2013 Fourth International Conference on Intelligent Control and Information Processing (ICICIP)
June 9 – 11, 2013, Beijing, China
Optimization Algorithms for Home Energy
Resource Scheduling in presence of data uncertainty
Stefano Squartini # , Matteo Boaro, Francesco De Angelis, Danilo Fuselli, Francesco Piazza
Department of Information Engineering, Università Politecnica delle Marche
Via Brecce Bianche 1, 60131 Ancona, Italy
# s.squartini@univpm.it
Abstract—Smart Home Energy Management is a very hot
topic for the scientific community and some interesting solutions
have also recently appeared on the market. One key issue is
represented by the capability of planning the usage of energy
resources in order to reduce the overall energy costs. This
means that, considering the dynamic electricity price and the
availability of adequately sized storage system, the expert system
is supposed to automatically decide the more convenient policy
for energy management from and towards the grid. In this work
a comparison among different linear and nonlinear methods
for home energy resource scheduling is proposed, considering
the presence of data uncertainty into account. Indeed, whereas
the employment of advanced optimization frameworks can take
advantage by their inherent offline approach, the need to forecast
the energy price and the amount of self-generated power. A
residential scenario, in which a system storage and renewable
resources are available and exploitable to match the user load
demand, has been considered for performed computer simula-
tions: obtained results show how the offline approaches provide
good performance also in presence of uncertain data.
Index Terms—Energy Resource Scheduling, Linear Program-
ming, Particle Swarm Optimization, Adaptive Critic Design,
Electricity Price Forecasting, Solar Irradiation Forecasting.
I. I NTRODUCTION
The concept of smart grid encompasses a large variety of
electrical power engineering issues, and different solutions
can be proposed for each specific application. In this re-
gard, Computational Intelligence has shown to represent a
remarkably powerful technological paradigm to be involved on
purpose [1]. In particular, one relevant application area that has
attracted the attention of many scientists and technicians is the
optimal energy management, which can be applied to system
of different size, from the very large (like big distribution
architectures) [2] to the very small (as in the self-powered
sensor networks) [3]). The authors here refer to the micro-
grid level (like for residential or domestic scenarios). The
main objective consists in reducing costs and thus avoiding
energy waste. To achieve that, several optimization algorithms
have been proposed over the years, and they are typically
oriented to schedule the activities of energy resources and
of the electrical appliances in order to minimize the overall
energy cost due to the grid connection. In this work we
specifically focus on the Energy Resource problem, assuming
that the load is given (or pre-defined by the user) [4], [5]. On
978-1-4673-6249-8/13/$31.00 ©2013 IEEE
purpose, different optimization techniques have been appeared
in the literature so far: linear programming techniques [6],
Particle Swarm Optimization (PSO) [7], Fuzzy-Logic [8],
Artificial Neural Networks [9], and also Adaptive Dynamic
Programming (ADP) [10]-[14].
In this paper, the home energy system is assumed to
be connected to the main grid and also equipped with a
photovoltaic (PV) system and a battery in order to allow some
energy saving. The load requirements must be always satisfied
by suitably managing renewable energy, battery and electrical
grid in order to reduce the costs. Therefore an optimal battery
controller must be obtained, whose control policy is to min-
imize the energy cost imported from the grid managing the
battery actions (charge/discharge) and knowing the forecasted
renewable resources, load profile and energy price. In this
work a comparison among six different methods for battery
management, the best promising chosen from the literature,
is proposed: an overview for each technique is provided and
also a comparison is reported, in terms of advantages and
disadvantages. The same approach has been followed in [17]:
now the presence of data uncertainty is considered, over a
certain time horizon, which makes the optimization task much
closer to real situations but also more difficult to carry out.
In particular, two forecasting problems are taken into account:
one is relative to the electricity price, assuming that it presents
a dynamic behaviour (as in the US electricity market), and
the other regards the solar irradiation. The performance of the
addressed algorithms is then evaluated by suitable computer
simulations under realistic conditions.
The analytical issues related to each optimization algorithm
are discussed in Section II and the simulated home energy sys-
tem is described in Section III. Section IV deal with the data
uncertainty in electricity price and solar irradiation whereas
Section V is about the performed computer simulations and
related results. Section VI draws the work conclusions.
II. O PTIMIZATION A LGORITHMS
A. Linear Programming Technique: LP algorithm
The implemented algorithm is based on the “Linear Pro-
gramming” (LP) paradigm. From a general perspective, its ob-
jective consists in maximizing or minimizing a given function
according to the following constrained scheme:
323
 max f (x) = cT x or min f (x) = cT x
subject to Ax  b or Ax b
where x 0, x 2 Rn⇥1 , A 2 Rm⇥n , b 2 Rm⇥1 , c 2 Rn⇥1 .
The cost function within this optimization framework is
defined as follows:
T ⇥ ⇤
U(t) = Â { L(t) Ru (t) + u(t) ·C(t)}
t=1
where L(t) is the load demand at temporal slot t, Ru (t) is the
amount of renewables exploited at time t, u(t) is the amount
of energy used for charging or discharging the system storage
(therefore it can be positive or negative), C(t) is the electricity
cost at time t and T is the work horizon.
The constraints of the LP problem, for 1  t  T , are
the following (see notation used in Section III for battery
parameters):
• Positiveness of cost function: L(t) Ru (t) + u(t) 0.
• Exploited renewable energy: Ru (t)  R(t) (where R(t) is
the total renewable energy available at time t.
• Charge and discharge limits (assuming a sampling time
of 1 hour): u(t)  Chrate and u(t)  Dhrate , in charging
and discharging conditions repsectively..
• Battery level (SL(t) is the system storage energy at time
t): SL(t) = SL(t 1) + u(t).
• Battery level limits: SLMIN  SL(t)  SLMAX .
B. Particle Swarm Optimization algorithm
PSO is a technique inspired to certain social behaviors,
and it is used to explore a search parameter space to find
values allowing to minimize an objective function [16]. The
PSO algorithm works by maintaining simultaneously various
candidate solutions (particles in the swarm) in the search
space. In PSO, the coordinates of each particle represent a
possible solution associated with two vectors, the position x
and velocity v vectors in N-dimensional search space. A swarm
consists of a number i of particles that fly through the feasible
solution space to find the optimal one. Each particle updates
its position xi on the basis of its own best exploration pi , its
best swarm overall experience pg , and its previous velocity
vector vi (k 1) according to (2) and (3).
xi (k) = xi (k1) + vi (k)
⇥ ⇤
vi (k) = vi (k 1)+r1 · rand1 · pi xi (k 1) +
⇥ ⇤
+ r2 · rand2 · pg xi (k 1)
where r1 and r2 are two positive correction factors, k is the
iteration step while rand1 and rand2 are two random numbers
[0.0, 1.0]. The PSO algorithm can be described in general as
follows:
1) For each particle, randomly initialize the position and
velocity vectors with the same size as the problem
dimension.
2) Measure the fitness (utility function value) of each
particle and store the particle with the best fitness value
(minimum utility function value).
3) Update velocity and position vectors according to (2)
and (3) for each particle.
4) Repeat steps 2 and 3 until a termination criterion is
satisfied.
As already done in [12], [13] we introduce in (4) an utility
(1) function that must be minimized for each temporal slot t.
U(t) =
s
C(t) 2 2
[L(t) R(t) + u(t)] · + SLcap [SL(t) + u(t)]
Cmin
(4)
where u(t) is the optimized value of battery charge (u(t) > 0)
or discharge (u(t) < 0) that must be found by the algorithm for
each time t, SLcap is the battery capacity, SL(t) is the actual
battery level and Cmin the minimum electricity price. Mini-
mizing U(t) means charging the battery when the renewable
contribution is high and/or when cost is low, and discharging
the battery when the available renewable energy is lower than
the load and/or the cost is high. Obviously u(t) must satisfy
the two battery constraints discussed above: if this is not the
case then the obtained solution u(t) is not valid and must be
discarded. So the function is multiplied by a penalty factor
which is typically set to an high value. This battery controller
works in online way, because the cost function is evaluated
step by step without knowing the energy profiles over the work
horizon.
C. Extended Particle Swarm Optimization algorithm
Similar to the scheme proposed in Section II-B, an extended
version of PSO has been realized. The operation is not online
anymore, but offline in order to give an optimal solution on an
extended period, for which all scenario profiles are considered
in the work horizon, as well as the forecasted data about
renewable energy. Differently from (4), the utility function
adopted in this case does not include the battery terms, and
a sum over the entire period is considered, in order to fulfill
the optimization process over the entire work horizon T , as
follows:
T q
2
(2)
U(t) = Â [L(t) R(t) + u(t)] ·C(t) (5)
t=1
D. Adaptive Dynamic Programming
(3)
Combining approximate dynamic programming and rein-
forcement learning, Werbos proposed a new optimization
technique [15], whose goal is to design an optimal control
policy, which can be able to minimize a given cost function
called “utility function” (especially in nonlinear and noisy
environments), adapting two neural networks: the Action Net-
work and the Critic Network. The Action Network, taking the
current state as input, has to drive the system to a desired one,
providing a control to the latter. The Critic Network, knowing
324
the state and the control provided by the Action Network,
checks its performance and return to the Action Network
a feedback signal to reach the optimal state over time. To
check Action performance, the Critic Network approximates
the Bellman equation defined as follows:
• on) because only few actions can be done by the controller, in
J(t) = Â g iU(t + i) (6)
i=0
where g is the discount factor (0, 1] and U(t) is the utility
function.
As already implemented in [12], [13], works inspired by the
one proposed in [10], an Action-Dependent Heuristic Dynamic
Programming (ADHDP) model free approach is adopted for
the design of an optimal controller, whose goal is to manage
the battery, knowing forecasted data (Load, Price, Renewable
Energy), in order to save money during an overall time-
horizon. The input to the Action network is the system state
x(t), and the output u(t) is the amount of energy used to
charge or discharge the battery; the input of the Critic Network
consists of the current system state and the current control
provided by the Action Network.
The used Critic network is composed by 15 linear neurons
in input, 40 sigmoidal hidden neurons and 1 linear in output,
while Action network by 4 linear neurons in input, 40 sig-
moidal hidden neurons and 1 linear in output. In this study
the proposed utility function U(t) is the following:
q ⇥ ⇤ 2
U(t) = L(t) R(t) + u(t) ·C(t) (7)
where u(t) is the optimized value of battery charge (u(t) > 0)
or discharge (u(t) < 0) that must be found for each time t.
Obviously u(t) must satisfy the battery constraints discussed in
Section III. When the utility function is minimized the control
policy is optimal and the cost is the lowest.
The online training is based on the “Backpropagation”
algorithm:
1) The Action and Critic weights are initialized before the
training: with random values [-1,1] or pre-trained with
extended PSO.
2) Train Critic Network refreshing its weights using com-
puting Critic error (Ec ), then refresh Action Network
computing Action error (Ea ).
3) Evaluate the system performance computing the total
cost to minimize in the work horizon. If the cost
decreases, the control policy is improving and the new
action weights are the best; if not, revert to old action
weights and add a small random perturbation. Then
restart the training from Step 2.
As the algorithm optimizes the utility function defined over
all the time horizon T , it can be considered an offline approach.
Moreover, it must be said that the initial computational cost
of the ADHDP algorithm is not very small, but it has the
advantage to adapt itself quite quickly when the time horizon
and the scenario change. Indeed, the optimization process can
continue from the best weights of the neural networks stored
325
on the previous training step, and it does not need to restart
from the beginning like the other proposed methods.
E. Self-Learning Procedure Based on ADHDP Scheme
Like in [11] this optimization procedure is based on a
simplified ADHDP scheme (and named s-ADHDP from now
fact the battery is limited to a ternary choice (charge, discharge
or idle). In this way we consider only a critic network in
the scheme. If a network is trained correctly, whenever power
demand occurs the critic network verifies which is the action
that involves the smallest output value, so the most convenient
action is chosen. The training procedure is the following:
1) Data are collected: the action is taken randomly, the state
is characterized by the cost rate, the load profile, the
battery level and the renewable energy;
2) Compute U(t) and Q(t) in order to obtain the target,
since the training is based on the mapping: {x(t
1); u(t 1)} ! {U(t) + gQ(t)}, where x(t 1) and
u(t 1) are the previous state and control, U(t) is the
actual utility function, g 2 (0, 1] is a discount factor and
Q(t) is the actual critic network output;
3) The critic network is trained with the “Levenberg-
Marquardt backpropagation” algorithm;
4) Eventually the neural network can be re-trained when-
ever there are consistent changes in the scenario.
The utility function that we want to minimize is:
U(t) = [L(t) R(t) + u(t)] ·C(t) (8)
where u(t) = u0 (t)q(t)
is the battery charge (u(t) > 0) or
discharge (u(t) < 0), u0 (t) is the battery action (1, 1, 0) and
q(t) is the charging/dicharging battery quantity. Also in this
case u(t) must satisfy the usual battery constraints. Also this
battery control strategy is offline.
III. S IMULATED H OME E NERGY S YSTEM
The proposed home model is composed of a main electrical
grid, external PV array, storage system and Power Manage-
ment Unity (PMU), that ensures the meeting of load demand.
As reported in Fig. 1, PMU unit (energy scheduler) manages
the energy flows: battery can be charged from the grid and/or
from PV, moreover if necessary it can be discharged to supply
the load. If there is exceeded energy from PV not usable from
the system, it is sold to the main grid. In addition the battery
must satisfy the following constraints:
1) The charging and discharging rate can not be exceeded.
2) Battery level must be always included between the upper
and lower bound.
All the simulations reported in Section V refer to the same
scenario: a system storage is supposed to be available, as
well as renewable resources deriving from solar energy. We
consider an area of 40 m2 covered by photovoltaic (PV) panels,
whose efficiency is 19 %, and irradiation data is taken from
NREL database 1 . The available renewable energy is computed
1 National Renewable Energy Laboratory (NREL) of U.S. Department of
Energy:http://www.nrel.gov/rredc/.
 The battery efficiency has been considered equal to 100%.
Renewable Grid Battery IV. D EALING WITH DATA U NCERTAINTY
Unidirectional Bidirectional A more realistic scenario where the system must work
Power Flow Power Flow with forecasted data is considered in this work, with the aim
to minimize the total energy cost. This is what effectively
Power Management Unit done in this section, where forecasted irradiation profiles and
unitary electricity prices are inserted within the addressed
offline optimization schemes and new computer simulations
Unidirectional
Power Flow
performed to assess their effectiveness.
A. Electricity Price
LOAD
According to the literature [19], [20] the prediction error
that affects the price in a day-ahead scenario is in the worst
case around 12%. A uniformly distributed noise has been
Fig. 1. Power Flows added to the original data related to the electricity price profile
used in simulations, taken from [21], in order to mimic such
a worst case condition. As for the load profile mentioned
with R(t) = SI(t) · h pv · A pv , where SI(t) is the Solar Irradiation above, also these data are relative to the US scenario. They are
(i.e., the amount of direct and diffused solar energy received reported in Fig. 3, where both original and uncertain electrivity
on a horizontal surface during a 60-minute period, expressed price curves are plotted.
inin kW h/m2 ) at time t, h pv is the efficiency of the PV and A pv
is the total area of the PV panel in m2 . The power load profile 8
Electricity Price
Original

considered in our simulated scenario is the same addressed in Forecasted

[11]: it represents a typical US home case study and data are 7

reported in Fig. 2.
6

Power Load
Dollar Cent \ kWh

9 5

8
4

3
6
kW

5 2

4
1
0 10 20 30 40 50 60 70 80 90 100
Hour
3

2
0 10 20 30 40 50 60 70 80 90 100 Fig. 3. Electricity price profiles: original and forecasted data are reported.
Hour

Fig. 2. Power Load profile B. Solar Irradiation

The algorithm used for solar irradiation forecasting is the
In Tab. I the system storage parameters are reported, and one described in [23], where an RBFN (Radial Basis Function
since the resolution time used is one hour, kW h and kW agree Network) is used, i.e. an ANN (Artificial Neural Network)
so we can consider the same unit of measurement both for which uses the RBF (Radial Basis Function) as activation
energy and power parameters. function:
N
TABLE I
BATTERY PARAMETERS ( IN kW ).
ye = f (x) = l0 + Â li f (kx ci k) (9)
i=1

SL0 SLMIN SLMAX SLcap Chrate Dhrate

f (kx ci k) = exp((kx ci k)2 /b ) (10)
5 1 9 10 1 1 where x is the ANN input vector, ci the ith RBF centroid and
li the ith output weights.
In Tab. I SL0 , SLMIN and SLMAX are respectively the initial, These are the main issues related to the algorithm (more
minimum and maximum State of Charge (SoC), SLcap is the details can be found in [23]):
battery capacity, while Chrate and Dhrate are the maximum • Five inputs have been used for solar irradiation data fore-
charge and discharge rate of the considered storage system. casting: Day of the Year, Hour of the Year, Sky Clearness

326
 Index, Environmental Temperature, Wind Speed; the three
latter represents the weather forecasts data which are
obtained by a local weather database 2 ;
• The network has been first pre-trained through the BP
(Back Propagation) algorithm, by using the Photovoltaic
Geographical Information System online data 3 - note that
this operation is typically done when no specific sensors
and measures are available for those PV plants whose
power production wants to be forecasted;
• The training has been performed by means of a joint
Minimal Resource Allocating Network and Adaptive Ex-
tended Kalman Filter algorithm, and using as target the
solar irradiation data related to a PV plant located in Jesi,
Italy, over one entire year (2010).
• Day-ahead solar irradiation forecasts have been per-
formed by using the trained RBF and measured solar
irradiation data in 2011; corresponding weather forecasts
have been used as input.
Note that different meteorological information is used as
ANN input. Solar power production profiles (original and
forecasted), used for our simulations, are reported in Fig. 4. As
pointed out already, the training and testing data are obtained
from italian databases. However, even though electricity price
and power load profiles refer to a US context, the power
production data can be considered as realistic for such a
scenario. This aspect does not affect anyway the comparative
evaluation purposes of the present work for the optimization
algorithms under investigation.
8000
Energy Production
7000
6000
5000
kWh
4000
3000
2000
1000
0
0 10 20 30 40 50 60 70
Hour
Fig. 4. Power production profiles: original and forecasted data are reported.
V. C OMPUTER S IMULATIONS
In this section the performance of aforementioned algo-
rithms are assessed from the money saving perspective. An
heuristic optimization approach is taken here as reference,
namely “baseline approach”. It works as follows:
• if the load is greater than the available renewable energy,
the difference is supplied discharging the battery (accord-
ing with Dhrate in Tab. I). If the battery support is not
enough, the needed energy to supply totally the load is
imported from the main grid;
2 “Il Meteo” website - http://www.ilmeteo.it .
3 PVGIS website - http://www.re.jrc.ec.europa.eu/pvgis .
• if the available renewable energy is greater than the
load demand, the surplus is used to charge the battery
(according with Chrate in Tab. I). If the battery is already
full or the surplus is greater than the charging rate, the
amount of energy in excess, not usable in other ways, is
sold to the main grid.
Four different 24-h time horizons (1-24h, 25-48h, 49-72h,
73-96h) have been considered in our simulations and the
money saving percentage with respect to the energy cost
obtained by means of the baseline approach is reported in
Tab. II and Tab. III, when original and forecasted data are
respectively used.
Looking at the results reported in these Tables, it is ev-
ident that the LP offline algorithm provides the best solu-
tion. Furthermore it has no convergence problems and the
computational cost is very low, which makes it well suited
to implementation on low-power HW/SW platform also in
presence of reduced energy data sampling times. However,
whnever extensions to the model are needed, the linear as-
sumptions behid the LP optimization scheme can represent a
strong limitation and different nonlinear methodologies should
be chosen, like the PSO and the ADHDP based algorithms.
The PSO approach here introduced is an online algorithm
able to work without forecasted data, and it optimizes step
by step a given utility function, with very low computational
complexity. For this reason it is not possible to offer an optimal
solution over a large time horizon, so the cost reduction is
limited. Different is the case of Extended PSO, which gives
a good solution over the considered work horizon, due to its
Original
offline nature.
Forecasted As mentioned, the Extended PSO is used to pretrain the
neural networks used in ADP method. The ADP, adapting the
Action and Critic weights, can improve the performance of
the Extended PSO and find a better solution with an higher
saving. The initial computational cost of the ADP algorithm
is not really low, but it has the advantage to adapt itself
quite quickly when the time horizon and the scenario change.
Finally, the self-learning procedure based on ADHDP scheme
80 90 100
offers a trade off between the goodness of the solution and
the computational cost: slightly sacrificing the cost reduction,
a much shorter time spent for the neural network training can
be obtained.
It is important to remark that the same trend is registered
when the original historical or the forecasted data are used.
This means that, even in presence of data uncertainty, the of-
fline methods allow to sensibly outperfom the online ones and
thus achieving a consistent money saving in the realistic home
energy management system taken as case study. Of course,
the importance of using adequate computational intelligence
algorithm for reliable electricity price and solar irradiation
forecasting is fundamental, and the approach here considered
for the solar power production is a remarkable example.
VI. C ONCLUSIONS
A comparison among different optimization techniques for
the energy resource scheduling in a smart home environment
327
 TABLE II
M ONEY SAVING ( IN PERCENTAGE ) COMPARED TO BASELINE ALGORITHM . Thermal Constraints, IEEE Transactions on Industrial Informatics, in
Historical data case study. press, 2012.
[6] Morais, H., Kádár, P., Faria, P., Vale, Z.A , Khodr, H.M.: Optimal
scheduling of a renewable micro-grid in an isolated load area using
01 24h 25 48h 49 72h 73 96h AV G mixed-integer linear programming, Renewable Energy - An International
Journal, Volume 35, issue 1, pages: 151-156, 2009.
LP 5.2% 5.1% 9.1% 8.5% 7.0% [7] Gudi, N., Wang, L., Devabhaktuni, V., Depuru, S.S.S.R.: A Demand-Side
ADHDP 4.9% 4.7% 8.6% 8.0% 6.6% Management Simulation Platform Incorporating Optimal Management
O f f PSO 4.5% 4.0% 8.0% 7.8% 6.1% of Distributed Renewable Resources. Proceedings of Power Systems
s ADHDP 4.5% 3.7% 7.4% 7.5% 5.8% Conference and Exposition (PSCE), pp.1-7, 2011.
On PSO 0.8% 1% 2.0% 1.5% 1.3% [8] Liang, R.H., Liao, J.H.: A Fuzzy-Optimization Approach for Generation
Scheduling with Wind and Solar Energy Systems. IEEE Transactions on
TABLE III Power Systems, Volume 22, Issue 4, pp. 1665-1674, 2007.
M ONEY SAVING ( IN PERCENTAGE ) COMPARED TO BASELINE ALGORITHM . [9] Vale, Z.A., Faria, P., Morais, H., Khodr, H.M., Silva, M., Kadar,
Forecasted data case study. P.: Scheduling Distributed Energy Resources in an Isolated Grid: An
Artificial Neural Network Approach, IEEE Power and Energy Society
General Meeting pp. 1-7, 2010.
01 24h 25 48h 49 72h 73 96h AV G [10] Welch, R.L., Venayagamoorthy, G.K.: Energy Dispatch Controllers for a
Photovoltaic System, Engineering Applications of Artificial Intelligence,
LP 3.5% 3.7% 7.1% 7.0% 5.3%
pp. 249-261, 2008.
ADHDP 3.1% 3.4% 6.6% 6.5% 4.9%
[11] Huang, T., Liu, D., Residential Energy System Control and Management
O f f PSO 2.4% 3.0% 5.7% 6.2% 4.3%
using Adaptive Dynamic Programming, Proceedings of Intenational
s ADHDP 2.4% 3.0% 5.4% 5.7% 4.2%
Joint Conference on Neural Networks (IJCNN), pp. 119-124, 2011.
On PSO 0.8% 1% 2.0% 1.5% 1.3%
[12] Fuselli, D., De Angelis, F., Boaro, M., Liu, D., Wei, Q., Squartini,
S., Piazza, F.: Optimal Battery Management with ADHDP in Smart
Home Environments, Advances in Neural Networks - ISNN 2012, LNCS
Springer, Volume 7368, 2012.
has been addressed in this work, and an evaluation from a [13] Fuselli, D., De Angelis, F., Boaro, M., Squartini, S., Liu, D., Wei, Q.,
monetary perspective has been provided in order to highlight Piazza, F.: Action Dependent Heuristic Dynamic Programming for Home
the performance in comparison with a baseline method based Energy Resource Scheduling, International Journal of Electrical Power
and Energy Systems, Volume 48, pp.148–160, 2013.
on heuristics. In particular, the computer simulations have been [14] M. Boaro, D. Fuselli, F. De Angelis, D. Liu, Q. Wei, and F. Piazza,
carried out considering the fact that the electricity price and the Adaptive Dynamic Programming Algorithm for Renewable Energy
solar irradiation profiles over a 24-hour horizon are not a-priori Scheduling and Battery Management, Cognitive Computation, 2012.
[15] Werbos, P.J.: Approximate Dynamic Programming for Real-Time Con-
known can be forecasted. Indeed, most of the investigated trol and Neural Modeling. Handbook of Intelligent Control, 1992.
optimization techniques operate offline and rely on day-ahead [16] Del Valle, Y., Venayagamoorthy, G.K., Mohagheghi, S., Hernandez, J.C.,
data which are by nature uncertain. Harley, R.G.: Particle Swarm Optimization: Basic Concepts, Variants
and Applications in Power Systems, IEEE Transactions on Evolutionary
Experimental results have shown that, even if a decrease Computation, Volume 12, Issue 2, pp. 171-195, 2008.
of performace is registered when forecasted data are used, all [17] F. De Angelis, M. Boaro, D. Fuselli, S. Squartini, F. Piazza, A Compar-
offline methods significantly outperform the online ones, and ison Between Different Optimization Techniques for Energy Scheduling
in Smart Home Environment, in Neural Nets and Surroundings -
specially the baseline approach. The LP algorithm offers the 22nd Italian Workshop on Neural Nets, WIRN 2012, Apolloni, Bassis,
best solution, and also a very low complexity if compared to Morabito, Esposito Eds., Springer series in Smart Innovation, Systems
the others. However its linearity assumption does not make it and Technology, 2012.
[18] C. Unsihuay-Vila, A. Zambroni de Souza, J. Marangon-Lima, and P.
well suited for extensions. This is clearly not the case for Balestrassi, Electricity demand and spot price forecasting using evolu-
the ADHDP algorithm which give very close performance tionary computation combined with chaotic nonlinear dynamic model,
in the present case study. Indeed, as future works, more International journal of electrical power and energy systems, vol. 32,
no. 2, pp. 108-116, 2010.
complex residential scenarios for optimal energy management [19] Conejo, A.J., Plazas, M.A., Espinola, R. and Molina, A.B., Day-ahead
will be taken into consideration, by including, on one hand, electricity price forecasting using the wavelet transform and ARIMA
complementary energy resources (like wind turbines or micro- models, IEEE Transactions on power systems, vol. 20, no. 2, pp. 1035-
1042, 2005.
chp systems) and, on the other, the battery size issue [24]. [20] Li, G., Liu, C.-C., Mattson, C. and Lawarree, J., Day-ahead electricity
price forecasting in a grid environment, IEEE Transactions on power
R EFERENCES systems, Vol. 22, N. 1, p.266, vol. 22, no. 1, pp. 266-274, 2007.
[21] Lee, T.Y., Operating schedule of battery energy storage system in a
[1] Venayagamoorthy, G.K.: Potentials and Promises of Computational time-of-use rate industrial user with wind turbine generators: a multi-
Intelligence for Smart Grids. IEEE Power and Energy Society General pass iteration particle swarm optimization approach, IEEE Transactions
Meeting, pp. 1-6, 2009. Energy Convers, vol. 22, no. 3, pp. 774–782, 2007.
[2] M. Kamh and R. Iravani. A sequence frame-based distributed slack bus [22] Marquez, M. and Coimbra, C., Forecasting of global and direct solar
model for energy management of active distribution networks. Smart irradiance using stochastic learning methods, ground experiments and
Grid, IEEE Transactions on, 3(2):828–836, 2012. the NWS database, Solar Energy, vol. 85, no. 5, pp. 746-756, 2011.
[3] M. Severini, S. Squartini and F. Piazza, Energy Aware Lazy Scheduling [23] L. Ciabattoni, M. Grisostomi, G. Ippoliti, S. Longhi, Solar irradiation
Algorithm for Energy-Harvesting Sensor Nodes, Neural Computing and Forecasting for PV System by Fully Tuned Minimal RBF Neural
Applications, in press, 2012. Network, in Neural Nets and Surroundings - 22nd Italian Workshop
[4] De Angelis, F., Boaro, M., Fuselli, D., Squartini, S., Piazza, F., Wei, Q., on Neural Nets, WIRN 2012, Apolloni, Bassis, Morabito, Esposito Eds.,
Ding, W.: Optimal Task and Energy Scheduling in Dynamic Residential Springer series in Smart Innovation, Systems and Technology, 2012.
Scenarios, Advances in Neural Networks - ISNN 2012, LNCS Springer, [24] S. Squartini, D. Fuselli, M. Boaro, F. De Angelis, F. Piazza, Home
Volume 7368, pp. 650-658, 2012. Energy Resource Scheduling Algorithms and their dependency on the
[5] F. De Angelis, M. Boaro, D. Fuselli, S. Squartini, F. Piazza, Q. Wei, Battery Model, Proceedings of IEEE Symposium Series on Computa-
Optimal Home Energy Management under Dynamic Electrical and tional Intelligence, Singapore, 2013, to appear.

328