋅ 𝑑𝑘,𝑛 + 𝜌0𝑠,𝑦 , ∀𝑠 ∈ 𝑆, 𝑦 ∈ 𝑌 , 𝑢𝑘,𝑛 ∈ 𝑈𝑘 (A5) 𝐷𝑉𝑙,𝑛 ≥ 0, ∀𝑙, 𝑛 (A14)
𝑘∈𝐾 𝑛∈𝑁
𝜌𝑠,𝑦,𝑘,𝑛 = 0, ∀𝑛 > 𝑦 (A6) A.3. Reformulation of electricity generation demand and electricity storage
requirement constraints
Constraints (A7) and (A8) represent the affine decision rules for the
decommissioning of power generation capacity each year. edisi,t,y is the The electricity generation demand constraint (5) would become con-
retired generation capacities using technology i with an age of t in year straint (A15) by plugging in the affine decision rules (A1), (A3), (A5),
y, while 𝜂 i,t,y,k,n and 𝜂 0 i,t,y are the corresponding coefficients based on (A7), and (A9).
the affine decision rule. ( )
∑∑ ∑ ∑ ∑
∑ ∑ −𝜌𝑠,𝑦,𝑘,𝑛 + −𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 ⋅ 𝑢𝑘,𝑛 + 𝑢𝑘=1,𝑛=𝑦
𝑒𝑑𝑖𝑠𝑖,𝑡,𝑦 = 𝜂𝑖,𝑡,𝑦,𝑘,𝑛 ⋅ 𝑢𝑘,𝑛 + 𝜂𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 , 𝑢𝑘,𝑛 ∈ 𝑈𝑘 (A7) 𝑘 𝑛 𝑠∈𝑆 𝑖∈𝐼𝑒 𝑡∈𝑇
( ) (A15)
𝑘∈𝐾 𝑛∈𝑁
∑ 0 ∑ ∑
≤ 𝜌𝑠,𝑦 + 𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦
0 , ∀𝑦 ∈ 𝑌 , 𝑢 ∈ 𝑈
𝜂𝑖,𝑡,𝑦,𝑘,𝑛 = 0, ∀𝑛 > 𝑦 (A8) 𝑠∈𝑆 𝑖∈𝐼𝑒 𝑡∈𝑇

14
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

To derive the robust counterpart based on constraint (A15), dual Constraint (9) that stands for the deployment targets of renew-
variables ZVAl,n,y are introduced corresponding to the uncertainty set able generation technologies can be equivalently reformulated into con-
constraints given by Eq. (2) and constraint (A15) for different time pe- straints (A30)-(A32) based on affine decision rules, where DVIi,l,n,y rep-
riod y. By taking the dual of the left-hand side maximization of con- resents the dual variables introduced corresponding to the constraints
straint (A15), the reformulation of constraint (5) is presented in con- for uncertainty set given by Eq. (2) for renewable power generation
straints (A16)–(A19). technology i in year y.
∑∑ ∑ ∑∑ ∑∑ ∑
𝜔𝑙,𝑛 ⋅ 𝑍𝑉 𝐴𝑙,𝑛,𝑦 ≤ 𝜌0𝑠,𝑦 + 𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦
0
, ∀𝑦 ∈ 𝑌 (A16) 𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝐼𝑖,𝑙,𝑛,𝑦 ≤ 𝛼𝑖,𝑡,𝑦
0
− 𝑇 𝐴𝑅𝑖,𝑦 ∀𝑖, 𝑦 (A30)
𝑙 𝑛 𝑠∈𝑆 𝑖∈𝐼𝑒 𝑡∈𝑇 𝑙 𝑛
∑ ∑𝑡∈𝑇
( ) 𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝐼𝑖,𝑙,𝑛,𝑦 ≥ −𝛼𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑖, 𝑦, 𝑘, 𝑛 (A31)
∑ ∑ ∑∑ 𝑙 𝑡∈𝑇
𝜃𝑙,𝑘,𝑛 ⋅ 𝑍𝑉 𝐴𝑙,𝑛,𝑦 ≥ −𝜌𝑠,𝑦,𝑘,𝑛 + −𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 ,
𝑙 𝑠∈𝑆 𝑖∈𝐼𝑒 𝑡∈𝑇 𝐷𝑉 𝐼𝑖,𝑙,𝑛,𝑦 ≥ 0, ∀𝑖, 𝑙, 𝑛, 𝑦 (A32)
∀{𝑦, 𝑛, 𝑘} − {𝑘 = 1, 𝑛 = 𝑦} (A17) The reformulation of electricity storage deployment constraint (10)
is presented in constraints (A33)-(A35), by using affine decision rules
∑ and deriving the dual problem. DVKl,n,y represents the dual variables
𝜃𝑙,𝑘=1,𝑛=𝑦 ⋅ 𝑍𝑉 𝐴𝑙,𝑛=𝑦,𝑦
that are introduced corresponding to the uncertainty set constraints for
𝑙
( ) year y.
∑ ∑∑ ∑∑ ∑∑
≥ 1+ −𝜌𝑠,𝑦,𝑘=1,𝑛=𝑦 + −𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘=1,𝑛=𝑦 , ∀𝑦 ∈ 𝑌 (A18) 𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝐾𝑙,𝑛,𝑦 ≤ 𝛿𝑖,𝑡,𝑦
0
− 𝑇 𝐴𝑅𝑆𝑦 , 𝑦 ∈ 𝑌 (A33)
𝑠∈𝑆 𝑖∈𝐼𝑒 𝑡∈𝑇 𝑙 𝑛 𝑖∈𝐼 𝑡∈𝑇
𝑠
∑ ∑∑
𝑍𝑉 𝐴𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 (A19) 𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝐾𝑙,𝑛,𝑦 ≥ −𝛿𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑘, 𝑛, 𝑦 ∈ 𝑌 (A34)
𝑙 𝑖∈𝐼𝑠 𝑡∈𝑇
In terms of the electricity storage requirement constraint, by plug-
𝐷𝑉 𝐾𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 ∈ 𝑌 (A35)
ging the affine decision rules in constraint (6) and taking the dual, the
reformulation can be derived as constraints (A20)-(A22), where DVEl,n,y
denotes the dual variables following the uncertainty set constraints in A.5. Reformulation of transition relationship constraints
Eq. (2) for stage y.
By plugging the affine decision rules (A1), (A3), (A5), (A7), and (A9)
∑∑ ∑∑ ∑∑
𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝐸𝑙,𝑛,𝑦 ≤ 𝛿𝑖,𝑡,𝑦
0
− 𝑅𝐸𝑄𝑀𝑖 ⋅ 𝛼𝑖,𝑡,𝑦
0
, ∀𝑦 ∈ 𝑌 into constraints (11)-(12) that represent initialization of variables re-
𝑙 𝑛 𝑖∈𝐼𝑠 𝑡∈𝑇 𝑖∈𝐼𝑒𝑖 𝑡∈𝑇 garding electricity generation and storage capacities, the two constraints
(A20) can be reformulated into constraints (A36)-(A39).

𝛼𝑖,𝑡,𝑦=𝑌 𝐼 𝑁 𝐼 ,𝑘,𝑛 = 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑡, 𝑘, 𝑛 (A36)


∑ ∑∑ ∑∑
𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝐸𝑙,𝑛,𝑦 ≥ 𝑅𝐸𝑄𝑀𝑖 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 − 𝛿𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑦, 𝑘, 𝑛
𝑙 𝑖∈𝐼𝑒𝑖 𝑡∈𝑇 𝑖∈𝐼𝑠 𝑡∈𝑇 𝛿𝑖,𝑡,𝑦=𝑌 𝐼 𝑁 𝐼 ,𝑘,𝑛 = 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑡, 𝑘, 𝑛 (A37)
(A21)
𝛼𝑖,𝑡,𝑦
0
=𝑌 𝐼 𝑁 𝐼 = 𝐼 𝑁 𝐼 𝐸𝑖,𝑡 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 (A38)
𝐷𝑉 𝐸𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 (A22)
𝛿𝑖,𝑡,𝑦
0
=𝑌 𝐼 𝑁 𝐼 = 𝐼 𝑁 𝐼 𝐵𝑖,𝑡 , ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 (A39)
A.4. Reformulation of target constraints on renewable electricity generation By applying the affine decision rules to the transition relationship
and emissions constraints, the reformulation of capacity transition relationship con-
straints (13) and (14) can be derived as eqs. (A40)-(A43).
Constraints (A23)-(A25) are the reformulation of the annual green-
house gas emissions target constraint (7), by plugging in the affine de- 𝛼𝑖,𝑡−1,𝑦−1,𝑘,𝑛 − 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 − 𝜂𝑖,𝑡,𝑦,𝑘,𝑛 = 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ≥ 1, 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛
cision rules and deriving the dual problem. DVFl,n,y represents the dual (A40)
variables that are introduced corresponding to the constraints for un-
certainty set given by Eq. (2) for stage y.
∑∑ ∑∑ 𝛿𝑖,𝑡−1,𝑦−1,𝑘,𝑛 − 𝛿𝑖,𝑡,𝑦,𝑘,𝑛 − 𝜇𝑖,𝑡,𝑦,𝑘,𝑛 = 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ≥ 1, 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛
𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝐹𝑙,𝑛,𝑦 ≤ 𝐸 𝐿𝑀 𝑇𝑦 − 𝐸 𝐸𝑖,𝑡,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦
0
, ∀𝑦 ∈ 𝑌 (A23)
𝑙 𝑛 𝑖∈𝐼𝑒 𝑡∈𝑇 (A41)
∑ ∑∑
𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝐹𝑙,𝑛,𝑦 ≥ 𝐸 𝐸𝑖,𝑡,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑦, 𝑘, 𝑛 (A24)
𝑙 𝑖∈𝐼𝑒 𝑡∈𝑇
𝜂𝑖,𝑡,𝑦
0
+ 𝛼𝑖,𝑡,𝑦
0
− 𝛼𝑖,𝑡
0
−1,𝑦−1
= 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ≥ 1, 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A42)

𝐷𝑉 𝐹𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 (A25)


𝜇𝑖,𝑡,𝑦
0
+ 𝛿𝑖,𝑡,𝑦
0
− 𝛿𝑖,𝑡
0
−1,𝑦−1
= 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ≥ 1, 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A43)
After deriving the robust counterpart, the reformulation of the re-
The retirement constraints (15) and (16) can be reformulated into
newable electricity generation target constraint (8) is shown in con-
eqs. (A44)-(A47), following the affine functions of variables in eqs. (A1),
straints (A26)-(A29), where DVHl,n,y stands for the dual variables based
(A3), (A5), (A7), and (A9).
on the uncertainty set constraints in Eq. (2) for year y.
∑∑ ∑∑
𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝐻𝑙,𝑛,𝑦 ≤ 𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦
0
, ∀𝑦 (A26) 𝛼𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 −1,𝑦−1,𝑘,𝑛 − 𝜂𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 ,𝑦,𝑘,𝑛 = 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛
𝑙 𝑛 𝑖∈𝐼𝑒𝑟 𝑡∈𝑇 (A44)
∑ ∑∑
𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝐻𝑙,𝑛,𝑦 ≥ −𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑘, 𝑛, 𝑦 − {𝑘 = 1, 𝑛 = 𝑦}
𝑙 𝑖∈𝐼𝑒𝑟 𝑡∈𝑇 𝛿𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 −1,𝑦−1,𝑘,𝑛 − 𝜇𝑖,𝑡=𝐿𝐼𝐹 𝐸𝑖 ,𝑦,𝑘,𝑛 = 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛
(A27) (A45)
∑ ∑∑
𝜃𝑙,𝑘=1,𝑛=𝑦 ⋅ 𝐷𝑉 𝐻𝑙,𝑛=𝑦,𝑦 ≥ 𝑅𝑃 𝑆𝐺 + −𝐹𝑖,𝑦 ⋅ 𝛼𝑖,𝑡,𝑦,𝑘=1,𝑛=𝑦 , ∀𝑦 (A28) 𝜂𝑖,𝑡 =𝐿𝐼𝐹 𝐸 ,𝑦 − 𝛼𝑖,𝑡=𝐿𝐼𝐹 𝐸 −1,𝑦−1 = 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }
0 0
(A46)
𝑖 𝑖
𝑙 𝑖∈𝐼𝑒𝑟 𝑡∈𝑇

𝐷𝑉 𝐻𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 (A29) 𝜇𝑖,𝑡


0
=𝐿𝐼𝐹 𝐸 ,𝑦 − 𝛿𝑖,𝑡=𝐿𝐼𝐹 𝐸 −1,𝑦−1 = 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }
0
(A47)
𝑖 𝑖

15
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

A.6. Reformulation of scheduling constraints ∑


𝜃𝑙,𝑘=2,𝑛=𝑦 ⋅ 𝐸 𝑉 𝐸 𝐵𝑙,𝑛=𝑦,𝑦
The reformulation of scheduled generation addition constraints (17) 𝑙
∑ ∑( )
is presented in constraints (A48)-(A50), by using the affine decision rules ≥ 𝛼𝑖,𝑡,𝑦,𝑘=2,𝑛=𝑦 − 𝛼𝑖,𝑡,𝑦−1,𝑘=2,𝑛=𝑦 − 1,
and deriving the dual problem. DVLl,n,i,y denotes the dual variables based 𝑖∈{𝑊 𝑂𝑁,𝑊 𝑂𝐹 𝐹 } 𝑡∈𝑇
on the uncertainty set constraints in Eq. (2) for generation technology i ∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A62)
in time period y, corresponding to constraint (17).
∑∑
𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝐿𝑙,𝑛,𝑖,𝑦 ≤ 𝛼𝑖,𝑡
0
− 𝑃 𝑅𝑂𝐺𝑖,𝑦 , ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 (A48) 𝐸 𝑉 𝐸 𝐵𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }; (A63)
=0,𝑦
∑ 𝑙 𝑛
Constraints (A64)-(A67) represent the robust counterpart for the
𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝐿𝑙,𝑛,𝑖,𝑦 ≥ −𝛼𝑖,𝑡=0,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑒 , 𝑦 ∈ 𝑌 , 𝑘, 𝑛 (A49)
𝑙 market availability constraint of solar PVs, namely constraint (22).
𝐷𝑉 𝐿𝑙,𝑛,𝑖,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑖 ∈ 𝐼𝑒 , 𝑦 (A50) EVECl,n,y represents the dual variable following the uncertainty set con-
straints in Eq. (2) for time period y.
Similarly, the scheduled electricity storage addition constraint (18) ∑∑
are reformulated as constraints (A51)-(A53), where DVMl,n,i,y represents 𝜔𝑙,𝑛 ⋅ 𝐸 𝑉 𝐸 𝐶𝑙,𝑛,𝑦
the dual variables that are introduced corresponding to the uncertainty 𝑙 𝑛
∑ ∑ ∑ ∑
set constraints for storage technology i in stage y. ≤− 𝛼𝑖,𝑡,𝑦
0
+ 𝛼𝑖,𝑡,𝑦
0
−1
,
∑∑ 𝑖∈{𝑆 𝑈 𝑁𝐷,𝑆 𝑈 𝑁𝑈 } 𝑡∈𝑇 𝑖∈{𝑆 𝑈 𝑁𝐷,𝑆 𝑈 𝑁𝑈 } 𝑡∈𝑇
𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝑀𝑙,𝑛,𝑖,𝑦 ≤ 𝛿𝑖,𝑡
0
=0,𝑦
− 𝑃 𝑅𝑂𝐵𝑖,𝑦 , ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 (A51)
𝑙 𝑛 ∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A64)

𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝑀𝑙,𝑛,𝑖,𝑦 ≥ −𝛿𝑖,𝑡=0,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 , 𝑘, 𝑛 (A52)
𝑙 ∑ ∑ ∑( )
𝐷𝑉 𝑀𝑙,𝑛,𝑖,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑖 ∈ 𝐼𝑠 , 𝑦 (A53) 𝜃𝑙,𝑘,𝑛 ⋅ 𝐸 𝑉 𝐸 𝐶𝑙,𝑛,𝑦 ≥ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 − 𝛼𝑖,𝑡,𝑦−1,𝑘,𝑛 ,
𝑙 𝑖∈{𝑆 𝑈 𝑁𝐷,𝑆 𝑈 𝑁𝑈 } 𝑡∈𝑇
By applying the affine decision rules, the scheduled decommission- ∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛 − {𝑘 = 3, 𝑛 = 𝑦} (A65)
ing constraint (19) can be reformulated into constraints (A54)-(A56).
DVNl,n,i,t,y represents the dual variables following the uncertainty set

constraints in Eq. (2) for generation technology i with an age of t in 𝜃𝑙,𝑘=3,𝑛=𝑦 ⋅ 𝐸 𝑉 𝐸 𝐶𝑙,𝑛=𝑦,𝑦
time period y. 𝑙
∑∑ ∑ ∑( )
𝜔𝑙,𝑛 ⋅ 𝐷𝑉 𝑁𝑙,𝑛,𝑖,𝑡,𝑦 ≤ 𝜂𝑖,𝑡,𝑦
0
− 𝑃 𝑅𝑂𝐷𝑖,𝑡,𝑦 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡, 𝑦 (A54) ≥ 𝛼𝑖,𝑡,𝑦,𝑘=3,𝑛=𝑦 − 𝛼𝑖,𝑡,𝑦−1,𝑘=3,𝑛=𝑦 − 1,
𝑖∈{𝑆 𝑈 𝑁𝐷,𝑆 𝑈 𝑁𝑈 } 𝑡∈𝑇
∑ 𝑙 𝑛
𝜃𝑙,𝑘,𝑛 ⋅ 𝐷𝑉 𝑁𝑙,𝑛,𝑖,𝑡,𝑦 ≥ −𝜂𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡, 𝑦, 𝑘, 𝑛 (A55) ∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A66)
𝑙
𝐷𝑉 𝑁𝑙,𝑛,𝑖,𝑡,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑖 ∈ 𝐼𝑒 , 𝑡, 𝑦 (A56) 𝐸 𝑉 𝐸 𝐶𝑙,𝑛,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }; (A67)

A.7. Reformulation of market availability constraints By applying affine decision rules to constraint (23), its reformulation
can be derived as constraints (A68)–(A70), where EVGl,n,i,y stands for
Constraint (20) that stands for the market availability for non- the dual variable based on the uncertainty set constraints in Eq. (2) for
intermittent electricity generation technologies can be reformulated into storage technology i in time period y.
constraints (A57)–(A59) using affine decision rules. EVEAl,n,i,y denotes ∑∑ ∑ ∑
the corresponding dual variable based on the uncertainty set constraints 𝜔𝑙,𝑛 ⋅ 𝐸𝑉 𝐺𝑙,𝑛,𝑖,𝑦 ≤ 𝑀𝐴𝐵𝑖,𝑦 − 𝛿𝑖,𝑡,𝑦
0
+ 𝛿𝑖,𝑡,𝑦
0
−1
,
𝑙 𝑛 𝑡∈𝑇 𝑡∈𝑇
in Eq. (2) for generation technology i in time period y.
∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A68)
∑∑ ∑ 0 ∑ 0
𝜔𝑙,𝑛 ⋅ 𝐸 𝑉 𝐸 𝐴𝑙,𝑛,𝑖,𝑦 ≤ 𝑀𝐴𝐺𝑖,𝑦 − 𝛼𝑖,𝑡,𝑦 + 𝛼𝑖,𝑡,𝑦−1 ,
𝑙 𝑛 𝑡∈𝑇 𝑡∈𝑇 (A57)
∀𝑖 ∈ 𝐼𝑒 − {𝑊 𝑂𝑁, 𝑊 𝑂𝐹 𝐹 , 𝑆𝑈 𝑁𝐷, 𝑆𝑈 𝑁𝑈 }, 𝑦 ∈ 𝑌 − {𝑌 𝐼𝑁𝐼} ∑ ∑ ∑
𝜃𝑙,𝑘,𝑛 ⋅ 𝐸𝑉 𝐺𝑙,𝑛,𝑖,𝑦 ≥ 𝛿𝑖,𝑡,𝑦,𝑘,𝑛 − 𝛿𝑖,𝑡,𝑦−1,𝑘,𝑛 ,
∑ ∑ ∑ 𝑙 𝑡∈𝑇 𝑡∈𝑇
𝜃𝑙,𝑘,𝑛 ⋅ 𝐸 𝑉 𝐸 𝐴𝑙,𝑛,𝑖,𝑦 ≥ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 − 𝛼𝑖,𝑡,𝑦−1,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛 (A69)
𝑙 𝑡∈𝑇 𝑡∈𝑇
∀𝑖 ∈ 𝐼𝑒 − {𝑊 𝑂𝑁, 𝑊 𝑂𝐹 𝐹 , 𝑆𝑈 𝑁𝐷, 𝑆𝑈 𝑁𝑈 }, 𝑦 ∈ 𝑌 − {𝑌 𝐼𝑁𝐼}, 𝑘, (A58)
𝑛
𝐸𝑉 𝐺𝑙,𝑛,𝑖,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑖 ∈ 𝐼𝑠 , 𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }; (A70)

𝐸 𝑉 𝐸 𝐴𝑙,𝑛,𝑖,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑖 ∈ 𝐼𝑒 A.8. Reformulation of electricity importation and non-negativity constraints


−{𝑊 𝑂𝑁, 𝑊 𝑂𝐹 𝐹 , 𝑆𝑈 𝑁𝐷, 𝑆𝑈 𝑁𝑈 }, 𝑦 ∈ 𝑌 − {𝑌 𝐼𝑁𝐼}; (A59)
Through plugging in the affine decision rules and taking the dual,
The reformulation of the market availability constraint for wind tur- the electricity importation constraint (24) can be reformulated into con-
bines, namely constraint (21), is presented as constraints (A60)-(A63), straints (A71)-(A73). EVGl,n,i,y denotes the corresponding dual variables
by plugging in the affine decision rule and taking the dual. EVEBl,n,y is following the uncertainty set constraints for importation source s in
the dual variable introduced corresponding to the uncertainty set con- stage y.
straints for year y. ∑∑
𝜔𝑙,𝑛 ⋅ 𝐸𝑉 𝐾𝑙,𝑛,𝑠,𝑦 ≤ 𝐼 𝑀 𝑃 𝑀𝑠,𝑦 − 𝜌0𝑠,𝑦 , ∀𝑠 ∈ 𝑆, 𝑦 ∈ 𝑌 (A71)
∑∑ 𝑙 𝑛
𝜔𝑙,𝑛 ⋅ 𝐸 𝑉 𝐸 𝐵𝑙,𝑛,𝑦 ∑
𝑙 𝑛 𝜃𝑙,𝑘,𝑛 ⋅ 𝐸𝑉 𝐾𝑙,𝑛,𝑠,𝑦 ≥ 𝜌𝑠,𝑦,𝑘,𝑛 , ∀𝑠 ∈ 𝑆, 𝑦 ∈ 𝑌 , 𝑘, 𝑛 (A72)
∑ ∑ ∑ ∑ 𝑙
≤− 𝛼𝑖,𝑡,𝑦
0
+ 𝛼𝑖,𝑡,𝑦
0
,
∀𝑖∈{𝑊 𝑂𝑁,𝑊 𝑂𝐹 𝐹 } 𝑡∈𝑇 ∀𝑖∈{𝑊 𝑂𝑁,𝑊 𝑂𝐹 𝐹 } 𝑡∈𝑇
−1 𝐸𝑉 𝐾𝑙,𝑛,𝑠,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑠, 𝑦 (A73)
∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 } (A60) By applying the affine decision rules and deriving the correspond-
∑ ∑ ∑( ) ing dual problems, the non-negative constraints (25)–(29) can be refor-
𝜃𝑙,𝑘,𝑛 ⋅ 𝐸 𝑉 𝐸 𝐵𝑙,𝑛,𝑦 ≥ 𝛼𝑖,𝑡,𝑦,𝑘,𝑛 − 𝛼𝑖,𝑡,𝑦−1,𝑘,𝑛 ,
𝑙 𝑖∈{𝑊 𝑂𝑁,𝑊 𝑂𝐹 𝐹 } 𝑡∈𝑇 (A61) mulated into their robust counterpart, namely constraints (A74)-(A88),
∀𝑦 ∈ 𝑌 − {𝑌 𝐼 𝑁 𝐼 }, 𝑘, 𝑛 − {𝑘 = 2, 𝑛 = 𝑦} where AVAl,n,i,t,y , AVBl,n,i,t,y , AVCl,n,i,t,y , AVDl,n,i,t,y , and AVEl,n,s,y stand

16
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

for the dual variables for constraints (25), (26), (27), (28), and (29), [17] Kittner N, Lill F, Kammen DM. Energy storage deployment and innovation
respectively. for the clean energy transition. Nat Energy 2017;2(9):17125. doi:10.1038/nen-
∑∑ ergy.2017.125.
𝜔𝑙,𝑛 ⋅ 𝐴𝑉 𝐴𝑙,𝑛,𝑖,𝑡,𝑦 ≤ 𝛼𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (A74) [18] Ekren O, Ekren BY. Size optimization of a PV/wind hybrid energy conversion sys-
𝑛
tem with battery storage using simulated annealing. Appl Energy 2010;87(2):592–8.
𝑙
∑∑ doi:10.1016/j.apenergy.2009.05.022.
𝜔𝑙,𝑛 ⋅ 𝐴𝑉 𝐵𝑙,𝑛,𝑖,𝑡,𝑦 ≤ 𝛿𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (A75) [19] Jia L, Tong L. Renewables and storage in distribution systems: centralized
𝑙 𝑛
∑ ∑ vs. decentralized integration. IEEE J Sel Areas Commun 2016;34(3):665–74.
𝜔𝑙,𝑛 ⋅ 𝐴𝑉 𝐶𝑙,𝑛,𝑖,𝑡,𝑦 ≤ 𝜂𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑒 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (A76) doi:10.1109/JSAC.2016.2525638.
𝑙 𝑛 [20] Heffron R, Körner M-F, Wagner J, Weibelzahl M, Fridgen G. Industrial demand-side
∑∑ flexibility: a key element of a just energy transition and industrial development. Appl
𝜔𝑙,𝑛 ⋅ 𝐴𝑉 𝐷𝑙,𝑛,𝑖,𝑡,𝑦 ≤ 𝜇𝑖,𝑡,𝑦
0
, ∀𝑖 ∈ 𝐼𝑠 , 𝑡 ∈ 𝑇 , 𝑦 ∈ 𝑌 (A77) Energy 2020;269:115026. doi:10.1016/j.apenergy.2020.115026.
𝑙 𝑛
∑∑ [21] Fusco F, Nolan G, Ringwood JV. Variability reduction through optimal combina-
𝜔𝑙,𝑛 ⋅ 𝐴𝑉 𝐸𝑙,𝑛,𝑠,𝑦 ≤ 𝜌0𝑠,𝑦 , ∀𝑠 ∈ 𝑆, 𝑦 ∈ 𝑌 (A78) tion of wind/wave resources – an Irish case study. Energy 2010;35(1):314–25.
𝑙 𝑛 doi:10.1016/j.energy.2009.09.023.
∑ [22] Freris L, Infield D. Renewable energy in power systems. Wiley; 2008.
𝜃𝑙,𝑘,𝑛 ⋅ 𝐴𝑉 𝐴𝑙,𝑛,𝑖,𝑡,𝑦 ≥ −𝛼𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡, 𝑦, 𝑘, 𝑛 (A79) [23] Warren K, et al. Managing uncertainty in electricity generation and demand fore-
𝑙
∑ casting. IBM J Res Dev 2016;60(1) 8:1-8:13. doi:10.1147/JRD.2015.2496822.
𝜃𝑙,𝑘,𝑛 ⋅ 𝐴𝑉 𝐵𝑙,𝑛,𝑖,𝑡,𝑦 ≥ −𝛿𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑠 , 𝑡, 𝑦, 𝑘, 𝑛 (A80) [24] Pillot B, Al-Kurdi N, Gervet C, Linguet L. An integrated GIS and robust optimiza-
𝑙 tion framework for solar PV plant planning scenarios at utility scale. Appl Energy
∑ 2020;260:114257. doi:10.1016/j.apenergy.2019.114257.
𝜃𝑙,𝑘,𝑛 ⋅ 𝐴𝑉 𝐶𝑙,𝑛,𝑖,𝑡,𝑦 ≥ −𝜂𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑒 , 𝑡, 𝑦, 𝑘, 𝑛 (A81)
[25] Moret S, Babonneau F, Bierlaire M, Maréchal F. Decision support for strategic energy
𝑙
∑ planning: a robust optimization framework. Eur J Oper Res 2020;280(2):539–54.
𝜃𝑙,𝑘,𝑛 ⋅ 𝐴𝑉 𝐷𝑙,𝑛,𝑖,𝑡,𝑦 ≥ −𝜇𝑖,𝑡,𝑦,𝑘,𝑛 , ∀𝑖 ∈ 𝐼𝑠 , 𝑡, 𝑦, 𝑘, 𝑛 (A82) doi:10.1016/j.ejor.2019.06.015.
𝑙 [26] Moret S, Babonneau F, Bierlaire M, Maréchal F. Overcapacity in European power

𝜃𝑙,𝑘,𝑛 ⋅ 𝐴𝑉 𝐸𝑙,𝑛,𝑖,𝑡,𝑦 ≥ −𝜌𝑠,𝑦,𝑘,𝑛 , ∀𝑠, 𝑦, 𝑘, 𝑛 (A83) systems: analysis and robust optimization approach. Appl Energy 2020;259:113970.
doi:10.1016/j.apenergy.2019.113970.
𝑙
[27] Shang C, You F. A data-driven robust optimization approach to scenario-
𝐴𝑉 𝐴𝑙,𝑛,𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑙, 𝑛, 𝑡, 𝑦 (A84) based stochastic model predictive control. J Process Control 2019;75:24–39.
doi:10.1016/j.jprocont.2018.12.013.
𝐴𝑉 𝐵𝑙,𝑛,𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑙, 𝑛, 𝑡, 𝑦 (A85) [28] Shang C, You F. Data analytics and machine learning for smart process man-
ufacturing: recent advances and perspectives in the big data era. Engineering
2019;5(6):1010–16. doi:10.1016/j.eng.2019.01.019.
𝐴𝑉 𝐶𝑙,𝑛,𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑒 , 𝑙, 𝑛, 𝑡, 𝑦 (A86) [29] Ning C, You F. A transformation-proximal bundle algorithm for multistage adaptive
robust optimization and application to constrained robust optimal control. Automat-
𝐴𝑉 𝐷𝑙,𝑛,𝑖,𝑡,𝑦 ≥ 0, ∀𝑖 ∈ 𝐼𝑠 , 𝑙, 𝑛, 𝑡, 𝑦 (A87) ica 2020;113:108802. doi:10.1016/j.automatica.2019.108802.
[30] Pilpola S, Lund PD. Analyzing the effects of uncertainties on the mod-
elling of low-carbon energy system pathways. Energy 2020;201:117652.
𝐴𝑉 𝐸𝑙,𝑛,𝑠,𝑦 ≥ 0, ∀𝑙, 𝑛, 𝑠, 𝑦 (A88) doi:10.1016/j.energy.2020.117652.
[31] Pye S, Sabio N, Strachan N. An integrated systematic analysis of uncer-
References tainties in UK energy transition pathways. Energy Policy 2015;87:673–84.
doi:10.1016/j.enpol.2014.12.031.
[1] UNFCCC. Adoption of the Paris agreement. Available: https://undocs.org/ [32] X.A. Sun and Á. Lorca, "Robust optimization in electric power systems operations,"
FCCC/CP/2015/L.9/Rev.1; 2015, Accessed on: Mar. 10, 2020. in Integration of large-scale renewable energy into bulk power systems: from planning
[2] IEA. World energy outlook 2019. Available: https://www.iea.org/reports/world- to operation, P. Du, R. Baldick, and A. Tuohy, editors Cham: Springer International
energy-outlook-2019; 2019, Accessed on: Mar. 10, 2020. Publishing, 2017, pp. 227–58.
[3] Huberty M, Zysman J. An energy system transformation: framing research choices [33] Lorca Á, Sun XA, Litvinov E, Zheng T. Multistage adaptive robust op-
for the climate challenge. Res Policy 2010;39(8):1027–9. timization for the unit commitment problem. Oper Res 2016;64(1):32–51.
[4] Zeyringer M, Price J, Fais B, Li P-H, Sharp E. Designing low-carbon power systems for doi:10.1287/opre.2015.1456.
Great Britain in 2050 that are robust to the spatiotemporal and inter-annual variabil- [34] Zhao C, Wang J, Watson J, Guan Y. Multi-Stage robust unit commitment considering
ity of weather. Nat Energy 2018;3(5):395–403. doi:10.1038/s41560-018-0128-x. wind and demand response uncertainties. IEEE Trans Power Syst 2013;28(3):2708–
[5] Jacobson MZ. Clean grids with current technology. Nat Clim Change 2016;6(5):441– 17. doi:10.1109/TPWRS.2013.2244231.
2. doi:10.1038/nclimate2926. [35] J. Sathaye and A.H. Sanstad, "Bottom-up energy modeling," in "En-
[6] K. Glueck and L. Friedman. Biden announces $2 trillion climate plan. Avail- cyclopedia of energy," LBNL Report #: LBNL-54851, 2004, Available:
able: https://www.nytimes.com/2020/07/14/us/politics/biden-climate-plan.html; https://escholarship.org/uc/item/3wm7q17c.
2020, Accessed on: Jan 3, 2021. [36] Ning C, You F. Data-driven adaptive nested robust optimization: general modeling
[7] New York State Senate. Senate bill S6599. Available: https://www.nysenate. framework and efficient computational algorithm for decision making under uncer-
gov/legislation/bills/2019/s6599; 2019, Accessed on: Jan 3, 2021. tainty. AIChE J 2017;63(9):3790–817. doi:10.1002/aic.15717.
[8] Bazmi AA, Zahedi G. Sustainable energy systems: role of optimization modeling [37] Ning C, You F. Data-driven decision making under uncertainty integrating robust op-
techniques in power generation and supply—a review. Renew Sustain Energy Rev timization with principal component analysis and kernel smoothing methods. Com-
2011;15(8):3480–500. doi:10.1016/j.rser.2011.05.003. put Chem Eng 2018;112:190–210. doi:10.1016/j.compchemeng.2018.02.007.
[9] Hoggett R. Technology scale and supply chains in a secure, afford- [38] Shang C, Huang X, You F. Data-driven robust optimization based on kernel learning.
able and low carbon energy transition. Appl Energy 2014;123:296–306. Comput Chem Eng 2017;106:464–79. doi:10.1016/j.compchemeng.2017.07.004.
doi:10.1016/j.apenergy.2013.12.006. [39] Ning C, You F. Optimization under uncertainty in the era of big data and deep learn-
[10] Zhang MM, Wang Q, Zhou D, Ding H. Evaluating uncertain investment decisions in ing: when machine learning meets mathematical programming. Comput Chem Eng
low-carbon transition toward renewable energy. Appl Energy 2019;240:1049–60. 2019;125:434–48. doi:10.1016/j.compchemeng.2019.03.034.
doi:10.1016/j.apenergy.2019.01.205. [40] Supekar SD, Skerlos SJ. Analysis of costs and time frame for reducing CO2 Emis-
[11] Svobodova K, Owen JR, Harris J, Worden S. Complexities and contradictions in the sions by 70% in the U.S. auto and energy sectors by 2050. Environ Sci Technol
global energy transition: a re-evaluation of country-level factors and dependencies. 2017;51(19):10932–42. doi:10.1021/acs.est.7b01295.
Appl Energy 2020;265:114778. doi:10.1016/j.apenergy.2020.114778. [41] Ning C, You F. A data-driven multistage adaptive robust optimization framework
[12] Jacobson MZ, et al. 100% clean and renewable wind, water, and sunlight for planning and scheduling under uncertainty. AIChE J 2017;63(10):4343–69.
(WWS) all-sector energy roadmaps for the 50 United States. Energy Environ Sci doi:10.1002/aic.15792.
2015;8(7):2093–117. doi:10.1039/C5EE01283J. [42] Ben-Tal A, Goryashko A, Guslitzer E, Nemirovski A. Adjustable robust so-
[13] Heras J, Martín M. Social issues in the energy transition: effect on lutions of uncertain linear programs. Math Program 2004;99(2):351–76.
the design of the new power system. Appl Energy 2020;278:115654. doi:10.1007/s10107-003-0454-y.
doi:10.1016/j.apenergy.2020.115654. [43] Zappa W, Junginger M, van den Broek M. Is a 100% renewable Euro-
[14] Zhao N, You F. Can renewable generation, energy storage and energy efficient tech- pean power system feasible by 2050? Appl Energy 2019;233-234:1027–50.
nologies enable carbon neutral energy transition? Appl Energy 2020;279:115889. doi:10.1016/j.apenergy.2018.08.109.
doi:10.1016/j.apenergy.2020.115889. [44] Maïzi N, et al. Maximizing intermittency in 100% renewable and reliable power
[15] Jacobson MZ, Delucchi MA, Cameron MA, Frew BA. Low-cost solution to the systems: a holistic approach applied to Reunion Island in 2030. Appl Energy
grid reliability problem with 100% penetration of intermittent wind, water, 2018;227:332–41. doi:10.1016/j.apenergy.2017.08.058.
and solar for all purposes. Proc Natl Acad Sci USA 2015;112(49):15060–5. [45] M. Kintner-Meyer et al., "National assessment of energy storage for grid balancing
doi:10.1073/pnas.1510028112. and arbitrage," 2013.
[16] Hart EK, Jacobson MZ. The carbon abatement potential of high pene- [46] Nicoletti J, Ning C, You F. Incorporating agricultural waste-to-energy pathways into
tration intermittent renewables. Energy Environ Sci 2012;5(5):6592–601. biomass product and process network through data-driven nonlinear adaptive robust
doi:10.1039/C2EE03490E. optimization. Energy 2019;180:556–71. doi:10.1016/j.energy.2019.05.096.

17
N. Zhao and F. You Advances in Applied Energy 2 (2021) 100019

[47] Charnes A, Cooper WW. Deterministic equivalents for optimizing and satisficing un- [56] Ning C, You F. Data-driven adaptive robust unit commitment under wind
der chance constraints. Oper Res 1963;11(1):18–39. doi:10.1287/opre.11.1.18. power uncertainty: a Bayesian nonparametric approach. IEEE Trans Power Syst
[48] You F, Grossmann IE. Design of responsive supply chains un- 2019;34(3):2409–18. doi:10.1109/TPWRS.2019.2891057.
der demand uncertainty. Comput Chem Eng 2008;32(12):3090–111. [57] New York Independent System Operator. 2020 load & capacity data. Avail-
doi:10.1016/j.compchemeng.2008.05.004. able: https://www.nyiso.com/documents/20142/2226333/2020-Gold-Book-Final-
[49] Yue D, You F. Optimal supply chain design and operations under multi-scale un- Public.pdf; 2020, Accessed on: Jan 3, 2021.
certainties: nested stochastic robust optimization modeling framework and solution [58] Pehl M, Arvesen A, Humpenöder F, Popp A, Hertwich EG, Luderer G. Under-
algorithm. AIChE J 2016;62(9):3041–55. doi:10.1002/aic.15255. standing future emissions from low-carbon power systems by integration of life-
[50] Gao J, You F. A stochastic game theoretic framework for decentralized optimiza- cycle assessment and integrated energy modelling. Nat Energy 2017;2(12):939–45.
tion of multi-stakeholder supply chains under uncertainty. Comput Chem Eng doi:10.1038/s41560-017-0032-9.
2019;122:31–46. doi:10.1016/j.compchemeng.2018.05.016. [59] NREL. 2019 annual technology baseline. Available: https://atb.nrel.gov/
[51] Tong K, You F, Rong G. Robust design and operations of hydrocarbon biofuel supply electricity/2019/; 2019, Accessed on: Jan 3, 2021.
chain integrating with existing petroleum refineries considering unit cost objective. [60] U.S. Department of Energy. 2018 wind technologies market report. Available:
Comput Chem Eng 2014;68:128–39. doi:10.1016/j.compchemeng.2014.05.003. https://www.energy.gov/sites/prod/files/2019/08/f65/2018%20Wind%20
[52] Mulvey JM, Vanderbei RJ, Zenios SA. Robust optimization of large-scale systems. Technologies%20Market%20Report%20FINAL.pdf; 2019, Accessed on.
Oper Res 1995;43(2):264–81. doi:10.1287/opre.43.2.264. [61] U.S. Energy Information Administration. Annual solar photovoltaic module shipments
[53] Yue X, Pye S, DeCarolis J, Li FGN, Rogan F, Gallachóir BÓ. A review of approaches report. Available: https://www.eia.gov/renewable/annual/solar_photo/; 2019, Ac-
to uncertainty assessment in energy system optimization models. Energy Strategy cessed on.
Rev 2018;21:204–17. doi:10.1016/j.esr.2018.06.003. [62] A. Brooke, D.A. Kendrick, A. Meeraus, and R.E. Rosenthal, "GAMS, a user’s guide,"
[54] U.S. Energy Information Administration. U.S.-Canada electricity trade increases. Avail- 1988.
able: https://www.eia.gov/todayinenergy/detail.php?id=21992; 2015, Accessed [63] Pleßmann G, Blechinger P. How to meet EU GHG emission reduction targets? A
on: Jan 3, 2021. model based decarbonization pathway for Europe’s electricity supply system until
[55] Blei DM, Jordan MI. Variational inference for Dirichlet process mixtures. (in en). 2050. Energy Strategy Rev 2017;15:19–32. doi:10.1016/j.esr.2016.11.003.
Bayesian Anal 2006;1(1):121–43. doi:10.1214/06-BA104.

18

You might also like