Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Applied Energy 253 (2019) 113568

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Optimal scheduling for microgrids with hydrogen fueling stations T


considering uncertainty using data-driven approach

Xiong Wua, , Shixiong Qia, Zhao Wanga, Chao Duana, Xiuli Wanga, Furong Lib
a
School of Electrical Engineering, Xi’an Jiaotong University, China
b
Department of Electronic and Electrical Engineering, University of Bath, BA2 7AY, Bath, UK

H I GH L IG H T S

• Coordinated scheduling model of hydrogen fueling stations and microgrid is proposed.


• AUncertainties
data-driven approach is presented to characterize the uncertainty.
• A Wassersteinofmetric
renewable energy power, electrical and hydrogen load are considered.
• Several affine policy based
based distributionally robust optimization approach is applied.
• techniques are developed to simplify the problem.

A R T I C LE I N FO A B S T R A C T

Keywords: Hydrogen fueling stations can convert electricity to hydrogen with onsite hydrogen production technologies.
Microgrids When using renewables and cheap grid energy, hydrogen fueling stations present an economical and secure
Hydrogen fueling stations option for hydrogen production and supply for hydrogen vehicles. Microgrids with hydrogen fueling stations can
Hydrogen vehicles thus serve dual functions: producing hydrogen and meeting the demand of hydrogen vehicles. Previous re-
Scheduling
searches focus on hydrogen production and consumption in isolation, this paper integrates the two issues for
Uncertainty
Data-driven
achieving whole-system efficiency for the micrgrids. This paper proposes an optimal scheduling model for mi-
crogrids with hydrogen fueling stations taking consideration of wide range of uncertainties: renewable gen-
eration, electrical and hydrogen loads and electricity price. The uncertainty of renewable energy power and
loads is addressed using a data-driven chance constrained approach; while the uncertainty of electricity price is
handled with a distributionally robust optimization approach based on Wasserstein metric. When mitigating
uncertainty, the proposed approaches only utilize historical data and do not need any prior knowledge about the
true probability distribution of the uncertainty. Moreover, affine policy based techniques are developed to
convert the complicated optimization problem with huge uncertain constraints into a tractable mixed integer
linear programming problem. Finally, numerical analyses are carried out on an IEEE 33-bus distribution mi-
crogrid, the results show that the proposed model not only coordinate the distributed energy resources and
hydrogen fueling stations optimally, but also enables the schedule to strike a balance between the conservatism
and optimism against uncertainties.

1. Introduction fuel cell electric vehicles are gaining extensive attentions for their su-
perior features in fast refueling rate and long range [1]. On the other
The threat of environmental pollution necessitates the rapid devel- hand, the deployment of hydrogen fueling stations (HFSs) in the
opment of renewable energy generation and environment-friendly transportation systems is indispensable for the proliferation of HVs [2].
public transportation. Nowadays, conventional internal combustion How to produce hydrogen and deliver the hydrogen to HFSs is a major
engine vehicles are gradually replaced by low-carbon electric vehicles concern for investors. Generally, hydrogen can be produced through the
to mitigate air pollution and the climate change challenge. Among the thermal, electrolytic or biochemical processes [3]. Among the hydrogen
emerging electric vehicle technologies, hydrogen vehicles (HVs) such as production technologies, the power-to-hydrogen (PtH) technology is of


Corresponding author.
E-mail address: wuxiong@mail.xjtu.edu.com (X. Wu).

https://doi.org/10.1016/j.apenergy.2019.113568
Received 19 April 2019; Received in revised form 7 July 2019; Accepted 15 July 2019
Available online 30 July 2019
0306-2619/ © 2019 Elsevier Ltd. All rights reserved.
X. Wu, et al. Applied Energy 253 (2019) 113568

Nomenclature Parameters and constants

Sets P active power


Q reactive power
G set of distributed generators (DGs) Y hydrogen demand
H set of hydrogen fueling stations (HFSs) ρ price
W set of renewable energy sources n hydrogen molar flow
L set of distribution lines β, γ confidence level
B set of buses π distribution coefficient
B0 set of root buses E pressure of the storage tanks
T set of time intervals in the scheduling period η efficiency
P ambiguity set of the probability distribution R resistance
X reactance
Indices Rh gas constant
T temperature inside the hydrogen tank
 probability distribution or probability measure Mol molar mass of hydrogen
h script for HFSs LHV low heat value of hydrogen
el script for electrolyzers U volume of the hydrogen tank
hv script for hydrogen vehicles (HVs) V voltage
g script for DGs ΔV voltage deviation
gd script for utility grid
up , dn script for upward/downward Decision variables
pr script for electricity price
w script for renewable energy sources x binary variable indicating the operating state of DGs.“1”:
l script for distribution lines Active, “0”: Inactive
ld script for loads y variable related to hydrogen amount
in script for injection c variable related to cost
+ ,− script for export/import p variable related to active power
su, sd script for start-up/shut-down q variable related to reactive power
¯
(·)̲ ,(·) script for lower/upper bounds r variable related to reserve
φ (·) ,θ (·) children/parents of buses e variable related to pressure of the storage tank
δ (·) adjacency of buses or HFSs v voltage magnitude

(·) variables related to uncertainty s power forecast error of wind turbines (WTs) or loads

(·) observed sample ϕ aggregated net power forecast error
· number of elements α participation factor

interest to renewable energy utilization by electrolyzing water [4]. The transporting hydrogen through pipelines needs a huge initial invest-
produced hydrogen is stored in the pressurized tanks as gas. Subse- ment [6]. Consequently, distributed production of hydrogen with dis-
quently, it is either converted back to electricity through fuel cells or tributed energy resources (DERs) can be an effective way to reduce the
transported to HFSs by trucking or pipelines to fulfill HVs [5]. The hydrogen production and particularly transportation cost when HFSs
hydrogen can either be produced centrally in large scale and then coordinate with DERs in the form of microgrid. Microgrids, as smart
transported to HFSs or directly generated on site at HFSs locally. forms of distributed energy aggregators in power systems [7], exhibit
Commonly, the mass production of hydrogen is more economic. How- substantial potential to integrate DERs (such as distributed generators
ever, the extra transportation cost may offset the economic advantage if (DGs) and wind turbines (WTs)) with hydrogen production and fueling.
transporting hydrogen through trucking. On the other hand, The HFSs can be owned by the microgrid operator [8]. The operator

WT DG DG Price forecast

Load forecast

Utility gird
EMS
Weather forecast

PtH PtH

Hydrogen demand
HV HV
HFS 1 HFS 2

Fig. 1. A microgrid with HFSs controlled by an EMS.

2
X. Wu, et al. Applied Energy 253 (2019) 113568

maximizes its benefit by coordinating the HFSs and DERs, e.g., making 2) The approaches to handle uncertainties
use of internal renewable energy to generate hydrogen and sell it to
HVs. On the other hand, the HFSs can be independent or profit-driven There are two main uncertainties in the hydrogen-based microgrid,
utilities [9]. They participate in power markets to contract for cheaper namely, the uncertainty of renewable energy generation and loads, and
energy, produce hydrogen and sell it to HV for profits. In addition, they the uncertainty of electricity price. Most of related studies perform the
can profit from participating in auxiliary services such as reserve and stochastic programming optimization to handle the uncertainty. For
regulation services [10]. Although independent, they can also utilize example, Refs. [23,26,27] assume the uncertainties of renewable en-
renewable energy from the microgrid in a cooperative manner [11]. In ergy and load obey the normal distribution, and corresponding sce-
this paper, the HFSs are assumed to be owned by the microgrid op- narios are generated to simulate the uncertainty. However, the true
erator. An energy management system (EMS) supervises the operation probability distribution is not known in fact and only some sample data
of the microgrids, and makes the generation schedule of DERs and HFSs are available. Besides scenario based techniques, the chance con-
according to the forecast of electricity price, electrical and hydrogen strained approach is also an effective way to model the uncertainty. Ref.
loads, as shown in Fig. 1. Since microgrids are confronted with various [28] develops a unit commitment model to solve the day ahead sche-
uncertainties from renewable energy generation, electrical and hy- duling problem considering the uncertain loads and wind power with
drogen loads and electricity price, it is extremely challenging to manage the chance constrained method. Approximated approaches are com-
the microgrids with HFSs economically considering wide range of fac- monly used to solve this problem. However, the problem usually be-
tors. To address this issue, the paper proposes a scheduling model to comes computational challenging due to the nature of approximation
optimize the operation of microgrids with HFSs considering different [29]. Robust optimization approaches [30] also receive much attention
uncertainties. in uncertainty modelling. However, the proper estimation of the un-
Related works are summarized from two main aspects: certain interval is not an easy task. Irrational estimation may be too
conservative to compromise the optimization results. In recent years, a
1) Models associated with hydrogen management novel modeling paradigm, namely, distributionally robust optimization
is proposed to hedge against the ambiguity of probability distribution
A number of researches study the centralized hydrogen generation [31]. It constructs an ambiguity set, e.g. the moment based ambiguity
and transportation problems. Ref. [12] optimizes the hydrogen trans- set [32], based on historical data containing the possible information of
portation system by establishing a model aiming to minimize the uncertainty. This method is free of presumption of probability dis-
transportation cost of compressed gas trucks. Ref. [13] develops a tribution and thus gains much attention recently [33]. Ref. [34] applies
comprehensive model to optimize the design, planning and operation of a distributionally robust optimization method to a microgrid scheduling
a hydrogen value chain. Most of the researches however devote their problem. Ref. [35] further applies this method to a hybrid AC/DC mi-
attention to distributed generation of hydrogen in the context of hy- crogrid under a multi-time-scale framework. However, the studied
drogen-based microgrids. Moreover, the majority of them model the microgrids are both in islanded mode thus not influenced by electricity
microgrid sizing or scheduling with an electrolyzer/hydrogen tank/fuel prices. Further studies are needed to explore the influence of the un-
cell system, which can convert generated hydrogen back to electricity certain electricity price on the microgrid scheduling in the inter-
through fuel cells [14–16]. For example, Ref. [17] develops a model for connected mode.
sizing a stand-alone microgrid considering hydrogen loads and hy- Compared with conventional scheduling of electrolyzer/hydrogen
drogen storage systems. Ref. [18] presents a multi-criteria approach to tank/fuel cell system based microgrids, researches in the scheduling of
design a hybrid renewable microgrid including wind turbine, photo- microgrids with HFSs are very limited. Refs. [10,25,26] make inter-
voltaic panels, fuel cell, electrolyser, and hydrogen tank. Comparatively esting exploration into this topic, however the uncertainty and the
speaking, the scheduling topics are more frequent. Ref. [19] studies the coordination of DERs and HFSs are not investigated. Additionally, in
coordinated operation strategy of a multi-energy storage system which most literature, the uncertainty of renewable energy power and load is
incorporates the hydrogen generation. However, the uncertainties of modeled based on the assumption that it obeys a pre-determined
renewable power and electricity price are ignored. Ref. [20] performs probability distribution. Moreover, the uncertainty of electricity price is
an optimal scheduling of hydrogen-based microgrids with hybrid sto- commonly ignored. This paper proposes a novel scheduling model for
rage systems. Ref. [21] develops a real time energy management microgrids with HFSs considering the coordination of DERs and HFSs as
strategy for a renewable energy hydrogen based microgrid. The model well as a range of uncertainties. To our best knowledge, the joint op-
predictive control approach is applied to deal with the unpredictability eration of microgrid and HFSs has not been addressed before. The un-
of renewable energy in both references. Ref. [22] applies a fuzzy-based certainty is handled with a data-driven approach, which does not re-
approach to the real-time power management of such systems. In Ref. quire any prior knowledge about the true probability distribution of the
[23], a novel energy management strategy is proposed to control an uncertainty. To sum up, the contributions of this paper are:
isolated microgrid with the electrolyzer/hydrogen tank/fuel cell
system. Ref. [24] further takes the combined heat and power into ac- (1) An optimization model is proposed for guiding the scheduling of
count in the scheduling model. microgrids with HFSs. The model, which minimizes the total op-
In comparison, only a limited number of research optimize micro- erational cost under the worst-case distribution of electricity price
grid scheduling considering HVs to be supplied directly through on-site during the scheduling period, provides detailed power allocations
HFSs. In Ref. [25], a HFS with on-site hydrogen production is designed for DERs, optimal hydrogen generation for HFSs and optimal
by electrolysis of water using renewable energy. The operation opti- charging pattern for HVs in the day.
mization is carried out to schedule the HFS. However, the uncertainties (2) A data-driven approach is presented to characterize the uncertainty.
of renewable energy power and electricity price are not modeled. Ref. Specifically, the uncertainty of renewable energy power and load is
[10] proposes a scheduling model for private HFSs to serve both the addressed using a chance constrained approach; while the un-
transport sector and the electricity market operator. The paper mainly certainty of electricity price is handled with a distributionally ro-
concerns the operation of HFSs and ignores the coordination of DERs in bust optimization approach based on Wasserstein metric. To our
microgrids. Ref. [26] optimizes the home energy systems including knowledge, this is the first time that the Wasserstein metric based
electric and hydrogen vehicles and provides optimal charging pattern approach is introduced to deal with the uncertainty of electricity
for both electric and hydrogen vehicles. The uncertainty of wind-solar price in power system.
power is characterized by Gaussian probability distribution function. (3) Several affine policy based techniques are developed to interpret
Nevertheless, the true probability distribution is not known in fact. the uncertainty-related inequalities in an explicit manner and the

3
X. Wu, et al. Applied Energy 253 (2019) 113568

Wasserstein metric based methods are adopted to reformulate the management cost and the profit obtained from charging HVs, as stated
complex optimization problem into a tractable mixed integer linear in (4).
programming (MILP) problem.
1) Constraints of DG
The remainder of this paper is laid out as follows. Detailed opti-
mization models are provided in Section 2. Algorithms proposed for The active power output of DGs is limited by their upper/lower
solving the models are presented in Section 3. Case studies are illu- bounds, as indicated in (5). Similarly, the bounds of reactive power
strated in Section 4, and finally, conclusions are given in Section 5. output of DGs are shown in (6). The logical relation among the start-up
indicator, the shut-down indicator, and the on/off indicator of DGs is
2. Proposed model revealed in (7). The upward and downward reserves of DGs are used to
balance possible power fluctuations of renewable energy sources and
2.1. Basic scheduling model of the microgrid with HFSs loads, and they are nonnegative as indicated in (8).

P̲ itg x itg + ritg, dn ⩽ pitg ⩽ Pig x itg − ritg, up ∀ i ∈ G, ∀ t ∈ T (5)


In general, the goal of the microgrid operation is to minimize the
overall cost during the scheduling period. However, the uncertainty of 0 ⩽ qitg ⩽ Qi x itg
g
∀ i ∈ G, ∀ t ∈ T (6)
electricity price may influence the power trading strategy with the
utility grid. This paper develops a risk aversion strategy by considering x itg, su − x itg, dn = x itg − x ig, t − 1 ∀ i ∈ G, ∀ t ∈ T (7)
the worst distribution of the electricity price based on distributionally
robust optimization. Accordingly, the objective function of the problem ritg, dn ⩾ 0; ritg, up ⩾ 0 ∀ i ∈ G, ∀ t ∈ T (8)
is to minimize the expected total operational cost during the scheduling
period under the worst distribution of the electricity price, as for-
2) Constraints of HFS
mulated in (1).
J = min sup {  (cgd + cg + ch)} The HFSs produce hydrogen by consuming electric power from the
 ∈P (1)
utility grid or renewable energy sources through electrolyzers. The
produced hydrogen will then be stored in the pressurized tanks and
cgd = ∑∼
ρ tpr ptgd
t∈T (2) used for charging HVs. Fig. 2 shows the onsite hydrogen production
process in HFSs.
cg = ∑ ∑ (fig (pitg ) + ρig, su x itg, su + ρig, sd x itg, sd + ρig, om ) The AC power is converted to DC power through AC-DC converters
t∈T i∈G (3) firstly. The DC power is then used to electrolyze water by electrolyzers
and generate both hydrogen and oxygen. The hydrogen molar flow of
ch = ∑ ∑ (−ρ h yithv + ρih, su x ith, su + ρih, sd x ith, sd + ρih, om ) the electrolyzer is relevant to the AC power, as shown in (9).
t∈T i∈H (4)
ηel P el
where  denotes the expectation operator;  denotes the probability nel =
LHV h (9)
distribution of the electricity price; P denotes the ambiguity set of the
probability distribution; fig () represents the fuel cost function of DGs. The hydrogen molar flow can be adjusted through controlling the
The total operational cost includes the energy trading cost with the input power flow. The hydrogen is then compressed in the compressor
utility grid, the operation cost of DGs, and the operation cost of HFSs. and stored in the storage reservoir as gas. An important control variable
The microgrid is assumed to be a price-taker, it either buy or sell power of the storage reservoir is the tank pressure, which is a measure of the
from or to the utility grid with real-time electricity price. The energy hydrogen amount in the vessels. The tank pressure can be expressed as
trading cost with the utility grid is stated in (2). Specifically, the op- the function of the hydrogen inflow, outflow and the pressure at the
eration cost of DGs includes the fuel cost, the start-up and shut-down previous time, as indicated in (10).
cost, and the operation and management cost, as shown in (3). The
RhT h el
operation cost of HFSs is the difference between the aggregated sum of et = et − 1 + (n − nhv )
Uh (10)
the start-up and shut-down cost of electrolyzers, the operation and

- Power +

e- e-
O2 H2
AC
DC Anode - Catode +

AC Grid Converter Electrolyzer

Compressor

HV HFS H2 Storage Reservoir


Fig. 2. Onsite hydrogen production process in HFSs.

4
X. Wu, et al. Applied Energy 253 (2019) 113568

The consumed power by the electrolyzers is limited by their bounds, of local DERs minus the total power consumption of HFSs and loads, as
as shown in (11). Similarly, the logical relation among the state in- specified in (22)–(25). Note that the power injection of the utility grid is
dicators of the electrolyzer is revealed in (12), and the nonegativity of also considered at the root bus. The active (or reactive) power inflow is
reserves is indicated in (13). The storage capacity of the HFS is char- equivalent to the power outflow at each bus, as shown in (26) and (27).
acterized in terms of the storage tank pressure. The tank pressure is a The voltage difference between the two terminal buses of a branch is
function of the produced and consumed hydrogen at the current time related to the active and reactive power flows of the branch, as speci-
and the pressure at the previous time as shown in (14). The produced fied in (28). The bounds of branch active power, reactive power and bus
hydrogen is associated with the power consumption by the electrolyzers voltage are presented in (29)–(32). It should be noted that V0 is the
considering the conversion efficiency; while the consumed hydrogen is reference voltage. If the bus is the root bus, then the bus voltage is set to
equal to the total hydrogen consumption of the HVs. The hydrogen the reference voltage as indicated in (31); otherwise, the bus voltage
storage levels, are restricted by their capacity limits as shown in (15). should be within a given interval as specified in (32).
Constraint (16) enforces the stored hydrogen at the end of the sche-
duling period to be equal to that at the beginning so as to facilitate the pjtin = ptgd + ∑i ∈ δg (j) pitg − ∑i ∈ δh (j) pith + ∑i ∈ δw (j) pitw − pitld
scheduling in the next period. ∀ j ∈ B0, ∀ t ∈ T (22)
− Pih x ih + rith, dn ⩽ −pith ⩽ −rith, up ∀ i ∈ H, ∀ t ∈ T (11) qjtin = qtgd + ∑i ∈ δg (j) qitg + ∑i ∈ δw (j) qitw − qitld
x ith, su − x ith, dn = x ith − x ih, t − 1 ∀ i ∈ H, ∀ t ∈ T (12) ∀ j ∈ B0, ∀ t ∈ T (23)

rith, dn ⩾ 0; rith, up ⩾ 0 ∀ i ∈ H, ∀ t ∈ T (13) pjtin = ∑i ∈ δg (j) pitg − ∑i ∈ δh (j) pith + ∑i ∈ δw (j) pitw − pitld
h h ∀ j ∈ B B0, ∀ t ∈ T (24)
RhT h ⎛ ηi pit ⎞
eith = eih, t − 1 + h h ⎜ h
− yithv ⎟ ∀ i ∈ H, ∀ t ∈ T
Ui Mol ⎝ LHV (14) qjtin = ∑i ∈ δg (j) qitg + ∑i ∈ δw (j) qitw − qitld

E̲ ih ⩽ eith ⩽ E ih ∀ i ∈ H, ∀ t ∈ T (15) ∀ j ∈ B B0, ∀ t ∈ T (25)

eih, T = eih,0 ∀ i ∈ H, ∀ t ∈ T (16) pjtin = ∑ l


pjk ,t − ∑ pijl , t ∀ j ∈ B, ∀ t ∈ T
k ∈ φ (j ) i ∈ θ (j ) (26)
3) Constraints of HV
qjtin = ∑ l
qjk ,t − ∑ qijl , t ∀ j ∈ B, ∀ t ∈ T
k ∈ φ (j ) i ∈ θ (j ) (27)
The traveling plan, arriving time and departure time of HVs are
private information and exhibit dramatic uncertainty. Recent studies
⎛p l R + q l X ⎟⎞
usually take the electric vehicles (EVs) or HVs as an entity and assume ij, t ij
⎜ ij, t ij
they are controlled by an aggregator [36,37]. However, not all the EVs vit − vjt = ⎝ ⎠ ∀ (i, j ) ∈ L
or HVs are willing to be controlled in reality. Hence, we assume the
V0 (28)
HFSs directly forecast the hydrogen demand according to their ex-
− Pijl ⩽ pijl , t ⩽ Pijl ∀ (i, j ) ∈ L (29)
perience and historical data. The consumed hydrogen is equal to the
hydrogen forecast in the basic model, as shown in (17). The uncertainty l l
− Qij ⩽ qijl , t ⩽ Qij ∀ (i, j ) ∈ L (30)
will be considered in the following section.
yithv = Yithv ∀ i ∈ H, ∀ t ∈ T (17) vit = V0 ∀ i ∈ B0 (31)

V0 − ΔV ⩽ vit ⩽ V0+ΔV ∀ i ∈ B B0 (32)


4) Constraints of utility grid:

Constraints (18) and (19) represent the bounds of the active (re- 2.2. A chance constrained approach to handle the uncertainty of renewable
active) power exchanged with the utility grid. energy power and load
− P gd ⩽ ptgd ⩽ P gd ∀t∈T (18)
The uncertainty of the renewable energy power and load can’t be
− Q gd ⩽ qtgd ⩽ Q gd ∀t∈T (19) ignored as it may have a significant influence on the operation. To
characterize these influences, a chance constrained approach is em-
ployed in this paper for its advantage in balancing the conservatism and
5) Constraints of renewable energy
optimism. To interpret the method clearly, only WTs are considered in
the microgrid. However, the method can also handle the cases which
The active power output of renewable energy sources follows the
include single photovoltaic arrays (PVs) or both WTs and PVs. In ad-
power forecast profile as shown in (20), while the reactive power
dition, the power forecast error of WTs or loads is assumed to follow the
output of renewable energy sources is considered to be adjustable, as
same distribution. This is consistent with practices as natural resources
stated in (21).
such as wind and irradiation at the same area are nearly the same.
pitw = Pitw ∀ i ∈ W, ∀ t ∈ T (20) Therefore, a similar error is shown in the prediction systems. The actual
power output of WTs is obtained by adding the forecast power and the
0⩽ qitw ⩽ Qitw ∀ i ∈ W, ∀ t ∈ T (21) forecast error, as shown in (33). Similarly, the actual load is shown in
(34). Moreover, the aggregated net power forecast error of the micro-
6) Constraints of power flow grid is specified in (35). In addition, the hydrogen load is also con-
sidered. The actual hydrogen demand of HVs is shown in (36).
The microgrid power flow is formulated with the LinDistFlow model, ∼
which is widely used in distribution systems [38–40]. The complete p itw = pitw + ∼
s tw Pitw ∀ i ∈ W, ∀ t ∈ T (33)
formulations are expressed in (22)–(32). The amount of active (or re- ∼
p itld = pitld + ∼
s tld Pitld ∀ j ∈ B, ∀ t ∈ T (34)
active) power injection at each bus equals the aggregated power output

5
X. Wu, et al. Applied Energy 253 (2019) 113568

∼ ∼
ϕt = ∼
s tld ∑ Pitld − ∼s tw ∑ Pitw ∀t∈T vjt − V0 = ∑i ∈ G (πijg ∼
p itg + πij′ g qitg )+
(35)
i∈B i∈W
∑i ∈ H π h (−∼
ijp h) + ∑
it (π w ∼
i∈W
p w + π ′ w q w )−
ij it ij it

y ithv = yithv + ∼
s thv Yithv ∀ i ∈ H, ∀ t ∈ T (36) ∑i ∈ B (πijld∼
p itld + πij′ ld qitld ) ∀ t ∈ T, ∀ j ∈ B (47)

An affine policy based approach [41,42] is employed to dispatch the Consequently, constraint is directly associated with the uncertainty,
power forecast error. It should be noted that the microgrid is viewed as as shown in (48).
a controllable entity and the internal power fluctuation needs to be ∼
handled itself. Therefore the total power forecast error is assigned to the
⎧ − ΔV¯ ⩽ ∑ (πijg (pitg + αijg ϕt ) + πij′ g qitg )+ ⎫
⎪ i∈G ⎪
actual power output of DGs and HFSs affinely through participation ⎪ ∼ ⎪
⎪ h h h ⎪
factors:

∑ πij (−pit + αit ϕt )+

∼ ∼ t i∈H

p itg = pitg + αitg ϕt ∀ i ∈ G, ∀ t ∈ T (37) ⎨ ∼w
∑ (πijw (pitw + s t Pitw ) + πij′w qitw )− ⎬
⎪ ⎪
⎪ i∈W ⎪

−∼ ⎪ ∼ld ⎪
p ith = −pith + αith ϕt ∀ i ∈ H, ∀ t ∈ T (38) ∑ (πijld (pitld + s t Pitld ) + πij′ ld qitld ) ⩽ ΔV¯
⎪ ⎪
⎩i∈B ⎭
The affine policy is also consistent with reality as the aggregated net
1−β ∀ t ∈ T, ∀ j ∈ B (48)
power forecast error can be dispatched to each DG and HFS in the
microgrid. (39) and (40) specify the nonnegativity of the participation
factors, and (41) indicates that the sum of the participation factors 2.3. Compact form of the model
equals 1.
The proposed model can be written as the following compact form:
αitg ⩾ 0 ∀ i ∈ G, ∀ t ∈ T (39)

min sup   ∼ {C(x , ξ )}
ξ
αith ⩾ 0 ∀ i ∈ H, ∀ t ∈ T (40)  ξ∼∈ P ξ∼ (49)

s. t . H(x ) ⩽ 0 (50)
∑ αitg + ∑ αith = 1 ∀t∈T
i∈G i∈H (41)
 τ {G(x , τ͠ ) ⩽ 0} ⩾ 1 − β ∀  τ ∈ Pτ (51)
Let t represent the joint probability distribution for random vari- ∼
where C(x , ξ ) denotes the objective function; x represents the de-
s tw and ∼
ables ∼ s tld . In (42) and (43), the adequacy of the upward (or ∼
downward) reserve of DGs and HFSs is ensured with a high probability. terministic variable vector; ξ represents the electricity price variable
Likewise, (44) ensures the safe interval of bus voltage and (45) guar- vector; τ͠ represents the WT power and load variable vector; H(x) de-
antees the safe interval of the tank pressure with a high probability. The notes the deterministic constraints; G(x , τ͠ ) denotes the chance con-
chance constraints (42)–(44) ensure the system security with a high straints; (49) represents (1); (50) stands for (5)–(8), (11)–(32),
probability, therefore it is more reliable for the system. (39)–(41); (51) corresponds to (42), (43), (46) and (48). Obviously, the
deterministic constraints are tractable MILP formulations. The chance

t { −ritg, dn ⩽ αitg ϕt ⩽ ritg, up} ⩾ 1 − β ∀ i ∈ G, ∀ t ∈ T (42) constraints and the objective function are intractable. However, it’s

found that the uncertainty-related vectors ξ and τ͠ are decoupled. Thus,
∼ the chance constraints and the objective function can be handled se-
t { −rith, dn ⩽ αith ϕt ⩽ rith, up} ⩾ 1 − β ∀ i ∈ H, ∀ t ∈ T (43)
parately. In the following section, state-of-art techniques are developed
t {V0 − ΔV ⩽ ∼
vit ⩽ V0 + ΔV } ⩾ 1 − β ∀ j ∈ B, ∀ t ∈ T (44) to convert the chance constraints and the objective function into
tractable MILP formulations. Then the problem becomes an MILP pro-
t {E̲ ih ⩽ ∼
e ith ⩽ E ih} ⩾ 1 − β − γ ∀ i ∈ H, ∀ t ∈ T (45) blem which can be solved by solvers such as CPLEX and GUROBI.

By combining (14) and (38), (45) can be rewritten as the explicit 3. Solution to the model
formulations which depend on the uncertainty:
3.1. Deterministic reformulation of the chance constraints
g g g∼ g g
⎧ − ΔV ⩽ ∑i ∈ G (πij (pit + αij ϕt ) + πij′ qit )+ ⎫
⎪ h∼

⎪ h h
∑i ∈ H πij (−pit + αit ϕt )+ ⎪ A Monte Carlo method (MCM) [43] using historical data is em-
t ⩾1−β ∀ ployed to reformulate the chance constraints into tractable MILP forms,
⎨ ∑i ∈ W (πijw (pitw + ∼ s tw Pitw ) + πij′ w qitw )− ⎬
⎪ ⎪ as implied in the following lemma.
ld ∼
⎪∑i ∈ B (πij (pit + s t Pitld ) + πij′ ld qitld ) ⩽ ΔV
ld ld ⎪
⎩ ⎭ Lemma 4.1. Let Stw (x ) , Stld (x ) denote the cumulative distribution functions

t ∈ T, ∀ j ∈ B (46) (CDFs) of random variables ∼ s tw , ∼
s tld respectively. If− ritg, dn ⩽ αitg ϕt ⩽ ritg, up ,
∀ i ∈ G, ∀ t ∈ T , ∀∼ w w w
s t ∈ [ s̲ t′ , st′ ], ∀∼ ld ld
s tld ∈ [ s̲ t′ , st′ ], where
However, (44) relies on the uncertainty implicitly. To interpret (44)
s̲ t′ w = Stw − 1 (β1 2) , st′ w = Stw − 1 (1 − β1 2) , s̲ t′ ld = Stld − 1 (β2 2) and
in a straightforward manner, the voltage should be reformulated as the
st′ ld = Stld − 1 (1 − β2 2) with β1 + β2 = β . Then (42) is satisfied.
expression associated with the stochastic power injection. It is fortunate
that such relation is demonstrated in the following lemma. Proof. Please see Appendix B.
Lemma 3.1. In the context of LinDistFlow model, the bus voltage is linearly Then (42) can be rewritten as the following deterministic forms:
correlated with the power injection of all buses in a radial distribution
system. ⎛ ⎞
− ritg, dn ⩽ αitg ⎜ s̲ t′ ld ∑ Pitld − st′w ∑ Pitw ⎟ ∀ i ∈ G, ∀ t ∈ T
Proof. Please see Appendix A. ⎝ i∈B i∈W ⎠ (52)

Note that the power injection derives from DERs, HFSs and loads, ⎛ ⎞
therefore the voltage can be expressed as a linear formulation of the αitg ⎜st′ ld ∑ Pitld − s̲ t′ w ∑ Pitw ⎟ ⩽ ritg, up ∀ i ∈ G, ∀ t ∈ T
⎝ i∈B i∈W ⎠ (53)
power output of DERs, HFSs and loads, as shown in (47).

6
X. Wu, et al. Applied Energy 253 (2019) 113568

Similarly, (43) can be addressed following the same method. With about the probability distribution.
regard to the voltage security, denote F (∼
s tw, ∼
s tld ) as follows:
∼ ∼
F (∼
s tw, ∼
s tld ) = ∑i ∈ G (πijg (pitg + αijg ϕt ) + πij′ g qitg ) + ∑i ∈ H πijh (−pith + αith ϕt )+ 3.2. Reformulation of the model based on Wasserstein metric
∑ (π w (p w + ∼
i∈W ij s w Pw ) + π ′ w q w )
it t it ij it The model is a distributionally robust optimization problem. A
− ∑i ∈ B (πijld (pitld + ∼
s tld Pitld ) + πij′ ld qitld ) Wasserstein metric based data driven approach is introduced to convert
(54) the problem into a tractable form. Compared with conventional mo-
ment-based approach which only includes limited probabilistic in-
Then (44) is reformulated as follows: formation such as mean and covariance, this method incorporates more
− ΔV ⩽ F (st′ w, st′ ld ) ⩽ ΔV , −ΔV ⩽ F (st′ w, s̲ t′ ld ) ⩽ ΔV probabilistic information. In addition, it is consistent with the practice
that the more data are collected, the more we know the true distribu-
− ΔV ⩽ F ( s̲ t′ w, st′ ld ) ⩽ ΔV , −ΔV ⩽ F ( s̲ t′ w , s̲ t′ ld ) ⩽ ΔV
tion [44]. To interpret the method explicitly, the Wasserstein distance is
∀ t ∈ T, ∀ j ∈ B (55) introduced firstly.
In a similar manner, let Sthv (x ) denote the CDFs of random variables The Wasserstein distance [45] provides an effective way to measure

s thv , s̲ t′ hv = Sthv − 1 (γ 2) , st′ hv = Sthv − 1 (1 − γ 2) , denote F ′ (∼
s tw, ∼
s tld, ∼
s thv ) as the statistical distance between two probability distributions, which is
follows: defined as follows:
h h h∼ Definition 4.1. The Wasserstein distance between two probability
RhT h ⎛ ηi (pit − αit ϕt ) ⎞
F ′ (∼
s tw, ∼
s tld, ∼
s thv ) = eih, t − 1 + − (yithv + ∼
s thv Yithv ) ⎟ distributions 1 and 2 is:
Ui Mol h ⎜
h LHV h
⎝ ⎠ ∼ ∼ ∼ ∼
(56) Dis (1, 1) = inf
Π
{∫ Ω2 }
∥ξ1 − ξ2 ∥ Π(dξ1, dξ2)
(58)
Then (46) is reformulated as follows: ∼ ∼
where Π is a joint distribution of ξ1 and ξ1 with marginal distributions 1

E̲ ih ⩽ F ′ ( s̲ t′ w, st′ ld, st′ hv ), F ′ (st′ w, s̲ t′ ld, s̲ t′ hv ) ⩽ E ih and 2 ; Ω is the support space of ξ ; · is the norm function, and we use
∀ t ∈ T, ∀ i ∈ H (57) 1-norm in this paper for its tractability in the model. Given a historical
data set with N samples, the true distribution  belongs to an ambiguity
Remark. The reformulation is essential to find the bounds of the set:
uncertainty at the confidence level based on available historical data. N) ⩽ ε }
P = { ∈ Φ: Dis( ,  (59)
The reformulation converts the chance constraints into tractable linear
counterparts eventually. This method greatly simplifies the problem where  N denotes the empirical distribution which derives from
and avoids highly demanding algorithm such as second order cone historical data of ξ ; Φ denotes the set of all probability distribution
programming, which is commonly used to handle chance constraints. In with support Ω; ε is the radius of the Wasserstein ball. Generally, it is a
addition, it’s data driven and independent on the prior knowledge sample size dependent function which decreases to 0 as N increases to

Start

Data preprocessing

WT and load data Electricity price data

Obtain the historical data of the WT power and load


forecast error Obtain the historical data of the electricity price
forecast error

Arrange the sample in an ascending order


Construct the ambiguity set of the electricity prices
Find the value of the inverse CDFs at the confidence levels according to their bounds
according to the MCM

Model building and solving

Reformulate the chance constraints as MILP equations

Reformulate the distributionally robust optimization problem


as an MILP problem

Enter the parameters of the microgrid scheduling model

Build the MILP model in Matlab/YALMIP

Solve the model by invoking CPLEX solver

End

Fig. 3. Flowchart of the proposed approach for solving the model.

7
X. Wu, et al. Applied Energy 253 (2019) 113568

infinity. One empirical formula of ε is given as follows: probability distribution of the electricity price. Moreover, the result
tends to be less conservative by incorporating more data, which is
2 1 ⎞ consistent with the practice that the more historical data is available,
ε = Dia log ⎛ ⎜ ⎟
N ⎝ 1 − χ⎠ (60) the more we know about the information.

where Dia is the diameter of the support of the random variable; χ is the
confidence level. Subsequently, a distributionally robust optimization 3.3. Summary of the solving process
approach based on the Wasserstein metric is developed to handle the
uncertainty of electricity price. In summary, the flowchart of the proposed approach is shown in
∼ Fig. 3. At the stage of data preprocessing, the WT and load data and
Construct the support set of the electricity price vector ξ according electricity price data are handled separately. For WT and load data, the
to its bounds, as indicated in (61): historical data of the forecast error for WT power and load are ranked
∼ ∼ ascendingly firstly. Subsequently, find the value of the inverse CDFs at
Ω = {ξ ∈ Rm : Bξ ⩽ b} (61)
the given confidence level using the MCM. For electricity price data,
where B, b are corresponding constant matrix and vector. Considering after obtaining the historical data of the electricity price forecast error,

the worst realization of the electricity price vector ξ in the ambiguity we construct the support set of the electricity prices according to their
set, the objective is expressed as follows: bounds.
Afterwards, the model building and solving procedure begins. At
J = min sup {  (cgd + cg + ch)}
 ∈P this stage, we reformulate the chance constraints and transform the
= min sup {  ( ∑t ∈ T ∼
ρ tpr ptgd + cg + ch)} distributionally robust optimization into an MILP problem. Finally, the
 ∈P parameters of the microgrid scheduling model are initialized. The

= min sup {  ( 〈z, ξ 〉 + cg + ch)} transformed MILP model is built in Matlab/YALMIP, and solved by
z∈Ξ  ∈P (62) CPLEX solver.
where z denotes the power vector exchanged with the utility grid; 〈 〉
denotes the scalar product operator; Ξ denotes the feasible region of z, 4. Case studies
which is confined by (5)–(8), (11)–(32), (39)–(41) and (52)–(57);.
Based on the Wasserstein distance and dual theory, the objective To verify the effectiveness of the proposed model, a IEEE 33-bus
function can be reformulated as the following tractable form, as in- distribution system based microgrid [46] is studied. The tested micro-
dicated in (63). The detailed proof can be referred to [45]. grid is displayed in Fig. 4. The regular load of the microgrid is about
3.35 MW. There are 2 DGs and 4 WTs in the microgrid. The parameters
1 N
⎧ inf ψε + ∑i = 1 κi of DERs are shown in Table 1. In addition, 3 HFSs are placed in the
z , ψ, κi, μi N
⎪ microgrid. Each HFS has an electrolyzer and a group of hydrogen sto-
⎪ s. t . z ∈ Ξ rage tanks, the parameters of which are illustrated in Table 2 based on

J=
⎨ cg + ch + 〈z, ξi〉 + 〈μi , b − Bξ î 〉 ⩽ κ i
̂ ∀i⩽N [10,47]. The real-time electricity price is procured from [48]. The
⎪ ∥BTμi − z∥∗ ⩽ ψ ∀ i ⩽ N selling price of hydrogen is 1.35 $/kg. The forecast error is simulated

⎪ μi ⩾ 0 ∀ i ⩽ N based on ARMA model, which is demonstrated in Fig. 5 when adding
⎩ the forecast value. 1000 samples are analyzed.

where ξi ̂ is the ith sample of ξ ; ψ is an auxiliary variable; κi, μi are The forecast WT power and load and their different confidence in-
sample-dependent auxiliary variable and vector. Since we use the 1- tervals are shown in Fig. 6. The hydrogen demand of three HFSs is also
norm to measure the Wasserstein distance in the uncertainty space, the shown in Fig. 7. The confidence level 1-β is set to 95% and 1-β-γ is set to
||·||∗ is the ∞ norm, whereby the optimization reduces to an MILP 96% in the studied case. The safe interval of voltage is [0.9, 1.1]. Nu-
problem, which could be solved directly by off-the-shelf solvers such as merical experiments are conducted in Matlab with YALMIP and solved
CPLEX. by CPLEX solver.

Remark. The advantage of the Wasserstein metric based approach lies


1) Generation schedule of DGs and HFSs
in that it can convert the original problem into a tractable linear model.
In addition, the method is data-driven without any presumption on the

Fig. 4. Structure of the IEEE 33-bus distribution system based microgrid.

8
X. Wu, et al. Applied Energy 253 (2019) 113568

Table 1
Technical Parameters of DERs.
Type Rated power (kW) Minimum power (kW) Start-up cost ($) Shut-down cost ($) Maintenance cost ($/h) Number

DG 1000 160 0.14 0.14 0.1 2


WT 500 – – – 0.1 4

Table 2 20
Technical Parameters of HFSs. HFS1 HFS2 HFS3
Forecast
Parameters Value Parameters Value Parameters Value
100%

Hydrogen Demand (kg)


Rh 8.314 (J/mol K) Uh 200 (m3) ηh 0.6 15 95%
LHVh 39.72 (kWh/kg) Th 313 (K) 400 (kW)
90%
P̄ h
85%
Molh 0.002 (kg/mol) Ēh 10 (bar) E̲h 2 (bar)
80%
ρh,su 0.14($) ρh,sd 0.14($) ρh,om 0.1($/h)
10

0
0 24;0 24;0 24
Time (h)
Fig. 7. Forecast hydrogen demand and their confidence intervals.

certain high-price periods (such as 8:00–14:00 and 17:00–21:00),


power generated from local DGs is relatively cheap and the DGs operate
with almost full power during these periods. It should be noted that
most of DGs and HFSs seldom operate at rated power simultaneously.
This is because the random WT power shrinks the available power of
dispatchable DGs and HFSs owing to the reserve requirement. The
Fig. 5. Forecast electricity price and its uncertain scenarios. generation of WTs is relatively cheap and they are utilized to generate
power to the fullest. Obviously, the resulting schedule demonstrates the
economic operation of the microgrid.
Table 3 lists the detailed costs of DGs and HFSs in the microgrid. It
shows that the majority of the costs are for the energy purchase cost
from the grid. In addition, the generation of DGs also costs a lot. Ac-
cording to Fig. 8, all the HFSs can produce hydrogen and sell it to HVs
to get a profit. Therefore, their costs are negative. It should be noted
that this doesn’t mean that the onsite hydrogen production in HFSs is an
economic way for generating hydrogen. This is because the cost doesn’t
take the power consumption cost for hydrogen production into account.
However, it’s considered in the microgrid since the DGs and HFSs are
viewed as a joint entity.
In fact, all the HFSs still can’t profit from electrolyzing water and
selling hydrogen even if they utilize a part of cheap energy from the
microgrid. This is mainly due to the high price of electricity and low
conversion efficiency of electrolyzing hydrogen. For example, the hy-
drogen production cost is nearly 5$/kg in current situation. However,
the selling price for hydrogen is only 1.35$/kg. With the improvement
Fig. 6. Forecast WT power and electrical load and their confidence intervals. of conversion efficiency and utilization of cheap electricity, the HFSs
are sure to get profits. To verify this, Fig. 9 plots the influence of the
Fig. 8 gives the generation schedule of DGs and power consumption electricity price and conversion efficiency on the cost of HFS1. As can
of HFSs in the microgrid in the day. Note that HFSs produce hydrogen be seen, when the price is fixed, the cost decreases as the efficiency
by consuming power, and thus their power output is negative. As can be increases. Similarly, when the efficiency is fixed, the cost increases as
seen in Fig. 8, the hydrogen is produced and stored mostly at 0:00–8:00, the electricity price rises. Generally, HFS1 can make a profit when the
14:00–18:00, and 22:00–24:00. This is partly due to the fact that the average price falls below 0.03 $/kWh in the studied case.
wind power is abundant then, and partly because the electricity price is
relatively low at that time. The majority of power supply is purchased 2) Hydrogen production and consumption of HFSs
from the utility grid then for the sake of low price. We find that DGs
which are off most of the time are started up to fulfill loads only when The hydrogen production and consumption as well as the state of
the electricity price rises during the peak hours. For example, during charge (SOC) of three HFSs in the day are shown in Fig. 10. Most of the

9
X. Wu, et al. Applied Energy 253 (2019) 113568

Fig.8. Generation schedule of DGs and power consumption of HFSs.

Table 3 delivered to the storage tanks continuously although it’s consumed by


Detailed costs of DGs and HFSs in the microgrid. HVs at the same time. Since the produced hydrogen can fill the gap
Type Fuel/Hydrogen Start-up/Shut- Operation/ Total cost
caused by the hydrogen consumption of HVs, the SOC can always
cost ($) down cost ($) Management cost ($) ($) maintain a high level. As the pressure of the storage tanks should be
kept to their initial level at the end of the day, all the SOC of the HFSs
Grid – – – 2792.9 returns to the same level at time 24.
DG1 1176.4 0.6 2.8 1179.8
DG2 1104.4 0.6 2.8 1107.8
HFS1 −147.7 0.1 2.8 −144.8 3) Voltage fluctuations of all buses
HFS2 −109.8 0.1 2.8 −106.9
HFS3 −150.0 0.1 2.8 −147.1 The voltage variations of buses during the day are shown in Fig. 11.
The minimum voltage and maximum voltage in the day is 0.941 p.u.
and 1.006 p.u. respectively. Since all the voltage is within the safe in-
terval, the network security is ensured. This is mainly because the re-
active power of DERs is adequate. Therefore, the voltage security
constraints have little influence on the scheduling results in the studied
case.

4) The verification of proposed chance constrained approach

To verify the proposed chance constrained approach, the generation


schedule of DGs and power consumption of HFSs regardless of the
uncertainty (referred to as deterministic case) are conducted and the
results are displayed in Fig. 12. As can be observed, the generation
behaviors of DGs and the power consumption patterns of HFSs are
nearly the same as those in the case using proposed approach (referred
to as chance constrained case). The DGs and HFSs mostly operate with
rated power once utilized in deterministic case. However, this strategy
leaves the microgrid exposed to great uncertainty as all the power
fluctuations have to be compensated by the utility grid, which disobeys
Fig. 9. Influence of the electricity price and efficiency on the cost of HFS1. the design of microgrids since microgrids are viewed as controllable
entities. In comparison, the proposed model takes the uncertainty of
hydrogen production and consumption behaviors happen during the renewable energy power and load into account. All DGs and HFSs spare
evening or morning. This is consistent with practices as the cheap wind some reserves for possible disturbances. In this way, the uncertainty of
power is abundant during those periods, when the low electricity prices the microgrid entity becomes much less.
also easily incur the hydrogen production. Although there are a few On the other hand, the case applied with robust optimization is
fluctuations in the SOC curve, all the SOC lies in the safe interval. conducted (referred to as robust case). The method considers the worst
Moreover, the minimum SOC of three HFSs is far from their minimum case of the uncertainty, and it's realized by setting the confidence level
bound (0.2 p.u.). This is because the hydrogen is produced and to 1. The scheduling results between the chance constrained and robust

10
X. Wu, et al. Applied Energy 253 (2019) 113568

Fig. 10. Hydrogen production and consumption as well as SOC of the HFSs in the day.

flexible measures are available, which leads to an increase in the op-


eration cost. Specifically, the confidence level increases to 1 means that
all constraints should be met completely. As can be seen, the objective
value increases sharply to the highest in this situation. The problem
becomes a robust optimization problem as the constraints should be
met even in the worst case, which consequently requires more costs.
Although it makes the microgrid more robust against the uncertainty, it
also compromises the operational cost. In this regard, a proper con-
fidence level is essential.

6) The verification of the proposed distributionally robust optimization

To validate the effectiveness of the proposed distributionally robust


optimization, cases with different electricity price sample sizes under
four different distributions of the renewable energy power and load
forecast error are conducted. The performance of the stochastic opti-
mization (SO) based method is utilized to compare with that of the
Fig. 11. Voltage fluctuations of all buses in the day.
proposed distributionally robust optimization (DRO) approach. The
results are compared in Fig. 15. As can be seen, the objective value
cases are compared in Fig. 13. derived from DRO is always lower than the practical objective value
It’s obvious that the results in the robust case are more conservative derived from SO. This is mainly because the proposed method considers
than those in the chance constrained case. The power output and the worst-case distribution in the ambiguity set. Consequently, the re-
consumption of DGs and HFSs in robust case is mostly less than that in sults are more conservative. However, the ambiguity set shrinks and
the chance constrained case. As a result, it leaves more reserves to the approximates the true distribution better by including more historical
microgrid, which also incurs a higher cost. data. As can be seen in Fig. 15, the gap between the objective values of
the DRO method and the SO method becomes smaller as the sample size
5) Influence of the confidence level increases. This reveals that the more data we use, the less conservative
the schedule is. Another phenomenon that can be observed is that the
Fig. 14 further demonstrates the impact of the confidence level on objective value tends to converge to a stable solution as the sample size
the objective function value under different distributions of the un- increases. This is because the proposed ambiguity set approximates the
certainty. Note that the uncertainty only refers to that of the renewable true distribution more accurately as historical data are increasingly
energy power and load here. It is found that the objective function incorporated, which also results in a more stable solution.
value increases as the confidence level increases under all distributions.
This is easy to understand as the chance-constrained inequalities can be 7) Computational time of the studied cases
violated with a lower probability with the increase of the confidence
level. Therefore, the reserve requirement becomes severer and fewer The data handling of chance constraints employs sorting algorithm

11
X. Wu, et al. Applied Energy 253 (2019) 113568

Fig. 12. Comparison of the scheduling results between the chance constrained and deterministic cases.

Fig. 13. Comparison of the scheduling results between the chance constrained and robust cases.

to find the boundary, which is very fast. However, the distributionally As can be observed, the computational time increases as more
robust optimization introduces many scenario variables, which may samples are included. This is mainly because the sample will bring more
increase the computational time. To explore its influence, the compu- variables to the optimization. However, the conservatism can be also
tational time of the studied cases with different sample sizes is also reduced by including more samples. Thanks to the transformation to an
displayed in Table 4. MILP problem, the computational time is still low as only 4 min is

12
X. Wu, et al. Applied Energy 253 (2019) 113568

Overall, we can conclude that the proposed model can effectively


schedule the microgrid with HFSs considering the uncertainties of re-
newable energy power, load and electricity price. The case studies have
validated the proposed model.

5. Conclusions

Faced with increasingly environmental pollution problems, hy-


drogen vehicles and corresponding hydrogen fueling stations have
gained widespread attention in recent years. To reduce the operation
costs, it’s a wise strategy to integrate hydrogen fueling stations with
onsite hydrogen production into microgrids. In this paper, a scheduling
model of a microgrid with hydrogen fueling stations is proposed to
make use of local distributed energy resources and hydrogen fueling
stations to produce hydrogen and fulfill the hydrogen demand of hy-
Fig. 14. Impact of the confidence level on the objective value under different drogen vehicles. The uncertainties of the renewable energy sources,
distributions. loads, and the electricity price are taken into account here. In parti-
cular, a data driven chance constrained approach is presented to handle
the uncertainty of renewable power and load. Some affine policy based
techniques and Monte Carlo method based approach are developed to
transform the chance constraints into tractable mixed integer linear
programming forms. Compared with conventional deterministic and
robust approaches, this method enables the schedule to strike a balance
between the conservatism and optimism. In addition, the developed
approach only utilizes historical data, and thus overcomes the short-
coming of conventional chance constrained approaches which rely on
the probability distribution of the uncertainty. Moreover, the proposed
Wasserstein metric based distributionally robust optimization provides
a risk aversion strategy to cope with the variable electricity price.
However, the conservativeness will shrink as more historical data are
incorporated. As illustrated by case studies, the proposed model not
only schedules the distributed energy resources in an optimal manner
but also arranges the hydrogen production and consumption of hy-
drogen fueling stations economically.
However, in this paper, we only consider the centralized operation,
which lacks the privacy protection mechanism of individuals. In this
regard, a distributed cooperation model of hydrogen fueling stations
Fig. 15. Comparison of the objective value between the DRO and SO.
and microgrids and a fair allocation method of mutual benefits will be
studied in our future work.
Table 4
Computational time of the cases with different sample sizes.
Sample size 10 50 100 500 1000 Acknowledgments

Computational Time (s) 0.4 2 4 59 222


This work was supported by National Natural Science Foundation of
China (51807149), Science and Technology Department of Shaanxi of
required even though 1000 samples are included. China (2017ZDCXL-GY-02-03).

Appendix

Appendix A. . Proof of lemma 3.1

Since the distribution system is radial, then |L |=|B |-1 holds. Disregarding the equality at the root bus and neglecting the subscript “t” in the
LinDistFlow model, the power balance equalities (26) and (27) can be rewritten as the matrix form, as indicated in (64). However, it should be noted
that the constant matrix in (64) is obtained assuming that bus 1 is the balance bus. In addition, line 1 is assumed to be the outgoing line of bus 1, and
line |L | is considered to be the incoming line of bus |B |. It is worth mentioning that the constant matrix here is only for illustration, the actual
constant matrix is closely related to the specific structure of the distribution system. Similarly, constraints (31) and (32) are rewritten as the form of
(65), where the constant matrix is also consistent with the assumptions mentioned above.

13
X. Wu, et al. Applied Energy 253 (2019) 113568

2 B ×1 2 L ×1
   
2 L ×2 B 2 B ×2 B
 ⎡ p1 ⎤ in   ⎡ p1l ⎤
⎡0 1 ⋯ 0 ⎤⎢ ⋮ ⎥ ⎡1 ⋯ 0 ⎤⎢ ⋮ ⎥
⎢⋮ ⋮ ⋱ ⋮ ⎥ ⎢ p in ⎥ ⎢⋮ ⋱ ⋮ ⎥ ⎢pl ⎥
⎢0 0 ⋯ 1 ⎥ ⎢ B ⎥ = ⎢0 ⋯ − 1 ⎥⎢ L⎥
⎢ 0 1 ⋯ 0⎥ ⎢ q in ⎥ ⎢ 1 ⋯ 0 ⎥ ⎢ ql ⎥
⎢ ⎢ 1 ⎥ ⎢ ⎥
⋮ ⋮ ⋱ ⋮⎥ ⎢ ⎥
⎢ ⋮ ⋱ ⋮ ⎥⎢ 1 ⎥
⎢ ⎥ ⋮ ⎢ ⎥ ⋮
⎣ 0 0 ⋯ 1 ⎦ ⎢ in ⎥ ⎣ 0 ⋯ − 1⎦ ⎢ l ⎥
⎢qB ⎥ ⎢qL ⎥ (64)
⎣ ⎦ ⎣ ⎦
2 L ×1
B ×2 L

 
  p l
⎡ 1 ⎤
  
B×B
 B× 1
 ⎡0 ⋯ 0 ⎤⎢ ⋮ ⎥
v − V ⎢⋮ ⋱ ⋮ ⎥⎢ l ⎥
⎡ 1 ⋯ 0 ⎤ ⎡ 1 0 ⎤ ⎢ 0 ⋯ R L V0 ⎥ ⎢pL ⎥
⎢⋮ ⋱ ⋮ ⎥ ⎢ ⋮ ⎥ = ⎢ ⎥⎢ l ⎥
0 ⋯ 0 ⎥ ⎢ q1 ⎥
⎣ 0 ⋯ − 1⎦ ⎢⎣ v B − V0 ⎦
⎥ ⎢
⋮ ⋱ ⋮
⎢ ⎥⎢ ⋮ ⎥
⎢ 0 ⋯ X L V0 ⎥ ⎢ ⎥
⎣ ⎦ ⎢ql ⎥
⎣ L⎦ (65)
(64) and (65) can be further written as (66) and (67) in a compact manner. [PQ]in represents the vector of joint active and reactive power
injection for all buses; [PQ]l denotes the vector of joint active and reactive power flow for all branches; [V ′] represents the vector of the voltage
deviation; A, C, D, F are constant matrixes.
C [PQ]in = A [PQ]l (66)

F [V ′] = D [PQ]l (67)
Since the distribution system is radial, the row vectors in matrix A, which record the connection information of the branches in the network, are
linear independent. Thus matrix A is invertible. In addition, the row vectors in matrix F, which are only related to corresponding buses, are also
linear independent, and we can safely conclude matrix F is invertible as well. Consequently, [PQ]in can be transformed to [V'] as (68):
[V ′] = F −1D [PQ]l = F −1DA−1C [PQ]in (68)
According to (68), the bus voltage is linearly correlated with the bus power injection, and the proof is completed.

Appendix B. . Proof of lemma 4.1

Firstly, (69) is proven as follows:


∼w ∼ld
t {∀ s t ∉ [ s̲ t′ w, s¯t′ w] or ∀ s t ∉ [ s̲ t′ ld, s¯t′ ld]}
∼w ∼w ∼ld ∼ld
= t {s t < s̲ t′ w} + t {s t > s¯t′ w} + t {s t < s̲ t′ ld} + t {s t > s¯t′ ld}
⩽ Stw ( s̲ t′ w ) + 1 − Stw (¯st′ w ) + Stld ( s̲ t′ ld ) + 1 − Stld (¯st′ ld )
= β1 + β2 = β (69)
According to (69), we can conclude (42) is naturally satisfied. As to the procurement of and Stw − 1 (β ) Stld − 1 (β ) ,
the MCM is utilized. Consider a
random variable ∼ τ whose observed sample set is available whereas its exact probability distribution is unknown. Assume the sample set
̂ , τ (2)
S′ = {τ (1) ̂ , ⋯, τ (̂ n) } is ascendingly ordered, Let F(x) =  ∗ {∼ τ ⩽ x } be the CDF of the true distribution  ∗. Then F −1 (β )=
max {τ (̂ k ): τ (̂ k ) ⩽ τ ([̂ βn]), k = 1, ⋯, n} according to the MCM, where “[]” denotes the rounding operator.

References scenarios: a bi-level programming approach via real-time pricing. Appl Energy
2018;232:54–68.
[10] El-Taweel NA, Khani H, Farag HEZ. Hydrogen storage optimal scheduling for fuel
[1] Chan CC. The state of the art of electric, hybrid, and fuel cell vehicles. Proc IEEE supply and capacity–based demand response program under dynamic hydrogen
2007;95:704–18. pricing. IEEE Trans Smart Grid 2018:1.
[2] Kendall K, Kendall M, Liang B, Liu Z. Hydrogen vehicles in China: replacing the [11] Li J, Zhang C, Xu Z, Wang J, Zhao J, Zhang Y-JA. Distributed transactive energy
Western model. Int J Hydrogen Energy 2017;42:30179–85. trading framework in distribution networks. IEEE Trans Power Syst
[3] da Silva Veras T, Mozer TS, da Silva César A. Hydrogen: trends, production and 2018;33:7215–27.
characterization of the main process worldwide. Int J Hydrogen Energy [12] Lahnaoui A, Wulf C, Heinrichs H, Dalmazzone D. Optimizing hydrogen transpor-
2017;42:2018–33. tation system for mobility by minimizing the cost of transportation via compressed
[4] Welder L, Ryberg DS, Kotzur L, Grube T, Robinius M, Stolten D. Spatio-temporal gas truck in North Rhine-Westphalia. Appl Energy 2018;223:317–28.
optimization of a future energy system for power-to-hydrogen applications in [13] Samsatli S, Samsatli NJ. The role of renewable hydrogen and inter-seasonal storage
Germany. Energy 2018;158:1130–49. in decarbonising heat–comprehensive optimisation of future renewable energy
[5] Cai G, Kong L. Techno-economic analysis of wind curtailment/hydrogen produc- value chains. Appl Energy 2019;233:854–93.
tion/fuel cell vehicle system with high wind penetration in China. CSEE J Power [14] Eriksson E, Gray EM. Optimization and integration of hybrid renewable energy
Energy Syst 2017;3:44–52. hydrogen fuel cell energy systems–a critical review. Appl Energy 2017;202:348–64.
[6] Liemberger W, Halmschlager D, Miltner M, Harasek M. Efficient extraction of hy- [15] Zhang W, Maleki A, Rosen MA, Liu J. Optimization with a simulated annealing
drogen transported as co-stream in the natural gas grid–the importance of process algorithm of a hybrid system for renewable energy including battery and hydrogen
design. Appl Energy 2019;233:747–63. storage. Energy 2018;163:191–207.
[7] Nwulu NI, Xia X. Optimal dispatch for a microgrid incorporating renewables and [16] Chen S, Kumar A, Wong WC, Chiu M-S, Wang X. Hydrogen value chain and fuel
demand response. Renew Energy 2017;101:16–28. cells within hybrid renewable energy systems: advanced operation and control
[8] Zhang M, Chen J. The energy management and optimized operation of electric strategies. Appl Energy 2019;233:321–37.
vehicles based on microgrid. IEEE Trans Power Deliv 2014;29:1427–35. [17] Li B, Roche R, Paire D, Miraoui A. Sizing of a stand-alone microgrid considering
[9] Li Y, Yang Z, Li G, Mu Y, Zhao D, Chen C, et al. Optimal scheduling of isolated electric power, cooling/heating, hydrogen loads and hydrogen storage degradation.
microgrid with an electric vehicle battery swapping station in multi-stakeholder Appl Energy 2017;205:1244–59.

14
X. Wu, et al. Applied Energy 253 (2019) 113568

[18] Bakhtiari H, Naghizadeh RA. Multi-criteria optimal sizing of hybrid renewable Trans Sustain Energy 2016;7:625–36.
energy systems including wind, photovoltaic, battery, and hydrogen storage with ɛ- [33] Duan C, Jiang L, Fang W, Liu J, Liu S. Data-driven distributionally robust energy-
constraint method. IET Renew Power Gener 2018;12:883–92. reserve-storage dispatch. IEEE Trans Ind Inf 2018;14:2826–36.
[19] Teng Y, Wang Z, Li Y, Ma Q, Hui Q, Li S. Multi-energy storage system model based [34] Shi Z, Liang H, Huang S, Dinavahi V. Distributionally robust chance-constrained
on electricity heat and hydrogen coordinated optimization for power grid flex- energy management for islanded microgrids. IEEE Trans Smart Grid 2018.
ibility. CSEE J Power Energy Syst 2019;5:266–74. [35] Qiu H, Gu W, Xu Y, Zhao B. Multi-time-scale rolling optimal dispatch for AC/DC
[20] Garcia-Torres F, Bordons C. Optimal economical schedule of hydrogen-based mi- hybrid microgrids with day-ahead distributionally robust scheduling. IEEE Trans
crogrids with hybrid storage using model predictive control. IEEE Trans Ind Sustain Energy 2018.
Electron 2015;62:5195–207. [36] Yao W, Zhao J, Wen F, Xue Y, Ledwich G. A hierarchical decomposition approach
[21] Petrollese M, Valverde L, Cocco D, Cau G, Guerra J. Real-time integration of optimal for coordinated dispatch of plug-in electric vehicles. IEEE Trans Power Syst
generation scheduling with MPC for the energy management of a renewable hy- 2013;28:2768–78.
drogen-based microgrid. Appl Energy 2016;166:96–106. [37] Fan H, Duan C, Zhang C-K, Jiang L, Mao C, Wang D. ADMM-based multiperiod
[22] Rouholamini M, Mohammadian M, Wang C, Gharaveisi AA. Optimal fuzzy-based optimal power flow considering plug-in electric vehicles charging. IEEE Trans
power management for real time application in a hybrid generation system. IET Power Syst 2017;33:3886–97.
Renew Power Gener 2017;11:1325–34. [38] Baran ME, Wu FF. Network reconfiguration in distribution systems for loss reduc-
[23] Cau G, Cocco D, Petrollese M, Kær SK, Milan C. Energy management strategy based tion and load balancing. IEEE Trans Power Deliv 1989;4:1401–7.
on short-term generation scheduling for a renewable microgrid using a hydrogen [39] Wang Z, Chen B, Wang J, Kim J, Begovic MM. Robust optimization based optimal
storage system. Energy Convers Manage 2014;87:820–31. DG placement in microgrids. IEEE Trans Smart Grid 2014;5:2173–82.
[24] “Energy independence and security act of 107,” https://www.gpo.gov/fdsys/pkg/ [40] Ding T, Lin Y, Li G, Bie Z. A new model for resilient distribution systems by mi-
BILLS-110hr6enr/pdf/BILLS-110hr6enr.pdf. crogrids formation. IEEE Trans Power Syst 2017:1.
[25] Xiao W, Cheng Y, Lee W, Chen V, Charoensri S. Hydrogen filling station design for [41] Jabr RA. Adjustable robust OPF with renewable energy sources. IEEE Trans Power
fuel cell vehicles. IEEE Trans Ind Appl 2011;47:245–51. Syst 2013;28:4742–51.
[26] Mehrjerdi H, Bornapour M, Hemmati R, Ghiasi SMS. Unified energy management [42] Lorca Á, Sun XA. Multistage robust unit commitment with dynamic uncertainty sets
and load control in building equipped with wind-solar-battery incorporating elec- and energy storage. IEEE Trans Power Syst 2017;32:1678–88.
tric and hydrogen vehicles under both connected to the grid and islanding modes. [43] Duan C, Jiang L, Fang W, Liu J. Data-driven affinely adjustable distributionally
Energy 2019;168:919–30. robust unit commitment. IEEE Trans Power Syst 2018;1.
[27] Bornapour M, Hooshmand R-A, Khodabakhshian A, Parastegari M. Optimal sto- [44] Duan C, Fang W, Jiang L, Yao L, Liu J. Distributionally robust chance-constrained
chastic coordinated scheduling of proton exchange membrane fuel cell-combined approximate ac-opf with wasserstein metric. IEEE Trans Power Syst
heat and power, wind and photovoltaic units in micro grids considering hydrogen 2018;33:4924–36.
storage. Appl Energy 2017;202:308–22. [45] Esfahani PM, Kuhn D. Data-driven distributionally robust optimization using the
[28] Wu Z, Zeng P, Zhang X-P, Zhou Q. A solution to the chance-constrained two-stage Wasserstein metric: performance guarantees and tractable reformulations. Math
stochastic program for unit commitment with wind energy integration. IEEE Trans Program 2018;171:115–66.
Power Syst 2016;31:4185–96. [46] Baran ME, Wu FF. Network reconfiguration in distribution systems for loss reduc-
[29] Nemirovski A, Shapiro A. Convex approximations of chance constrained programs. tion and load balancing. IEEE Trans Power Deliv 1992:1401–7.
SIAM J Optim 2006;17:969–96. [47] Nojavan S, Zare K, Mohammadi-Ivatloo B. Selling price determination by electricity
[30] Xiang Y, Liu J, Liu Y. Robust energy management of microgrid with uncertain re- retailer in the smart grid under demand side management in the presence of the
newable generation and load. IEEE Trans Smart Grid 2015;7:1034–43. electrolyser and fuel cell as hydrogen storage system. Int J Hydrogen Energy
[31] Bertsimas D, Gupta V, Kallus N. Data-driven robust optimization. Math Program 2017;42:3294–308.
2018;167:235–92. [48] Xiao Y, Wang X, Pinson P, Wang X. A local energy market for electricity and hy-
[32] Wang Z, Bian Q, Xin H, Gan D. A distributionally robust co-ordinated reserve drogen. IEEE Trans Power Syst 2018;33:3898–908.
scheduling model considering CVaR-based wind power reserve requirements. IEEE

15

You might also like