Accepted Manuscript

Optimal load dispatch of community microgrid with deep learning based solar
power and load forecasting

Lulu Wen, Kaile Zhou, Shanlin Yang, Xinhui Lu

PII: S0360-5442(19)30077-5

DOI: 10.1016/j.energy.2019.01.075

Reference: EGY 14548

To appear in: Energy

Received Date: 10 September 2018

Accepted Date: 14 January 2019

Please cite this article as: Lulu Wen, Kaile Zhou, Shanlin Yang, Xinhui Lu, Optimal load dispatch of
community microgrid with deep learning based solar power and load forecasting, Energy (2019),
doi: 10.1016/j.energy.2019.01.075

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Optimal load dispatch of community microgrid with deep learning based

solar power and load forecasting

Lulu Wen a,b, Kaile Zhou a,b,c*, Shanlin Yang a,b*, Xinhui Lu a,b

a. School of Management, Hefei University of Technology, Hefei 230009, China

b. Key Laboratory of Process Optimization and Intelligent Decision-making of Ministry of Education, Hefei

University of Technology, Hefei 230009, China

c. City University of Hong Kong, Kowloon, Hong Kong SAR, China

Abstract: A deep recurrent neural network with long short-term memory units (DRNN-LSTM)

model is developed to forecast aggregated power load and the photovoltaic (PV) power output in

community microgrid. Meanwhile, an optimal load dispatch model for grid-connected community

microgrid which includes residential power load, PV arrays, electric vehicles (EVs), and energy

storage system (ESS), is established under three different scheduling scenarios. To promote the

supply-demand balance, the uncertainties of both residential power load and PV power output are

considered in the model by integrating the forecasting results. Two real-world data sets are used to

test the proposed forecasting model, and the results show that the DRNN-LSTM model performs

better than multi-layer perception (MLP) network and support vector machine (SVM). Finally,

particle swarm optimization (PSO) algorithm is used to optimize the load dispatch of grid-connected

community microgrid. The results show that EES and the coordinated charging mode of EVs can

promote peak load shifting and reduce 8.97% of the daily costs. This study contributes to the optimal

* Corresponding authors.
E-mail addresses: [email protected] (K. Zhou), [email protected] (S. Yang).
1
 ACCEPTED MANUSCRIPT

load dispatch of community microgrid with load and renewable energy forecasting. The optimal

load dispatch of community microgrid with deep learning based solar power and load forecasting

achieves total costs reduction and system reliability improvement.

Keywords: Optimal load dispatch; community microgrid; load forecasting; solar power; deep

learning

1. Introduction

Energy and environmental issues are two major challenges for global sustainable development

[1]. In most countries, power supply still mainly relies on the combustion of traditional fossil fuels

which has brought serious environmental problems [2]. At the same time, electricity consumption

has increased significantly with the explosive growth of population and the rapid development of

economy, especially for the residential electricity consumption which shows an increasing trend and

has approximately accounted for 18.68% of total electricity consumption in China by 2014 [3].

In recent years, a new generation of smart grid system has been proposed and developed rapidly

to deal with these challenges [4]. As a part of smart grid, microgrid is a cluster of various distributed

generators (DGs), energy storage system (ESS), loads, and other monitoring and protection devices

[5]. A microgrid can meet more flexible and personalized power demand of a local area. By

interacting with the main power grid, it can significantly improve the energy efficiency [6].

Meanwhile, it can also promote the local consumption of distributed energy resources, thus

alleviating environmental pressures. According to whether microgrid interact with the main power

grid or not, there are mainly two operation modes, i.e. islanded mode and grid-connected mode [7].

With the wide deployment of intermittent renewable energy and the integration of increasingly
2
 ACCEPTED MANUSCRIPT

diverse loads in microgrid, microgrid energy management has drawn great attention [8]. Microgrid

energy management aims at promoting supply-demand balance, reducing operational costs and

pollutant emissions, and improving power system stability through the optimal dispatch of power

generation resources and demand-side loads etc. [9, 10]. The simplest dispatch model of community

microgrid only includes DGs (e.g. micro turbine and diesel generator) and loads [6]. Currently, the

studies on optimal load dispatch of community microgrid are becoming increasingly complex to

meet actual needs by considering more components. In order to reduce pollutant emissions caused

by conventional power generation and integrate renewable energy more efficiently, several

researches started to explore the community microgrid which contains renewable energy resources

[11, 12]. Meanwhile, with the development of battery related technologies, the cost of energy

storage is falling [13]. More and more studies attempt to explore the optimization of community

microgrid that contains ESS and electric vehicles (EVs) [14, 15].

ESS can smooth the output of renewable energy resources and shift the peak load in community

microgrid, thereby contributing to improve the stability of power system and reduce pollutant

emissions [16]. It has been an essential object in community microgrid load dispatch. In Ref. [14],

the authors evaluated the economics of providing demand response under a realistic tariff using a

range of storage technologies. The dynamic economic emission dispatch problem of microgrid

incorporating energy storage system was further explored in Ref. [17]. In order to integrate

renewable energy flexibly, Ansari et al. [18] investigated the optimal scheduling of battery storage

in distribution system with high photovoltaic (PV) arrays penetration. As for EVs, they are

important components in community microgrid for optimizing load dispatch, as they are regarded

as an effective way to reduce electricity consumption cost and improve stability of power system

3
 ACCEPTED MANUSCRIPT

through reasonable arrangement of charging and discharging activities. In Ref. [19], a novel

charging management strategy was proposed to reduce the negative effects of EVs on power system

stability. Moreover, Liu et al. [20] developed a load dispatch model to integrate the renewable

energy source using EVs as an ESS. There are also some research efforts that have focused on the

load dispatch optimization of community microgrid that contains plug-in hybrid EVs [21, 22].

However, few works have researched the community microgrid which contains power load,

renewable energy resources, EVs, and ESS simultaneously to meet future needs in more complex

context.

The inherent randomness and volatility of renewable energy pose many challenges to the load

dispatch of community microgrid. Therefore, a variety of studies has considered the intermittence

of renewable energy in the load dispatch model of microgrid [12, 23]. For instance, in Ref. [12], the

authors presented a robust optimization method considering disturbances of solar power for

scheduling microgrid economically. At the same time, some data driven methods have been used to

forecasting renewable energy, and thus to determine the uncertainty and provide the insights for

scheduling power load [23, 24]. However, these papers just considered the uncertainty of renewable

energy resources (i.e. supply side) in microgrid. The uncertainty and volatility of power load (i.e.

demand side) which can affect supply-demand balance, are rarely taken into account.

In traditional load dispatch model of community microgrid, residential power load is usually

given directly. Few studies considered the uncertainty and volatility of load demand which plays an

important role in promoting interaction and balance between supply and demand. Accurate load

forecasting which can present the load variation over a period in future, has provided the possibility

for matching supply and demand in community microgrid. Several methods, such as support vector

4
 ACCEPTED MANUSCRIPT

machine (SVM), artificial neural network (ANN), statistical regression (SR), grey model (GM),

decision tree (DR), and evolution algorithm (EA), have been used for load forecasting [25-27]. For

instance, in Ref. [28], the authors predicted air conditioning load using a SVM model with

parameters obtained through modified simulated annealing optimization. A decision tree was

employed to forecast building energy demand levels with various variables in Ref. [29]. In addition,

Kaboli et al. [30] presented an optimized gene expression programming to forecast long-term

electricity consumption. There are also some research efforts focused on the load forecasting using

SR and GM [31-33]. Nevertheless, these methods have certain limitation in learning the internal

relationship between residential power load and other factors, as well as in achieving more accurate

forecasting results. Since ANN can model more complex relations between power load and related

variables compared with other methods, it has been wildly used in the forecasting of residential

power load [34, 35]. However, general neural networks usually have weak learning ability and

cannot consider the time dependencies in power load profile, thus failing to achieve the best

forecasting results. Therefore, it is necessary to develop a deep neural network model which

performs better in feature representation and mapping as well as considers the time dependencies.

Overall, few of previous studies considered the uncertainties of renewable energy resources and

power load simultaneously. Besides, current forecasting methods have difficulty in achieving more

accurate forecasting results. To promote the balance and interaction between supply and demand

sides, both the volatility of renewable energy resources and power load should be taken into account

in the optimal load dispatch of community microgrid. Also, it is necessary to develop effective and

efficient method for the forecasting of renewable energy output and demand-side power load.

To bridge this gap, the optimal load dispatch model of a grid-connected community microgrid

5
 ACCEPTED MANUSCRIPT

was developed based on the forecasting of PV power output and residential power load. The key

contributions of this study are as follows.

(1) A deep recurrent neural network with long short-term memory units (DRNN-LSTM) was

developed to forecast power load of aggregated residential buildings and PV power output during a

short-term period.

(2) The optimal load dispatch model of a grid-connected community microgrid which includes

residential power load, PV arrays, EVs, and ESS, was established under three different scheduling

scenarios.

(3) The uncertainties of the PV power output and residential power load were considered in the

optimal load dispatch model simultaneously by using the forecasting results of the DRNN-LSTM

model.

The remainder of the study is organized as follows. Section 2 introduces the related methodology.

Section 3 provides the detailed description of the DRNN-LSTM model for forecasting residential

power load and PV power output and the optimization model for finding the optimal load dispatch

of community microgrid. The experimental results and discussions are given in Section 4. Section

5 presents the conclusions.

2. Methodology

2.1 Artificial neural networks (ANN)

ANN abstracts the information processing of human brain neural network, and establishes some

non-linear models to learn the relationships between input vectors and output values [36]. Different

connection modes will form different neural networks. A simple feedforward network with single

hidden layer is shown in Fig. 1.

6
 ACCEPTED MANUSCRIPT

… … output

Hidden
…
layer

… … Input

Fig. 1. A simple feedforward network with single hidden layer.

The output from a given node j in layer i can be expressed as

n
y j  f ( kj yk   j ) (1)
k 1

where f is activation function, kj is the weight parameter between node j in layer i and node k

in layer i-1, yk is the output from node k in layer i-1,  j is the bias of node j, n is the total

number of nodes in layer i-1.

When all the outputs are obtained, the error e is calculated as

e  ( y p  ya ) 2 (2)

where y p is the forecasting value and ya is the actual value.

Then, all the samples are processed in the training phase, and all the weight parameters are

updated as

e
kj  kj   (3)
kj

where  is learning rate.

2.2 Recurrent neural network (RNN) and Long Short-term Memory (LSTM)

Recurrent neural network (RNN) is neural sequence model that can model temporal dependencies

presented in time series data [37]. There are some ring structures in RNN, in order to remember

information at previous time steps. Therefore, RNN is capable of dealing with dynamical problems

7
 ACCEPTED MANUSCRIPT

of time series. For example, RNN has been widely applied in language modeling, handwriting

recognition, and speech recognition [38, 39]. In a simple RNN, the output at time t is determined by

both the input at this time step and the output at prior time step, and it can be expressed as

y ( t )  f (1 x ( t )  2 h ( t 1)   ) (4)

where f is activation function, 1 and 2 are weight parameters corresponding to different

time steps,  is the bias parameters, x (t ) is the input at time step t, h(t 1) is the output at time step

t-1, and y (t ) is the output at time step t.

However, RNN always consumes long time in the training through back-propagation. Besides,

the gradients are tend to vanish or explode over long time period, thus making the learning of long-

term dependencies difficult [40]. To address these problems, long short-term memory (LSTM) was

proposed to replace the conventional neuron in RNN. LSTM can adaptively scales the input and

output value, remembers or forgets the cell state value [41]. All these are achieved by using input,

forget and output gates respectively. The schematic of a LSTM unit is shown in Fig. 2

ht

Ct-1 Ct
× +
tanh
ft it pt
× ×
C't
σ σ tanh σ
ht-1 ht

Xt

Fig. 2. Schematic of a LSTM.

When the input X t enters LSTM, it passes through an activation function  which outputs ft .

f t   ( f  [ yt 1 , X t ]   f ) (5)

In (5),  f is weight parameter,  f is bias parameter, and yt 1 is the output from prior LSTM

unit. Then, the forget gate will discard some information from the transient cell state, and the input

8
 ACCEPTED MANUSCRIPT

gate will identify what new information remains in the cell state and update the cell state through

following functions.

it   (i  [ yt 1 , X t ]   i ) (6)

Ct  tanh(c  [ yt 1 , X t ]   c ) (7)

Ct  f t  Ct 1  it  Ct (8)

where i and c are weight parameters, i and c are bias parameters, it is the vector that

determine what value will be updated, Ct is the new candidate vector, Ct 1 is the cell state of

previous LSTM unit, and Ct is the new cell state. At last, yt will be output by two activation

functions in output gate, and it can be expressed as

zt   ( p  [ yt 1 , X t ]   p ) (9)

yt  zt  tanh(Ct ) (10)

where  p is weight parameter,  p is bias parameter, zt is a part of cell state that will be

output.

2.3 Particle swarm optimization (PSO)

Population based algorithm is a computational technology based on the law of biological group

behavior and is usually used to solve distributed problems [42]. As an important population based

algorithm, PSO has been widely used to deal with the load dispatch of microgrid [43]. It starts from

the random solution to find the optimal solution by iteration, and evaluates the quality of solutions

by their fitness [44]. Compared with genetic algorithm, PSO algorithm has fewer adjustable

parameters, thus making the algorithm more efficient [45, 46]. Since PSO is of strong capability

and high efficiency in solving optimization problems, this study uses PSO algorithm to deal with

the proposed optimal load dispatch model.

9
 ACCEPTED MANUSCRIPT

In PSO algorithm, each solution of the optimization problem can be described as a particle which

has a position vector and a velocity vector. Position vector represents a possible solution of the

optimization problem, and velocity vector determines the direction and magnitude of the change in

position. Position vector and velocity vector are adjusted dynamically according to previous best

results. Besides, there is a fitness function to estimate the present position.

Suppose there are m particles, the position of particle i can be expressed as X i  ( xi ,1 , xi ,2 , xi , D )

in D dimensional space, where 1  i  m . Vi  (vi ,1 , vi ,2 , vi , D ) represents the velocity of particle i.

Generally, the position and velocity of particles are taken in a continuous real space. The best

positions in the history of particle i are described as pi  ( pi1 , pi 2 , piD ) , and

p j  ( p j1 , p j 2 , p jD ) denotes the best positions of all particles, where 1  j  m . Then the

velocity and the position of particle i from iteration k to iteration k+1 can be updated according to

ViD  ViD  c1r1 ( piD  xiD )  c2 r2 ( p jD  xiD )

k 1 k k k k k

 k 1 k 1
(11)
 X iD  X iD  ViD
k

where r1 and r2 are random numbers that are distributed in [0, 1] uniformly, c1 and c2 are

learning parameters which represent the self-learning ability and social learning ability of particle

respectively, and  is inertia weight parameter. Small inertia weight parameter contributes to

convergence of PSO algorithm, and large inertia weight parameter is beneficial to jump out of local

optimal and facilitate global search [47]. Therefore, the selection of learning parameter and inertia

weight parameter is of great importance for the performance of PSO algorithm.

3. Model

This study developed a DRNN-LSTM model to forecast residential hourly power load and the

hourly PV power output over short-term horizon respectively. Then, an optimal load dispatch model

10
 ACCEPTED MANUSCRIPT

of a grid-connected community microgrid which contains residential power load, PV arrays, EVs,

and ESS, was established under different scheduling scenarios. In the load dispatch model, the

residential power load and the PV power output were obtained from the forecasting results of the

DRNN-LSTM model.

3.1 Forecasting model

3.1.1 Model inputs

Depending upon the object of forecasting, the input of the model is a given combination of

schedule variables, weather variables, and timescale variable.

The schedule variables considered in the DRNN-LSTM model include hour of day (0-23), day of

week (1-7), day of month (1-31), and month number (1-12). It is obvious that residential power load

has high correlation with schedule variables. For example, the power load is low over daytime of

weekday as residents go to work. In contrast, the power load is high during weekend when residents

spend much time staying at home. Since PV arrays can only generate electricity under sunlight, the

PV power output is more vulnerable to schedule variables.

The weather related variables in forecasting residential power load are different from those in

forecasting the PV power output. Previous study has shown that weather conditions, especially

temperature, has impact on power load [31]. In this study, the weather variables for forecasting

residential power load include temperature, humidity, and wind speed. Except these weather

variables, global horizontal radiation, and diffuse horizontal radiation are considered in forecasting

the PV power output.

As for timescale variable, 24-h was selected as the reference sequence length of inputs. It is

because that routine pattern of residential power load correspond to 24-h. In addition, the PV power
11
 ACCEPTED MANUSCRIPT

output is most pertinent on a daily timescale.

3.1.2 Model structure

The residential power load reflects the operation state of household appliances which are of high

instantaneity and volatility, thus making power load profile transient and non-linear. Therefore, the

forecasting model should be capable to model non-linear and regularity of power load profile, as

well as consider the time dependency existed in the data.

Compared with residential power load profile, PV power output profile shows certain periodicity

and linearity. However, PV power output is vulnerable to weather conditions. To promote the

integration of solar power and reduce waste, more accurate forecasting is required.

The DRNN-LSTM model was proposed to overcome these challenges. The proposed forecasting

model was a hybrid model which is consisted of input layer, RNN-LSTM layers, general perception

layers, and output layer. It has been proved that hybrid model often outperforms non-hybrid model

[40]. In order to achieve the best forecasting results on residential power load and PV power output,

the model adopted a six-layer network structure, including an input layer, two RNN-LSTM layers,

two simple hidden layers, and an output layer. Fig. 3 presents the schematic of the proposed DRNN-

LSTM model.

…… Layer 6

Two Hidden Layer Layer 4,5

…… Layer 3

…… Layer 2

…… Layer 1
t=1 t=2 t=3 t=24

Fig. 3. Schematic of the DRNN model.

12
 ACCEPTED MANUSCRIPT

The structure of the DRNN-LSTM model is described as follows. (i) Layer 1: The input at one-

hour resolution is introduced in layer 1. Each circle represents an input, including schedule variables

and weather variables at that time step. The training data contains S samples. S is the number of

24-h sequences (i.e. the total number of days in the training data). (ii) Layer 2 and Layer 3: They

are two RNN-LSTM layers, and each circle denotes an output vector. (iii) Layer 4, Layer 5, and

Layer 6: Layer 4 and Layer 5 are two simple hidden layers, and these two layers and Layer 6

correspond to a multi-layer feedforward network.

3.1.3 Model regulations

In the training of the DRNN-LSTM forecasting model, some regulations were made to improve

model performance. The following regulations were adopted in this study.

To eliminate the influence of dimension and accelerate training process, the training data was

converted into scalar value by min-max normalization which can be expressed as

xi  min
xi  (12)
max  min

where xi is the original value, min and max are minimum value and maximum value of the variable

where xi is located.

Deep neural network has suffered from overfitting, thereby causing serious errors in testing phase.

Weight decay regularization is an effective method to avoid overfitting by adding regularization

term to cost function [48]. It can weaken the impact of perturbations and noise existed in training

data. The error function can be expressed as


e  e( w)  wT w (13)
2S

where e is the original error function, w is the weight vector, and  was set as 0.01 which is

default in Keras [49].

13
 ACCEPTED MANUSCRIPT

Dropout is another way to overcome overfitting. It deletes a certain proportion of neural units in

neural network layers to change the network structure. Hence, it can adjust the learning ability of

neural network model adaptively, and learn more robust characteristics [50].

The Adam was used to optimize the weight parameters in each layer [51]. It is known that

conventional gradient descent algorithm takes a lot of time and computing resource in training phase.

Although distributed parallel training can accelerate model learning, the computational resources

needed are not reduced. Compared with conventional gradient descent algorithm, Adam algorithm

has fast convergence speed and is well suitable for dealing with problems that involve large datasets

and many parameters [51]. In addition, it designs an independent adaptive learning rate for different

parameters and requires little memory.

The setting of hyper-parameters will affect the performance of machine learning algorithms. They

are usually assigned empirically or determined by the model structure and data characteristics. The

hyper-parameters in the DRNN-LSTM model are shown as following: (i) Length of output from

Layer 2, Layer 3, Layer 4, and Layer 5; (ii) Activation function of Layer 4, Layer 5, and Layer 6;

(iii) Type of neuron nodes (i.e. LSTM or GRU) in Layer 2 and Layer 3; (iv) Dropout of Layer 2,

Layer 3, Layer 4, and Layer 5. To optimize the hyper-parameters, a global method is required.

Hyperopt, a python library for hyper-parameters optimization, provides an effective way to optimize

hyper-parameters in machine learning by parallel schemes [52]. It is a sequential model-based

optimization method which presents an optimized interface that accepts an evaluation function and

parameters space. Given the parameters space, the optimal structure for the DRNN-LSTM model

will be found automatically.

3.1.4 Model evaluation

14
 ACCEPTED MANUSCRIPT

To evaluate the performance of the proposed DRNN-LSTM model, four common metrics were

used.

(1) Root Mean Square Error (RMSE). RMSE is the root mean square error between actual value

and forecasting value in testing phase.


T
t 1
( ya  y p ) 2
RMSE= (14)
T

where ya is the actual value in testing phase, y p is the corresponding forecasting value, and T

is the total number of time steps.

(2) Mean Absolute Error (MAE). MAE represents the mean absolute error between actual value

and forecasting value in testing phase, and it can be expresses as follows.


T
ya  y p
MAE  t 1
(15)
T

The meanings of all the variables in Eq. (15) are the same as those in RMSE.

(3) Mean absolute percentage error (MAPE). MAPE denotes the mean absolute error in testing

phase, and the smaller value of MAPE represents better forecasting performance.

100% ya  y p
 t 1 y (16)
T
MAPE 
T a

The meanings of all the variables in Eq. (16) are the same as those in RMSE.

(4) Pearson Correlation Coefficient (PCC). PCC reflects the correlation between actual values

and forecasting values in testing phase.

E[( ya   ( ya ))( y p   ( y p ))]

PCC  (17)
y y a p

The meanings of all the variables in Eq. (17) are the same as those in RMSE.

3.2 Optimization model

In this section, the optimal load dispatch model of grid-connected community microgrid which

15
 ACCEPTED MANUSCRIPT

contains power load, PV arrays, EVs, and ESS, was established. Fig. 4 shows the structure of the

grid-connected community microgrid.

. . . ESS
PV power

Main
power
grid . . .
.
.
. . . .

. . .

Residential Community

Fig. 4. Structure of the grid-connected community microgrid model.

3.2.1 Electric vehicles charging model

The EVs in the grid-connected community microgrid, can act as a load or a power source by

charging or discharging to join the load dispatch of microgrid. The daily driving distance, the

starting charging time, and charging load of EVs are important factors for electric vehicles to access

microgrid.

The daily running distance of EV obeys lognormal distribution (i.e. distance : log  N ( 1 ,  12 ) )

which can be expressed as

1 (ln x1  1 ) 2 (18)
f distance ( x1 )  exp( )
2 1 x1 2 12

where 1  3.2 ,  1  0.88 , and x1 is the daily running distance [53]. The charging demand of EV

is determined by driving distance. At the same time, it was assumed that charging electricity of EV

is equal to daily electricity consumption in this study. Fig. 5 presents the probability distribution of

the daily driving distance of EVs.

16
 ACCEPTED MANUSCRIPT

Fig. 5. Probability distribution of the daily driving distance of EVs.

It can be seen from Fig. 5 that daily driving distance of an EV is mainly between 0 km and 100

km.

Since the microgrid is a grid-connected community microgrid, the charging situation in which

EVs owners start charging when arriving home in the evening according to their daily routine, was

studied. The last return time of an EV was defined as a normal distribution (i.e. t : N ( 2 ,  22 ) ) which

can be expressed as

 1 ( x  24   2 ) 2
 exp( 2 ), 0  x2   2  12
 2 2 x2 2 22
f t ( x2 )   (19)
 1 ( x   )2
exp( 2 2 2 ),  2  12  x2  24
 2 x 2 2
 2 2

where  2  17.6 ,  2  3.4 , and x2 is the last return time of an EV [53]. The probability

distribution of starting charging time of an EV under above definitions is shown in Fig. 6.

Fig. 6. Probability distribution of starting charging time of an EV.

The daily starting charging time and driving distance of EV are independent mutually, then the

charging load profiles of EV can be calculated by a random experiment. In this study, the total

charging load of EVs are obtained by Monte Carlo simulation which can produce charging load

17
 ACCEPTED MANUSCRIPT

profile for each EV.

At the same time, there are some constraints for the EVs charging model, including capacity

constraints and charge/discharge constraints.

(1) Capacity constraints of EVs. The ratio of the residual electricity to the rated capacity can be

used to estimate the charge state (CS) of EVs. The CS must be in a certain range in order to protect

the batteries.

CSimin  CSi  CSimax (20)

where CSi is the CS of i-th EV, CSimin represents the lower limits, and CSimax represents the

upper limits.

(2) Capacity constraints of EVs when leaving the microgrid. In order to ensure that EVs have

enough power for routine running, the following presents another constraint.

CSti  CStmin
i
(21)

min
where CSt is the CS of i-th EV when it leaves microgrid, and CSti
i
is the lower limit for routine

running of EVs.

(3) Charge/discharge constraints of EVs. To prevent the damage caused by high power of

charging or discharging, the charge/discharge constraint must be met [54].

Pi min  Pi EV  Pi max (22)

where Pi EV is charging or discharging power of the i-th EV, Pi min denotes the lower limits, and

Pi max denotes the upper limits.

3.2.2 Objective function

The objective of the optimal load dispatch model is to minimize the transaction cost between the

microgrid and the main power grid, depreciation cost of EVs and ESS, and the cost of pollutant
18
 ACCEPTED MANUSCRIPT

treatment. The operation and maintenance cost of PV arrays was not considered in the model.

(1) Transaction cost between the microgrid and the main power grid can be described as
T
C1   P (t ) St (23)
t 1

where P(t ) is the transmission electricity between microgrid and the main power grid. The

positive value of P(t ) denotes that the main power grid transmits electricity to microgrid, and the

negative value of P(t ) denotes that microgrid sells electricity to the main power grid. St is the

electricity price over period t. If St  0 , it indicates that microgrid purchases electricity from the

main power grid. When St  0 , it indicates that microgrid sells electricity to the main power grid.

(2) Depreciation cost of EVs can be expressed as

m
cr
C2   (
ti 2

i 1 E1  ti1
Pi EV (t ) dt ) (24)

where m is the total number of EVs, cr is the battery replacement cost of EV, E1 is total charge-

discharge capacity of EV in its life cycle, t i1

and t i2
are the starting and ending time of i-th EV

accessing to the microgrid, and Pi EV (t ) represents the charging or discharging power of i-th EV

during period t [43].

(3) Depreciation cost of ESS is defined as follows [55].

n
cs
C3   (
tj2

j 1 E2  t j1
PjESS (t ) dt ) (25)

where n is the total number of energy storage unit (ESU), cs is the battery replacement cost of

ESU, E2 is total charge-discharge capacity of ESU in its life cycle, t j1

and t j2
are the starting

and ending time of j-th ESU accessing to the microgrid, and PjESS (t ) represents the charging or

discharging power of j-th ESU during period t .

(4) Cost of pollutant treatment can be expressed as

K
C4   (c k  k ) P (26)
k 1

19
 ACCEPTED MANUSCRIPT

where k represents the k-th pollutant emission, K is the total number of pollutant emissions including

CO 2 , SO 2 , and NO x [43]. c k is the treatment cost of the k-th pollutant emission,  k is the

pollutant discharge coefficients of the k-th pollutant emission, and P is the transaction electricity

that was transmitted to microgrid from the main power grid.

Therefore, the total cost of the proposed model is consisted of transaction cost between microgrid

and the main power grid, depreciation cost of EVs and ESS, and treatment cost of pollutant

emissions. The total cost is presented as follows.

min  C1  C2  C3  C4 (27)

3.2.3 Constraints of the model

The constraints of the model will be given in this section, including electricity supply and demand

balance constraint, the constraint of transmission capacity between the microgrid and the main

power grid, as well as capacity and charge/discharge constraints of ESS. The following part provides

the description of each constraint.

(1) Electricity supply and demand balance constraint. The balance of electricity supply and

demand should be kept in the operation of microgrid, and the constraint can be described as:

P  Ppv  PEVs  PESS  Pload (28)

where P is the transmission power between microgrid and the main power grid, Psolar is the PV

power output, PEV is the net power output of EVs, PESS is the net power output of ESS, and Pload

is the total power load.

(2) Capacity constraints of ESS.

CS min
j  CS j  CS max
j
(29)

where CS j is the CS of j-th ESU, CS min

j
represents the lower limits, and CS max
j
represents the
20
 ACCEPTED MANUSCRIPT

upper limits.

(3) Charge/discharge constraints of ESS.

Pjmin  PjESS  Pjmax (30)

where PjESS is charging or discharging power of the j-th ESU, Pjmin denotes the lower limits, and

Pjmax denotes the upper limits.

(4) Transmission capacity constraints. The transmission power between microgrid and the main

power grid is restricted in a certain range to ensure the stability and reliability of power system.

 P max  P  P max (31)

where P max is the upper limit of transmission power.

4. Results and discussions

4.1 Data

The residential power load data was obtained from Dataport website which provides a vast

database of original and curated data, including utility market operations data to consumer

behavioral research data [56]. The aggregated power load of 40 residential buildings which were

randomly selected, was used in this study. The residential buildings are located in Austin, Texas,

including single-family homes, town homes, apartments, and stores. The power load profiles at 1-

hour resolution between January 1, 2018 and February 1, 2018 were used in the experiments. The

training data corresponding to the data during January 1, 2018, 00:00 AM and January 28, 2018,

23:00 PM, and the testing data corresponding to the data over January 29, 2018, 00:00 AM-February

1, 2018, 23:00 PM. In addition, the weather variables of input in forecasting residential power load

were obtained from MesoWest website corresponding to KATT station, Austin [57].

The proposed DRNN-LSTM model was also tested on the data obtained from Yulara PV plant
21
 ACCEPTED MANUSCRIPT

[58]. The output power data over January 1, 2018-February 1, 2018 of a rooftop PV whose array

rating is 106.6 kW, was used in the experiment. Meanwhile, the website has provided the related

weather data which is necessary for the forecasting model. The partition of training data and testing

data is the same as the partition of that in forecasting residential power load.

4.2 Parameters setup

The proposed forecasting model was implemented in Python by using the Keras API running on

a TensorFlow backend. Table 1 presents the search space for hyper-parameters optimization.

Table 1
Search space of hyper-parameters.
Hyper-parameter Range
Length of output from Layer 2, Layer 3, Layer 4, and Layer 5 [10, 200]
Types of neuron nodes in Layer 2 and Layer 3 [LSTM, GRU]
Activation functions in Layer 4, Layer 5, and Layer 6 [Sigmoid, ReLU]
Dropout in Layer 2, Layer 3, Layer 4, and Layer 5 [0, 1]

In Table 1, the ranges of output length from Layer 2, Layer 3, Layer 4, and Layer 5 are set

reasonably. To avoid gradient vanishing or exploding, the type of neuron nodes on Layer 2 and

Layer 3 are optimized from LSTM and GRU. It is because that both LSTM and GRU can consider

the time dependency in data, but they may have different effects in analyzing different datasets.

Meanwhile, the forecasting performance was estimated and optimized by using different activation

functions in Layer 4, Layer 5, and Layer 6.

4.3 Residential power load forecasting

The proposed DRNN-LSTM model was used to forecast the short-term aggregated power load

of 40 residential buildings in Austin, TX, the United States. Fig. 11 shows the training and testing

process of the DRNN-LSTM model using the aggregated residential power load data over January
22
 ACCEPTED MANUSCRIPT

1, 2018-February 1, 2018. Fig. 12 presents the forecasting results during January 29, 2018-February

1, 2018.

Training Testing

Fig. 11. Training and testing process of the DRNN-LSTM model using the aggregated residential power load data
over January 1, 2018-February 1, 2018.

Fig. 12. Forecasting results of aggregated residential power load using DRNN-LSTM during January 29, 2018-
February 1, 2018.

From Fig. 11-12, it can be seen that residential power load is of high volatility and fluctuates

between 17.36 kW and 83.02 kW. There are two power load peaks in the morning and evening

during which residents tend to consume more electricity as they may cook meals, washing clothes,

and showering etc. However, the power load profile did not show an obvious daily periodic trends.

It is because that residential electricity consumption behavior is of high randomness and usually

affected by interferences from environment. In Fig. 12, it can be seen that the forecasting profile

almost coincides with actual power load profile. It illustrates that the DRNN-LSTM model has great

capability in modeling complex residential power load and can achieve good performance.

To further demonstrate the forecasting performance, multi-layer perception (MLP) network and
23
 ACCEPTED MANUSCRIPT

support vector machine (SVM) were also used to forecast aggregated residential power load. Fig.

13 shows the forecasting results over January 29, 2018-February 1, 2018 using MLP and SVM

respectively. Table 4 presents the relative performances of the DRNN-LSTM, MLP, and SVM in

forecasting aggregated residential power load.

(a) MLP (b) SVM

Fig. 13. Forecasting results over January 29, 2018-February 1, 2018 using MLP and SVM respectively.

Table 4
Evaluation metrics values of the DRNN-LSTM, MLP and SVM in forecasting aggregated residential power load.
Models DRNN-LSTM MLP SVM
RMSE 2.987 5.654 6.191
MAE 2.365 4.383 4.488
MAPE 7.43% 12.68% 19.02%
PCC 0.925 0.701 0.646

From Fig. 13, it can be seen that MLP perform better than SVM in forecasting aggregated

residential power load, but they still cannot surpass the proposed DRNN-LSTM model. Table 4

shows that RMSE and MAE of the DRNN-LSTM model in testing phase are 2.987 and 2.365

respectively, which are significantly less than the errors obtained by MLP and SVM. Meanwhile,

the MAPE of the DRNN-LSTM model (7.426%) is less than other two methods as well. When the

PCC is close to 1, it represents good forecasting results. The PCC of the DRNN-LSTM model

achieves 0.925 which is higher than MLP (0.701) and SVM (0.646). This is because that the high

randomness and instantaneity of residential power load lead to complex characteristics of power

load profile, such as nonlinearity and aperiodicity. Therefore linear model (e.g. SVM) has some

24
 ACCEPTED MANUSCRIPT

limitations in forecasting residential power load. As for MLP, it cannot accommodate the time

dependency in the residential power load, thus influencing the forecasting performance.

4.4 Solar power forecasting

The DRNN-LSTM model was also used to forecast the PV power output including 324 panels in

SITD (Sails in the Desert), Yulara, Australia. Fig. 14 shows the training and testing process of the

DRNN-LSTM model using the data of PV power output over January 1, 2018-February 1, 2018.

Fig. 15 presents the forecasting results during January 29, 2018-February 1, 2018.

Training Testing

Fig. 14. Training and testing process of the DRNN-LSTM model using the data of PV power output over January
1, 2018-February 1, 2018.

Fig. 15. Forecasting results of the PV power output using DRNN-LSTM during January 29, 2018-February 1,
2018.

In Fig. 14-15, the profile of PV power output fluctuates slightly compared with the profile of

residential power load, and the fluctuation may be caused by some floating clouds. Besides, the

profile exhibits certain periodicity. Generally, the PV power output is 0 during 08:00 PM - 05:00

25
 ACCEPTED MANUSCRIPT

AM since there is no sunlight at that time. At around 12:00 AM, the PV power output reaches the

maximum that is about 100kW. From Fig. 14-15, it can be seen that the forecasting results almost

bring into correspondence with actual data in testing phase. In other words, the model performs well

in forecasting both PV power output and residential power load. It denotes that the DRNN-LSTM

has strong generalization ability.

At the same time, the forecasting results are compared with the results obtained by MLP and

SVM. Fig. 16 shows the forecasting results over January 29, 2018-February 1, 2018 using MLP and

SVM respectively. Table 5 presents the relative performances of the DRNN-LSTM, MLP, and SVM

in forecasting the PV power output.

(a) MLP (b) SVM

Fig. 16. Forecasting results using MLP and SVM respectively over January 29, 2018-February 1, 2018.

Table 5
Evaluation metrics values of the DRNN-LSTM, MLP and SVM in forecasting the PV power output.
Models DRNN-LSTM MLP SVM
RMSE 7.536 12.181 14.923
MAE 4.369 7.526 11.549
MAPE 15.87% 34.71% 38.96%
PCC 0.961 0.928 0.914

From Fig. 15-16, it can be seen that the DRNN-LSTM model achieves the best forecasting results

through comparing the forecasting power output profile with actual power output profile. Table 5

shows that RMSE, MAE, and MAPE of the DRNN-LSTM model are less than MLP and SVM.

Besides, the PCC of the DRNN-LSTM model is 0.961 which is higher than MLP (0.928) and SVM

26
 ACCEPTED MANUSCRIPT

(0.914). It can also be found that all the evaluation metrics values of MLP are just slightly better

than those of SVM. It is because that the power output profile of PV array is of certain linearity and

periodicity. The practices have shown the capability of MLP and SVM in dealing with simple

linearity problems.

Overall, the proposed DRNN-LSTM model is of capability to make more accurate forecasting

results than traditional method in forecasting both aggregated residential power load and the PV

power output over short-term period. It provides the possibility for integrating renewable energy

efficiently, reducing pollutant emissions, as well as promoting to implement more flexible demand

side management and maintain the stability of power system.

The forecasting results of residential power load and the PV power output on February 1, 2018

will be used as the given experimental setup in the optimal load dispatch model, thereby contributing

to promote interaction and supply-demand balance in the grid-connected community microgrid.

4.5 Optimization results

In this section, three scheduling scenarios were discussed in order to analyze the influence of EVs

and ESS on the dispatch results.

Scheduling scenario 1:

To prove the effect of the coordinated charging/discharging of EVs on shifting peak load and

improving power system stability, the charging/discharging mode of EVs in scheduling scenario 1

was uncoordinated. At the same time, the ESS was not connected to the microgrid.

Scheduling scenario 2:

Compared with scheduling scenario 1, the ESS was connected to the microgrid in scheduling

scenario 2.
27
 ACCEPTED MANUSCRIPT

Scheduling scenario 3:

In this scheduling scenario, the EVs were dispatched under coordinated charging/discharging

mode. Besides, the EVs was constrained to have enough power to meet the needs of residential daily

driving, and the residual capacity of EVs battery will participate in the load dispatch of microgrid.

Meanwhile, the ESS was considered in this scheduling scenario.

4.5.1 Experimental setup

Although EVs has developed rapidly, the EVs ownership is still limited. Therefore, this study

assumed that the permeability of EVs was 50%, i.e. 20 EVs were connected to the microgrid. All

the EVs were integrated as a whole and controlled by the centralized controller. Fig. 17 shows the

daily power load profiles under 0% and 50% permeability of EVs. It can be seen from the two power

load profiles that the load is low during deep night and noon, but is high during morning and evening.

When the permeability of EVs is 50%, the peak load during evening will increase significantly and

bring more burden on the main power grid. This is because that the residents were assumed to charge

their EVs when they arrive home in the evening, and the EVs adopt uncoordinated charging mode.

Table 6 shows the parameters of EVs and each EV has the same parameters. The replacement cost

of an EV battery is ¥ 30,000, and the total charge-discharge capacity in its life cycle is 240,000 kWh

[59].

Table 7 shows the parameters of ESS. The replacement cost of ESS is ¥ 120,000, and the total

charge-discharge capacity in its life cycle is 800,000 kWh [60].

The day-ahead hourly purchasing price and selling price of electricity are shown in Fig. 18. The

selling price was assumed to be 10% lower than the respective purchasing price.

In this study, the operation and maintenance costs of PV arrays is set to 0, the environmental
28
 ACCEPTED MANUSCRIPT

parameters of the main power grid are listed in Table 8 [43].

Fig. 17. Daily load profile under 50% permeability of EVs.

Table 6
Parameters of EVs.
Parameters Value
Battery capacity (kWh) 30
SC lower/upper limits (%) 10/90
Power load per 100 km (kWh/100 km) 15 10
Charging efficiency 0.9
Discharging efficiency 0.9
Charging power limit (kWh) 5
Charging power limit (kWh) 5
Replacement cost of battery (¥) 30000

Table 7
Parameters of ESS.
Parameters Value
Maximum capacity (kWh) 100
Initial storage (kWh) 0.5
Maximum charging rate (kW) 10
Maximum discharging rate (kW) 10
Charging efficiency 0.9
Discharging efficiency 0.9
Replacement cost of battery (¥) 120000

Fig. 18. Day-ahead hourly purchasing price and selling price of electricity.

Table 8

29
 ACCEPTED MANUSCRIPT

Parameters of pollutant emission.

Type Cost (¥/kg) Pollutant emission coefficient of the main power grid (g/kWh)
CO 2 0.21 889
SO 2 14.842 1.8
NO x 62.964 1.6

Generally, the dispatch cycle is set as one day, and 1-h is a dispatch period. The profiles of

residential power load and the PV power output in this load dispatch model, come from the

forecasting results on February 1, 2018.

4.5.2 Comparison of different scheduling scenarios

In this section, PSO algorithm was used to optimize the load dispatch under three different

scheduling scenarios. The particle population size and maximum iteration number of the algorithms

were set to 100 and 500 respectively. The optimal result, which reflects the best performance on the

load dispatch of microgrid, was obtained by 100 times run.

In scheduling scenario 1, the ESS was not considered in the microgrid. PV arrays provides the

renewable energy for the microgrid, and the main power grid plays a complementary role when the

PV power output cannot satisfy the load demand of microgrid. Besides, the EVs adopt

uncoordinated charging mode. It means that EVs just provides a part of load. The dispatch profiles

of the microgrid under scheduling scenario 1 are shown in Fig. 19.

Fig. 19. Dispatch profiles of the microgrid under scheduling scenario 1.

It is can be seen from Fig. 19 that PV arrays generates solar power over 05:00 AM-18:00 PM,
30
 ACCEPTED MANUSCRIPT

and reaches the peak output power (i.e. 59.96kW) at around 10:00 AM when the light condition is

the best. The load profile represents the total load of residential power load and the charging load

of EVs. There are two peaks in load profile which occur in the morning and evening respectively.

We can also see that the load is always above 22 kW. This may be because that some household

appliances are in operation even at deep night, such as refrigerator and air conditioner.

Meanwhile, it can be found that the quantity of electricity purchasing from the main power grid

declines with the PV power output increases. At about 10:00 AM, the amount of electricity

purchasing decreases to 0. When the value of grid profile is less than 0, it means that PV arrays can

generate enough solar power to satisfy the load, and the microgrid can sell extra electricity to the

main power grid. By 10:00 AM, the microgrid can sell electricity to the main power grid with the

maximum power output of 30 kW. After that, the PV power output will decline gradually as the

light illumination is weakening. Since the electricity purchasing price is high during 09:00AM-

13:00 PM, thus the PV arrays contributes to reduce the total costs of the microgrid. However, the

load in the evening is not optimized as the residual electricity of EVs does not participate in the

dispatch.

Compared with scheduling scenario 1, an EES whose maximum capacity is 100 kWh was adopted

in the community microgrid in scheduling scenario 2. The EVs still adopt uncoordinated charging

mode. Fig. 20 shows the dispatch profiles of the microgrid under scheduling scenario 2.

31
 ACCEPTED MANUSCRIPT

Fig. 20. Dispatch profiles of the microgrid under scheduling scenario 2.

In Fig. 20, when the value of ESS profile is above 0, it denotes that ESS is charging. When the

value of ESS profile is less than 0, it means that the ESS is discharging. From Fig. 18 and Fig. 20,

it can be seen that ESS is charging in the periods of low electricity purchasing price and discharging

in the periods of high electricity purchasing price. It can also be found that ESS is also charging in

the period of 9:00 AM and 13:00 PM, although the electricity purchasing price is high. It is because

that PV array generates excess electricity during this period, and the microgrid also sells electricity

to the main power grid to reduce costs.

Compared with scheduling scenario 1, the adoption of ESS in microgrid contributes to shift grid

peak load which occurs in the morning and evening, thus improving the stability of power system.

In addition, the ESS helps to integrate distributed renewable energy and reduce electricity demand

from the main power grid. Therefore, the operation costs of microgrid will be reduced and the

environmental problems will be mediated. However, due to the maximum capacity of the ESS

(100kWh) is limited, the load dispatch has not yet reached the optimal result.

In scheduling scenario 3, the EVs adopt coordinated charging mode. Meanwhile, the ESS is also

adopted in the microgrid. The dispatch profiles of the microgrid under scheduling scenario 3 are

shown in Fig. 21.

Fig. 21. Dispatch profiles of the microgrid under scheduling scenario 3.

32
 ACCEPTED MANUSCRIPT

It can be seen from Fig. 21 that the peak load of grid profile in the evening is further shifted to

deep night during which the electricity purchasing price is low. Since the EVs adopt coordinating

mode, their load profile has changed a lot. When the value of EV profile is above 0, the EVs are

charging electricity, and when the value is less than 0, the EVs are discharging electricity. We can

see that the EVs is charging at deep night and discharging in the evening. Therefore, the peak load

of grid during evening has been decreased significantly. It can also be found that the total

charging/discharging electricity of ESS is reduced compared with scheduling scenario 2. The reason

is that unit cost of charging/discharging electricity of ESS is greater than that of EVs and the EVs

will be given priority for dispatching.

From Fig. 19 and Fig. 21, it can be found that the PV power output can meet the load demand

in microgrid and sell electricity to the main power grid in 9:00 AM-13:00 PM, but the ESS and EVs

only charged a small amount of electricity during this period. It is because that the selling price of

electricity is high in this period, and selling the excess electricity to the main power grid is best

choice to reduce the total costs.

Table 9 shows the total cost under the three different scheduling scenarios.

Table 9
Total cost under different scheduling scenarios.
Scenarios Total cost
Scenario 1 582.13
Scenario 2 574.04
Scenario 3 529.89

It can be seen from Table 9 that the minimum dispatch cost was achieved under scheduling

scenario 3. This reveals that ESS and coordinated charging mode of EVs contribute to reduce the

total costs. Besides, the peak load of grid will be shifted substantially and the renewable energy will

33
 ACCEPTED MANUSCRIPT

be integrated efficiently in scheduling scenario 3, thereby contributing to improve the stability of

power system and reduce pollutant emissions. It can be believed that the battery replacement cost

of ESS and EVs will be further reduced with advances in technology. Larger scale ESS and EVs

with higher energy storage capacity can be applied in microgrid to reduce operation costs and shift

peak load.

To further reduce peak load of grid, power companies can also provide targeted demand side

strategies, such as increasing electricity purchasing price in peak load periods or providing

incentives for residents who response the demand side strategies. Therefore, the reliability and

stability of power system will be further improved.

5. Conclusions

In this study, a DRNN-LSTM model was proposed to forecast aggregated residential power load

and the PV power output over short-term period. Two real-world datasets were used to test the

DRNN-LSTM model. The forecasting results release that the DRNN-LSTM model performs better

than MLP network and SVM in forecasting the output power of PV array. It achieved more excellent

performance in forecasting aggregated residential power load with the MAPE of 7.43% since

traditional forecasting methods have limitations in modeling complex nonlinear problems and

cannot consider the time dependencies in the data. Therefore, the proposed DRNN-LSTM model

presents the potential for making more accurate short-term forecasting to support economical load

dispatch of community microgrid.

Furthermore, this paper explored the optimal load dispatch in a grid-connected community

microgrid under three different scheduling scenarios. In order to promote interaction and balance

between supply and demand, the uncertainties of both the PV power output and residential power
34
 ACCEPTED MANUSCRIPT

load were considered in the load dispatch optimization by using the forecasting results of the

proposed DRNN-LSTM model. At last, the load dispatch model was optimized by a particle swarm

optimization algorithm, and the simulation results show that EES and EVs can shift peak load,

improve the solar energy utilization, as well as reduce 8.97% of daily costs. In addition, the charging

mode of EVs was discussed. It shows that charging loads of EVs will be shifted form high-priced

periods to low-priced periods under the coordinated charging mode, thereby reducing the total costs

and improving the stability of power system.

However, there are still few limitations in this study. The forecasting model and the optimization

model are two separate part, and future study needs to explore the dynamic mechanism between

them systematically. Besides, this study achieved hourly load dispatch in community microgrid. To

optimize the load dispatch in real time, more fine-grained data will be used in future.

Acknowledgement

This work is supported by the National Natural Science Foundation of China (Nos. 71822104,

71690235, 71690230 and 71521001), Anhui Science and Technology Major Project (No.

17030901024), Hong Kong Scholars Program (No. 2017-167), China Postdoctoral Science

Foundation (No. 2017M612072), and the Fundamental Research Funds for the Central Universities

(No. JZ2018HGPA0271).

References

[1] Djørup S, Thellufsen JZ, Sorknæs P. The electricity market in a renewable energy system.

Energy. 2018;162:148-157.

[2] Pudleiner D, Pielli K. Energy efficiency as a means to expand energy access: A Uganda roadmap.

35
 ACCEPTED MANUSCRIPT

Energy Policy. 2018;120:354-364.

[3] Zhang Y-J, Peng H-R. Exploring the direct rebound effect of residential electricity consumption:

An empirical study in China. Applied energy. 2017;196:132-141.

[4] Kovacic Z, Giampietro M. Empty promises or promising futures? The case of smart grids.

Energy. 2015;93:67-74.

[5] Wang D, Qiu J, Reedman L, Meng K, Lai LL. Two-stage energy management for networked

microgrids with high renewable penetration. Applied Energy. 2018;226:39-48.

[6] Lu X, Zhou K, Yang S. Multi-objective optimal dispatch of microgrid containing electric

vehicles. Journal of Cleaner Production. 2017;165:1572-1581.

[7] Parhizi S, Khodaei A, Shahidehpour M. Market-Based Versus Price-Based Microgrid Optimal

Scheduling. IEEE Transactions on Smart Grid. 2018;9(2):615-623.

[8] Kuznetsova E, Li Y-F, Ruiz C, Zio E. An integrated framework of agent-based modelling and

robust optimization for microgrid energy management. Applied Energy. 2014;129:70-88.

[9] Khan MRB, Jidin R, Pasupuleti J. Multi-agent based distributed control architecture for

microgrid energy management and optimization. Energy Conversion and Management.

2016;112:288-307.

[10] Modiri-Delshad M, Kaboli SHA, Taslimi-Renani E, Rahim NA. Backtracking search algorithm

for solving economic dispatch problems with valve-point effects and multiple fuel options. Energy.

2016;116:637-649.

[11] Pappala VS, Erlich I, Rohrig K, Dobschinski J. A stochastic model for the optimal operation

of a wind-thermal power system. IEEE Transactions on Power Systems. 2009;24(2):940-950.

[12] Wang L, Li Q, Ding R, Sun M, Wang G. Integrated scheduling of energy supply and demand

36
 ACCEPTED MANUSCRIPT

in microgrids under uncertainty: A robust multi-objective optimization approach. Energy.

2017;130:1-14.

[13] Asadi H, Kaboli SHA, Mohammadi A, Oladazimi M. Fuzzy-control-based five-step Li-ion

battery charger by using AC impedance technique. Fourth International Conference on Machine

Vision (ICMV 2011): Machine Vision, Image Processing, and Pattern Analysis. 2012; 8349.

[14] Zheng M, Meinrenken CJ, Lackner KS. Smart households: Dispatch strategies and economic

analysis of distributed energy storage for residential peak shaving. Applied Energy. 2015;147:246-

257.

[15] Garcia-Torres F, Vilaplana DG, Bordons C, Roncero-Sanchez P, Ridao MA. Optimal

Management of Microgrids with External Agents including Battery/Fuel Cell Electric Vehicles.

IEEE Transactions on Smart Grid. 2018.

[16] Farrokhabadi M, König S, Cañizares CA, Bhattacharya K, Leibfried T. Battery energy storage

system models for microgrid stability analysis and dynamic simulation. IEEE Transactions on

Power Systems. 2018;33(2):2301-2312.

[17] Alham M, Elshahed M, Ibrahim DK, El Zahab EEDA. A dynamic economic emission dispatch

considering wind power uncertainty incorporating energy storage system and demand side

management. Renewable Energy. 2016;96:800-811.

[18] Ansari B, Shi D, Sharma R, Simoes MG. Economic analysis, optimal sizing and management

of energy storage for PV grid integration. IEEE/PES Transmission and Distribution Conference and

Exposition (T&D). 2016;1-5.

[19] Hashemi-Dezaki H, Hamzeh M, Askarian-Abyaneh H, Haeri-Khiavi H. Risk management of

smart grids based on managed charging of PHEVs and vehicle-to-grid strategy using Monte Carlo

37
 ACCEPTED MANUSCRIPT

simulation. Energy Conversion and Management. 2015;100:262-76.

[20] Liu C, Chau K, Diao C, Zhong J, Zhang X, Gao S, et al. A new DC micro-grid system using

renewable energy and electric vehicles for smart energy delivery. IEEE Vehicle Power and

Propulsion Conference. 2010;1-6.

[21] Kamankesh H, Agelidis VG, Kavousi-Fard A. Optimal scheduling of renewable micro-grids

considering plug-in hybrid electric vehicle charging demand. Energy. 2016;100:285-297.

[22] Morais H, Sousa T, Soares J, Faria P, Vale Z. Distributed energy resources management using

plug-in hybrid electric vehicles as a fuel-shifting demand response resource. Energy Conversion

and Management. 2015;97:78-93.

[23] Craparo E, Karatas M, Singham DI. A robust optimization approach to hybrid microgrid

operation using ensemble weather forecasts. Applied Energy. 2017;201:135-147.

[24] Gupta R, Gupta NK. A robust optimization based approach for microgrid operation in

deregulated environment. Energy Conversion and Management. 2015;93:121-131.

[25] Kaboli SHA, Fallahpour A, Kazemi N, Selvaraj J, Rahim N. An expression-driven approach

for long-term electric power consumption forecasting. American Journal of Data Mining and

Knowledge Discovery. 2016;1(1):16-28.

[26] Kaboli SHA, Selvaraj J, Rahim N. Long-term electric energy consumption forecasting via

artificial cooperative search algorithm. Energy. 2016;115:857-871.

[27] Russo A, Raischel F, Lind PG. Air quality prediction using optimal neural networks with

stochastic variables. Atmospheric Environment. 2013;79:822-830.

[28] Tao Y, Yan H, Gao H, Sun Y, Li G. Application of SVR optimized by Modified Simulated

Annealing (MSA-SVR) air conditioning load prediction model. Journal of Industrial Information

38
 ACCEPTED MANUSCRIPT

Integration. 2018.

[29] Yu Z, Haghighat F, Fung BC, Yoshino H. A decision tree method for building energy demand

modeling. Energy and Buildings. 2010;42(10):1637-1646.

[30] Kaboli SHA, Fallahpour A, Selvaraj J, Rahim N. Long-term electrical energy consumption

formulating and forecasting via optimized gene expression programming. Energy. 2017;126:144-

164.

[31] Chen J, Wang X, Steemers K. A statistical analysis of a residential energy consumption survey

study in Hangzhou, China. Energy and Buildings. 2013;66:193-202.

[32] Liu B, Nowotarski J, Hong T, Weron R. Probabilistic load forecasting via quantile regression

averaging on sister forecasts. IEEE Transactions on Smart Grid. 2017;8(2):730-737.

[33] Bahrami S, Hooshmand R-A, Parastegari M. Short term electric load forecasting by wavelet

transform and grey model improved by PSO (particle swarm optimization) algorithm. Energy.

2014;72:434-442.

[34] Biswas MR, Robinson MD, Fumo N. Prediction of residential building energy consumption: a

neural network approach. Energy. 2016;117:84-92.

[35] Lind PG, Vera-Tudela L, Wächter M, Kühn M, Peinke J. Normal Behaviour Models for Wind

Turbine Vibrations: Comparison of Neural Networks and a Stochastic Approach. Energies.

2017;10(12):1944.

[36] Hagan MT, Demuth HB, Beale MH. Neural network design. Boston: Pws Pub., 1996.

[37] Yogatama D, Dyer C, Ling W, Blunsom P. Generative and discriminative text classification

with recurrent neural networks. arXiv preprint arXiv:170301898. 2017.

[38] Zhang X-Y, Yin F, Zhang Y-M, Liu C-L, Bengio Y. Drawing and recognizing chinese

39
 ACCEPTED MANUSCRIPT

characters with recurrent neural network. IEEE Transactions on Pattern Analysis and Machine

Intelligence. 2018;40(4):849-862.

[39] Britz D, Goldie A, Luong T, Le Q. Massive exploration of neural machine translation

architectures. arXiv preprint arXiv:170303906. 2017.

[40] Rahman A, Srikumar V, Smith AD. Predicting electricity consumption for commercial and

residential buildings using deep recurrent neural networks. Applied Energy. 2018;212:372-385.

[41] Hochreiter S, Schmidhuber J. Long short-term memory. Neural computation. 1997;9(8):1735-

1780.

[42] Kaboli SHA, Selvaraj J, Rahim N. Rain-fall optimization algorithm: A population based

algorithm for solving constrained optimization problems. Journal of Computational Science.

2017;19:31-42.

[43] Lu X, Zhou K, Yang S, Liu H. Multi-objective optimal load dispatch of microgrid with

stochastic access of electric vehicles. Journal of Cleaner Production. 2018;195:187-199.

[44] Rafieerad A, Bushroa A, Nasiri-Tabrizi B, Kaboli S, Khanahmadi S, Amiri A, et al. Toward

improved mechanical, tribological, corrosion and in-vitro bioactivity properties of mixed oxide

nanotubes on Ti–6Al–7Nb implant using multi-objective PSO. Journal of the Mechanical Behavior

of Biomedical Materials. 2017;69:1-18.

[45] Sebtahmadi SS, Azad HB, Kaboli SHA, Islam MD, Mekhilef S. A PSO-DQ current control

scheme for performance enhancement of Z-source matrix converter to drive IM Fed by abnormal

voltage. IEEE Transactions on Power Electronics. 2018;33(2):1666-1681.

[46] Rafieerad A, Bushroa A, Nasiri-Tabrizi B, Fallahpour A, Vadivelu J, Musa S, et al. GEP-based

method to formulate adhesion strength and hardness of Nb PVD coated on Ti–6Al–7Nb aimed at

40
 ACCEPTED MANUSCRIPT

developing mixed oxide nanotubular arrays. Journal of the Mechanical Behavior of Biomedical

Materials. 2016;61:182-96.

[47] Trelea I C. The particle swarm optimization algorithm: convergence analysis and parameter

selection. Information Processing Letters, 2003, 85(6): 317-325.

[48] Girosi F, Jones M, Poggio T. Regularization theory and neural networks architectures. Neural

computation. 1995;7(2):219-269.

[49] Chollet F. Keras: The python deep learning library. Astrophysics Source Code Library, 2018.

[50] Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: A simple way

to prevent neural networks from overfitting. The Journal of Machine Learning Research.

2014;15(1):1929-1958.

[51] Kingma DP, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv:14126980.

2014.

[52] Bergstra J, Komer B, Eliasmith C, Yamins D, Cox DD. Hyperopt: a python library for model

selection and hyperparameter optimization. Computational Science & Discovery. 2015;8(1):014008.

[53] Huang Y, Guo C, Wang L, Bao Y, Dai S, Ding Q. A cluster-based dispatch strategy for electric

vehicles considering user satisfaction. Automation of Electric Power Systems. 2015;39(12):183-

191.

[54] Kavousi-Fard A, Abunasri A, Zare A, Hoseinzadeh R. Impact of plug-in hybrid electric

vehicles charging demand on the optimal energy management of renewable micro-grids. Energy.

2014;78:904-915.

[55] Bucciarelli M, Paoletti S, Vicino A. Optimal sizing of energy storage systems under uncertain

demand and generation. Applied Energy. 2018;225:611-621.

41
 ACCEPTED MANUSCRIPT

[56] Dataport, Pecan Street Inc. https://dataport.cloud, (accessed 15 April 2018).

[57] MesoWest. University of Utah Department of Atmospheric Sciences. http://mesowest.utah.edu,

(accessed 15 April 2018).

[58] DKASC. http://dkasolarcentre.com.au/locations, (accessed 16 April 2018).

[59] Automotive science and technology. http://www.pconline.com.cn/autotech/895/8959994.html,

(accessed 10 August 2018).

[60] ESCN. http://www.escn.com.cn/news/show-392547.html, (accessed 11 August 2018).

42
 ACCEPTED MANUSCRIPT

Highlights

 The residential power load and PV power output are forecasted.

 The proposed model outperforms other methods.

 The forecasting results are used as the input of the microgrid optimization model.

 The adoption of EVs and ESS in microgrid contributes to 8.97% cost reduction.