energies

Article
A Deep Neural Network-Based Optimal Scheduling
Decision-Making Method for Microgrids
Fei Chen 1 , Zhiyang Wang 2 and Yu He 2, *

1 China Southern Power Grid Guizhou Power Grid Co., Ltd., Guiyang 550003, China;
[email protected] (F.C.); [email protected] (Z.W.)
2 School of Electrical Engineering, Guizhou University, Guiyang 550025, China; [email protected]
* Correspondence: [email protected]

Abstract: With the rapid growth in the proportion of renewable energy access and the structural
complexity of distributed energy systems, traditional microgrid (MG) scheduling methods that
rely on mathematical optimization models and expert experience are facing significant challenges.
Therefore, it is essential to present a novel scheduling technique with high intelligence and fast
decision-making capacity to realize MGs’ automatic operation and regulation. This paper proposes
an optimal scheduling decision-making method for MGs based on deep neural networks (DNN).
Firstly, a typical mathematical scheduling model used for MG operation is introduced, and the
limitations of current methods are analyzed. Then, a two-stage optimal scheduling framework
comprising day-ahead and intra-day stages is presented. The day-ahead part is solved by mixed
integer linear programming (MILP), and the intra-day part uses a convolutional neural network
(CNN)—bidirectional long short-term memory (Bi LSTM) for high-speed rolling decision making,
with the outputs adjusted by a power correction balance algorithm. Finally, the validity of the model
and algorithm of this paper are verified by arithmetic case analysis.

Keywords: microgrid; optimal dispatch; convolutional neural network; deep bidirectional long-short
memory neural network; artificial intelligence

Citation: Chen, F.; Wang, Z.; He, Y. A

Deep Neural Network-Based 1. Introduction
Optimal Scheduling
MGs, characterized by cleanliness, low-carbon emissions, and openness, have gar-
Decision-Making Method for
nered significant attention due to the rapid development of renewable energy in recent
Microgrids. Energies 2023, 16, 7635.
years [1]. As a critical solution to improve the consumption of distributed power sources
https://doi.org/10.3390/en16227635
and the reliability of power supply, MGs have become essential to reducing fossil energy
Academic Editors: Andrea Fioravanti pollution and promoting sustainable development in China [2]. However, the intermittent,
and Fabrizio Ciancetta volatile, and uncertain nature of renewable energy sources poses significant challenges
to the stable operation of MGs [3]. Additionally, the expansion of the MG system and
Received: 13 October 2023
Revised: 13 November 2023
the increased number of its components also impose more demanding requirements on
Accepted: 16 November 2023
the optimal scheduling method. The traditional scheduling method based on numerical
Published: 17 November 2023 model optimization, scenario matching, and manual manning makes it difficult to meet
the demand. Hence, studying fast, accurate, and intelligent scheduling decision methods
holds immense practical value and significance [4].
Currently, the predominant method for solving MG optimal scheduling problems
Copyright: © 2023 by the authors. is a numerical calculation method grounded in optimization theory. Common optimal
Licensee MDPI, Basel, Switzerland. scheduling models encompass MILP [5], dynamic programming [6], distributed optimiza-
This article is an open access article tion [7], etc. Likewise, common model-solving algorithms involve intelligent algorithms [8],
distributed under the terms and
second-order cone relaxation methods [9], Lagrange relaxation methods, etc. However, as
conditions of the Creative Commons
the uncertainties of the source–load dual-side within MGs escalate, solving the optimal
Attribution (CC BY) license (https://
scheduling problem under such uncertainties becomes a more realistic and challenging
creativecommons.org/licenses/by/
research problem [10]. Some researchers construct uncertainty planning models. The
4.0/).

Energies 2023, 16, 7635. https://doi.org/10.3390/en16227635 https://www.mdpi.com/journal/energies

Energies 2023, 16, 7635 2 of 17

main modeling and solution methods are stochastic planning [11], chance-constrained
planning [12], etc. Among these, robust optimization [13,14] methods have been proven
to be an effective method of solving MG uncertain optimization problems. They aim
at optimal operation under the worst-case scenario. However, their overly pessimistic
view of uncertain variables may lead to solution results that are too conservative to be
economical. The mathematical models of these methods are relatively complex and com-
putationally expensive. The other researchers used a multi-timescale optimal scheduling
strategy [15], which can be classified into day-ahead scheduling and intra-day stage ac-
cording to the timescale. Among these, the model predictive control (MPC) technique is a
widely employed modeling approach [16]. How to enhance intra-day real-time scheduling
computational efficiency is still a challenge.
In summary, the traditional optimization theory-based scheduling methods rely on
strict mathematical derivation, requiring researchers to participate. With the MG evolving
into a new system characterized by increased uncertainty and complexity, the traditional
optimization scheduling methods are gradually becoming inadequate to meet the demands
of MG operation [17]. Several critical problems of this method are as follows:
(1) The traditional mathematical optimization methods cannot model the components
of the MG in a fast and refined manner, but it is difficult to describe the physical
characteristics of the actual operation of the components using a simplified model [18].
(2) The traditional mathematical MG scheduling models are often nonlinear and noncon-
vex, which is a typical nondeterministic polynomial problem (NP-hard). The problem
is demanding on the solution algorithm, and it is not easy to find the optimal solution.
(3) The computational process of traditional mathematical optimization methods is com-
plex and inefficient, and it is difficult to adapt to the real-time solution of optimization
scheduling problems with uncertainty under complex and variable system operating
conditions [19].
(4) The traditional mathematical optimization methods ignore the significance of histor-
ical data and historical decision-making plans and fail to make use of the valuable
historical decision-making data information accumulated during the system’s opera-
tion.
Recently, the rapid development of computer technology has made neural networks
(NN) an important driver of the new technological revolution and industrial change [20].
A new intelligent decision-making method using NNs based on big data technology may
be a more effective way of thinking, which may help to break through the limitations of
mathematical optimization solution methods. Unlike traditional optimization methods,
the decision-making method based on NNs no longer depends on specific mathematical
models or algorithms; instead, it is trained using extensive real data [21]. This method
can greatly simplify the process and complexity of modeling and solving the optimal
scheduling problem, and cope with various theoretical problems and challenges that keep
emerging through its self-learning and self-evolution process. It can potentially facilitate
the transition from manual supervision to machine intelligence-based monitoring in the
domain of MG scheduling. Moreover, when the data make centralized training of models
bitter due to factors such as privacy and size, the idea of distributed frameworks [22,23]
can also be referred to for decentralized training of small models and then aggregated to
a big model. This allows great flexibility in the implementation of the method. Figure 1
shows the transition from the traditional optimization method to the NN-based method.
 (1) Only load data are used as training inputs without considering the influence of other
system state data on the scheduling decision results. This approach cannot fully ex-
tract the feature information embedded in the valuable historical operation data.
(2) Only using a shallow or single network model to build the scheduling mapping rela-
tionship, the accuracy of the output results is low.
Energies 2023, 16, 7635 3 of 17
(3) The decision results from the NNs-based scheduling method will inevitably violate
some actual constraints, and there is no reasonable and efficient solution to this issue.

Figure
Figure 1. 1. Comparison
Comparison illustration
illustration between
between the traditional
the traditional optimization
optimization methodmethod
and theand the NN-based
NN-based
method.
method.

To address
Several thehave
scholars above issues, to
attempted this paper
utilize proposes
artificial a two-stage
intelligence optimal scheduling
(AI) techniques in
the field offor
method scheduling
MGs. The decisions.
proposed Themethod
literatureaims
[24] utilized long and
to enhance short-term memory
the effectiveness of the NNs-
(LSTM)
driventoscheduling
establish themethod
mapping from
and thesystem
MG’sload to unit
ability to output.
handle However,
uncertainthe constructedand ad-
fluctuations
network
dress the limitations of the traditional mathematical model-driven and manual[25]
structure is relatively simple, and the results are unconstrained. The literature participa-
uses a multi-layer perceptron (MLP) to learn and mimic the scheduling decision of a smart
tion scheduling methods. In the day-ahead part (1 h timescale), which does not require
grid, and an iterative algorithm is used to correct the output of the NNs so that it satisfies
high timeliness, a MILP model is used to obtain the MG’s operating plan. In the day-ahead
the actual constraints. The literature [26] applies a feedforward neural network (FNN)
part
for the(15 min timescale),
optimal scheduling a ofDNN scheduling
combined heat anddecision network
power (CHP) is used
systems, for fast-rolling
which enhances opti-
mization. The main contributions of this paper are as follows:
computational efficiency by about 7000 times while permitting suboptimal cost. Although
previous
(1) Anstudies haverolling
intra-day demonstrated that NNs
optimization are feasible
model based and effective
on DNNs in optimal
and big dataenergy
is proposed,
scheduling, the current research still faces some issues:
which is trained using the dataset clustered by the K-means algorithm to improve
(1) Only load data are and
generalizability usedaccuracy.
as training inputs without considering the influence of other
(2) A novel CNN-Bi LSTMscheduling
system state data on the schedulingdecision
decisionresults. This
network is approach
proposed,cannot
digging fully
deep fea-
extract the feature information embedded in the valuable historical operation
ture information in the system operation data by CNN and establishing the accurate data.
(2) Only using a shallow or single network model to build the scheduling mapping
mapping between input and output by Bi LSTM.
relationship, the accuracy of the output results is low.
(3) A power balance correction algorithm is proposed to fine-tune the DNN outputs to
(3) The decision results from the NNs-based scheduling method will inevitably violate
quickly
some satisfy
actual all practical
constraints, constraints.
and there is no reasonable and efficient solution to this issue.
ToThe proposed
address method
the above canthis
issues, effectively reduce the
paper proposes complexity
a two-stage of solving
optimal the optimal
scheduling
method for MGs. The proposed method aims to enhance the effectiveness of the NNs-driven the so-
scheduling problem and significantly improve computational efficiency (reducing
lution time
scheduling for intra-day
method and the rolling optimization
MG’s ability to handletouncertain
milliseconds), whichand
fluctuations alsoaddress
improves the
intelligence
the limitationslevel of traditional
of the MGs. The rest of this paper
mathematical is organized
model-driven andasmanual
follows: Section 2 presents
participation
scheduling methods. In the
the basic mathematical day-ahead part
optimization model(1 h for
timescale),
day-aheadwhichMG does not require
scheduling. high 3 pre-
Section
timeliness,
sents the DNN-based intra-day rolling optimal scheduling method. Sectionpart
a MILP model is used to obtain the MG’s operating plan. In the day-ahead 4 presents
(15 min timescale), a DNN scheduling decision network is used for fast-rolling
simulation experiments and analyses. Section 5 presents the conclusions of this paper. optimization.
The main contributions of this paper are as follows:
(1) An intra-day rolling optimization model based on DNNs and big data is proposed,
which is trained using the dataset clustered by the K-means algorithm to improve
generalizability and accuracy.
(2) A novel CNN-Bi LSTM scheduling decision network is proposed, digging deep feature
information in the system operation data by CNN and establishing the accurate
mapping between input and output by Bi LSTM.
Energies 2023, 16, 7635 4 of 17

(3) A power balance correction algorithm is proposed to fine-tune the DNN outputs to
quickly satisfy all practical constraints.
The proposed method can effectively reduce the complexity of solving the optimal
scheduling problem and significantly improve computational efficiency (reducing the
solution time for intra-day rolling optimization to milliseconds), which also improves the
intelligence level of MGs. The rest of this paper is organized as follows: Section 2 presents
the basic mathematical optimization model for day-ahead MG scheduling. Section
Energies 2023, 16, x FOR PEER REVIEW 3
4 of 17
presents the DNN-based intra-day rolling optimal scheduling method. Section 4 presents
simulation experiments and analyses. Section 5 presents the conclusions of this paper.
2.
2. Microgrid
Microgrid Day-Ahead
Day-Ahead Optimal
Optimal Scheduling
Scheduling Model
Model
2.1. Overall Composition Structure of Microgrid
2.1. Overall Composition Structure of Microgrid
The
The specific composition structure
specific composition structureof
ofthe
theMG
MGstudied
studiedininthis
thispaper
paperisisshown
shownininFigure
Figure2,
2, where the arrow indicates the direction of power
where the arrow indicates the direction of power flow.flow.

Structure of the MG.

Figure 2. Structure

This MG
This MG consists
consists of
of photovoltaic
photovoltaic panels
panels (PVs), wind turbines
(PVs), wind turbines (WTs),
(WTs), microturbine
microturbine
(MT), upper grid (UG), storage battery (SB), and power load. The power bus is
(MT), upper grid (UG), storage battery (SB), and power load. The power bus is the
the carrier
carrier
of all device power interactions. The power distribution of the entire MG is set by the
of all device power interactions. The power distribution of the entire MG is set by the
scheduling center and sent to each controllable device.
scheduling center and sent to each controllable device.
2.2. Objective Function
2.2. Objective Function
The objective of the MG’s day-ahead optimal scheduling is to minimize the total daily
The objective
operating cost. Theoftotal
the MG’s day-ahead
operating cost of optimal scheduling
the system comprisesis to
theminimize the
operating total
and daily
start-up
operating cost. The total operating cost of the system comprises the operating and
costs of MT, as well as the charging/discharging costs of SB and the purchase/sale cost of start-
up
UG.costs
The of MT, can
above as well as the charging/discharging
be expressed as: costs of SB and the purchase/sale cost
of UG. The above can be expressed as:
T N T
Min ∑ ∑ (CMT,i,t U ) )++ ∑ ((C
T N T
Min 
(C + U+MT,i,t
t =1 i =1
t =1 i =1
CSB,t++CCUG,t
MT,i ,t ) )
MT,i ,t 
t =1
t =1
SB,t UG,t
(1)
(1)

CMT,i,t = ai P2 2 + bi PMT,i,t + ci uon,i,t


 CMT,i ,t = aMT,i,t i PMT, + b P i ,t + ci uon,i ,t


 UMT,i,t =i ,t CUii uMT,
su,i,t
= CU (2)
 U MT,=i ,t CS
CSB,t usu,i ,t
SB Pi SB,t
 

 (2)
SB,t = CS
 CCUG,t = PUG,t et t
SB PSB,
 the operating CUG,t =cost PUG,of
where CMT,i,t , CSB,t , and CUG,t are  t et MT, SB, and UG at time t, respectively.
UMT,i,t is the start-up cost of ith MT at time t. Equation (2) is a detailed expression of the
where CMT,device,
cost of each i , t , CSB,t , andaiC
where , bUG,t arecithe
i , and are operating
the cost factors cost of ofMT,
ith MT,SB, and
and CU
UGi at time
is the t , re-
start-up
cost of ith MT. CS
spectively. U MT,i ,t SB and e are
is the start-up
t the cost factors of SB and the electricity price
cost of i th MT at time t . Equation (2) is a detailed ex- of UG at
time t. PMT,i,t , PSB,t , and PUG,t are the power of ith MT, SB, and UG at time t, respectively.
pression
uon,i,t andofuthe cost of each device, where ai , bMT
su,i,t are the binary variables for ith i , and ci areand
operating the starting
cost factors of ati th
states MT,t,
time
CU i is the start-up cost of i th MT. CSSB and et are the cost factors of SB and the
respectively.
and
electricity price of UG at time t . PMT,i ,t , PSB,t , and PUG,t are the power of i th MT, SB,
and UG at time t , respectively. uon,i ,t and usu,i ,t are the binary variables for i th MT op-
erating and starting states at time t , respectively.
Energies 2023, 16, 7635 5 of 17

The quadratic cost function of MT is linearly approximated by dividing it into

S-segmented linear functions using segmented linearization methods [27], which reduces
the complexity of solving the overall model. The linearization results are as follows:

eMT,i,t = ci uon,i,t + ∑S as PMT,i,t,s

C (3)
s =1

C MT,i,s+1 − C MT,i,s
as = (4)
PMT,i,s+1 − PMT,i,s
eMT,i,t is the MT cost function after the linearization segments, S is the total number
where C
of segments, as is the slope of each segment, PMT,i,s is the output of the ith MT in the sth
segment.

2.3. Constraints
MG’s day-ahead optimal scheduling constraints include power balance constraint,
controllable unit operation constraint, SB operation constraint, and UG operation constraint.

2.3.1. Power Balance Constraint

Equation (5) indicates that power production and consumption are balanced at all
moments in the MG.
N
∑ PMT,i,t + PSB,t + PUG,t + PWT,t + PPV,t = PLoad,t , ∀t (5)
i =1

where PWT,t , PPV,t , and PLoad,t are the power of WT, PV, and power load at time t, respec-
tively.

2.3.2. Controllable Unit Operating Constraints

Controllable unit (MT) operating constraints include output, ramp rate, and unit status
constraints. The above constraints are expressed as follows:
min max
PMT,i uon,i,t ≤ PMT,i,t ≤ PMT,i , ∀i, t (6)

down up
− PMT,i ≤ PMT,i,t − PMT,i,t−1 ≤ PMT,i , ∀i, t (7)

usu,i,t − usd,i,t = uon,i,t − uon,i,t−1 , ∀i, t (8)

usu,i,t + usd,i,t ≤ 1, ∀i, t (9)

min and Pmax are the minimum and maximum output power of ith MT. Pdown and
where PMT,i MT,i MT,i
up
PMT,i are the up and down ramp rate limitations of ith MT. usd,i,t is the binary variables for
ith MT stop state at time t.

2.3.3. Storage Battery Operating Constraints

SB operating constraints include charging and discharging state constraints, output
power constraints, capacity constraints, and capacity cycle constraints. The above con-
straints are expressed as follows:

ucha,t + udis,t ≤ 1, ∀t (10)

dis dis
0 ≤ PSB,t ≤ Pmax udis,t , ∀t (11)
Energies 2023, 16, 7635 6 of 17

char char
Pmax uchar,t ≤ PSB,t ≤ 0, ∀t (12)

dis char
PSB,t = ηSB ( PSB,t udis,t + PSB,t uchar,t ), ∀t (13)

ESB,t = ESB,t−1 + PSB,t , ∀t (14)

Emin ≤ ESB,t ≤ Emax , ∀t (15)

E0 = Eend (16)
where ucha,t and udis,t are the charging and discharging binary variables at time t, respec-
char , Pdis , Pchar , and Pdis are the charging and discharging power at time t and its
tively. PSB,t SB,t max max
limitation, respectively. ηSB is the power conversion efficiency. ESB,t is the capacity at time t
and Emin , Emax are the minimum and maximum capacity.

2.3.4. Power Contact Line Constraints

The constraints on UG are mainly reflected in the power transmission line which
connects it to the MG. The power transmission line constraints include output constraints
and purchase and sale power status constraints. The above constraints are expressed as
follows:
ubuy,t + usell,t ≤ 1, ∀t (17)

buy buy
0 ≤ PUG,t ≤ Pmax ubuy,t , ∀t (18)

sell sell
Pmax usell,t ≤ PUG,t ≤ 0, ∀t (19)

buy sell
PUG,t = PUG,t ubuy,t + PUG,t usell,t , ∀t (20)
where the ubuy,t and usell,t are the power purchase and sale binary variables at time t,
buy sell , Pbuy sell
respectively. PUG,t , PUG,t max , and Pmax are the power purchase, sale time t, and its
limitation, respectively.
So far, the MG’s day-ahead optimal scheduling model based on MILP is completely
constructed.

3. Deep Neural Network-Based Intra-Day Rolling Optimization Method

A data-driven DNN-based scheduling method is proposed in this paper to address
the shortcomings and difficulties of traditional methods in intra-day rolling optimization.
Instead of relying on specific mathematical models, it trains with large amounts of real
data and makes scheduling decisions by high-dimensional matrix multiplication [28].
This method can reduce the complexity of solving the optimal scheduling problem and
significantly improve computational efficiency.

3.1. Intra-Day MPC Rolling Optimization Forms

In this paper, the DNN scheduling decision network is used as an optimizer for MPC
to perform intra-day rolling optimization. MPC is an alternating process of continuously
rolling local optimization and continuously rolling control role implementation. By obtain-
ing ultra-short-term power forecast information in real-time during intra-day scheduling
and using the actual scheduling results and new forecast information as feedback, MPC
rolling optimization forms can greatly reduce the impact of MG uncertainties on optimal
 Energies
Energies 2023, 16,x7635
2023,16, FOR PEER REVIEW 7 of 17 7 of 1

operating scheduling. The general steps of MPC rolling optimization can be expressed as
Step 1. Based on the current moment and the current system state, the system stat
follows:
in the
Stepfuture period
1. Based is obtained
on the by a certain
current moment and theprediction model.
current system state, the system state in
Stepperiod
the future 2. Based on the system
is obtained stateprediction
by a certain in a future period, the optimization problem in tha
model.
Stepis2.solved
period Based on the system
to obtain the state in asequence
control future period, the period.
in that optimization problem in that
periodStep
is solved to obtain the control sequence in that period.
3. Only the action of the first moment of the control sequence is applied to th
Step
system, and3. Only
the the action
above of the
steps arefirst moment
repeated forofthe
thenext
control sequence is applied to the
moment.
system, and the above steps are repeated for the next moment.
3.2.Total
3.2. TotalFramework
Framework of the
of the DNN-Based
DNN-Based Intra-Day
Intra-Day Scheduling
Scheduling Decision
Decision Method Method
Theoverall
The overall framework
framework of the
of the DNN-based
DNN-based intra-day
intra-day scheduling
scheduling methodmethod
is shownisinshown in
Figure3,3,which
Figure which mainly
mainly includes:
includes: the training
the training dataset
dataset construction
construction stage,training
stage, offline offline training
stage, and online decision stage.
stage, and online decision stage.

Figure3.3.The
Figure The framework
framework of the
of the proposed
proposed intra-day
intra-day method.
method.

(1)
(1) Training
Training dataset
datasetconstruction
constructionstage. To improve
stage. the accuracy
To improve and reduce
the accuracy and the pressure
reduce the pressur
on the network’s generalizability, the numerous real operating data
on the network’s generalizability, the numerous real operating data of MG collectedof MG collected
are clustered by the K-means algorithm [29], dividing into different training sets. The
are clustered by the K-means algorithm [29], dividing into different training sets. Th
net system load demand Pnet Load , net
which is a 96-dimensional time series represented as
net
net system load demand
PLoad = PLoad − PWind − PPV , is P used
Load , which
as theisclustering
a 96-dimensional
index. time series represented a
PLoad =training
(2) Offline
net
PLoad − PWind − PA ,two-dimensional
stage. time series feature map containing the
is used as the clustering index.
system operation state PV
is constructed as the input for the CNN-Bi LSTM network.
(2) The
Offline training
optimal stage. plan
scheduling A two-dimensional time series
is the network’s output, feature
training map scheduling
multiple containing the sys
tem operation
decision networks state
withisdifferent
constructed
training as datasets.
the input for the CNN-Bi LSTM network. Th
optimal
(3) Online scheduling
decision plan
stage. The is the network’s
system’s output,
ultra-short-term training
prediction multiple
state scheduling
is combined with deci
the day-ahead operation plan and fed
sion networks with different training datasets. into the well-trained CNN-Bi LSTM network.
(3) The outputs
Online of the network
decision stage. Theare fine-tuned
system’s by a power correction
ultra-short-term algorithmstate
prediction to getisthe
combined
final scheduling decision.
with the day-ahead operation plan and fed into the well-trained CNN-Bi LSTM net
work. The outputs of the network are fine-tuned by a power correction algorithm to
get the final scheduling decision.

3.3. Introduction to Deep Neural Networks

Energies 2023, 16, 7635 8 of 17

Energies
Energies 2023,
2023, 16,
16, xx FOR
FOR PEER
PEER REVIEW
REVIEW 88 of
of 17
17
3.3. Introduction to Deep Neural Networks
3.3.1. Convolutional Neural Networks
The
The efficient
The efficient feature
efficient feature extraction
feature extraction ability
extraction ability of
ability of the
of the CNN
the CNN makes
makes itit
CNN makes the
it the most
the most widely
most widely used
widely usedused
model
model in
in the
the field
field of
of deep
deep learning.
learning. The
The CNN
CNN primarily
primarily comprises
comprises a
a
model in the field of deep learning. The CNN primarily comprises a convolutional layer convolutional
convolutional layer
layer
and
and aaapooling
and poolinglayer.
pooling layer. The
layer. The convolutional
The convolutional layer
convolutional layer performs
layer performs effective
performs effective nonlinear
effective nonlinear local
nonlinear local feature
local feature
feature
extraction
extraction using
extraction using convolutional
using convolutional kernels,
convolutional kernels, while
kernels, while the
while the pooling
the pooling layer
pooling layer compresses
layer compresses
compresses the the extracted
the extracted
extracted
features
features and
features generates
and generates
generates moremoresignificant
more significantfeature
significant featureinformation
feature information
information toto
to enhance
enhance
enhance generalization
generalization
generalization ca-
ca-
capability
pability [30].
[30]. TheThe basic
basic structure
structure ofof the
the CNN
CNN is is shown
shown in
pability [30]. The basic structure of the CNN is shown in Figure 4. in Figure
Figure 4.4.

Figure 4.
Figure4.
Figure The
4.The basic
Thebasic structure
basicstructure of
structureof the
ofthe CNN.
theCNN.
CNN.

3.3.2.
3.3.2.Bidirectional
3.3.2. BidirectionalLong
Bidirectional Longand
Long andShort-Term
and Short-TermMemory
Short-Term MemoryNetworks
Memory Networks
Networks
We
We start by introducing the LSTM network, whichcontains
We start
start by
by introducing
introducing the
the LSTM
LSTM network,
network, which
which containsforgetting
contains forgettinggates,
forgetting gates,input
gates, input
input
gates,
gates, and
and output
output gates,
gates, and
and the
the basic
basic structure
structureis
isshown
shown in
inFigure
Figure
gates, and output gates, and the basic structure is shown in Figure 5. 5.5.

Figure 5.
Figure5. The
5.The basic
Thebasic structure
basicstructure of
structureof the
ofthe LSTM.
theLSTM.
LSTM.
Figure
In
In Figure
InFigure
Figure5,5,σ
5, andand
tanhtanh
and tanh represent
represent
represent Sigmoid
Sigmoid
Sigmoid and
and Tanh
and Tanh Tanh activation
activation
activation functions,
functions,
functions, respec-
respec-
respectively.
tively.
tively. The
The calculation
calculation of
of the
the data
data within
within LSTM
LSTM is
is
The calculation of the data within LSTM is as follows:as
as follows:
follows:
f =  (W [h , x ] + b )
=W (W f [ht −11 , xtt ] + bff ) (21)
ft =ftσ
t (
f [htf −1t,−x t] + b f )
(21)
(21)

iitt ==((W
Wii[[hhtt−−11,, xxtt ]]++bbii)) (22)
(22)
i t = σ ( Wi [ h t − 1 , x t ] + bi ) (22)
ZZtt == tanh(
tanh(WWcc[[hhtt−−11,, xxtt ]]++bbcc)) (23)
(23)
Zt = tanh(Wc [ht−1 , xt ] + bc ) (23)
CCtt == fftt 
CCtt−−11 ++ iitt 
 ZZtt (24)
(24)
Ct = ft ⊗ Ct −1 + it ⊗ Zt (24)
oott ==((W
Woo[[hhtt−−11,, xxtt ]]++bboo)) (25)
(25)
ot = σ (Wo [ ht −1 , xt ] + bo ) (25)
hhtt == oott 
tanh(
tanh(CCtt )) (26)
(26)
ht = ot ⊗ tanh(Ct ) (26)
where W
where W and and bb denote
denote the the weight
weight matrix
matrix and
and bias
bias vector, respectively. 
vector, respectively.  repre-
repre-
where W and b denote the weight matrix and bias vector, respectively. ⊗ represents dot
sents
sents dot
dothtproduct.
product. hhtt−−11 and h denote
andthehttoutput of the
denote the output
andof
output of the
the last
last and
and current
current moments,
moments, Ctre-
re-
product. −1 and ht denote last current moments, respectively. − 1
spectively. C
C t−1 and CC denote the memory state of the last and current
spectively. t−1 and t denote the memory state of the last and current moments, re-
t moments, re-
spectively.
spectively. ZZtt isis the
the Intermediate
Intermediate state
state of
of the
the network. and oott denote
network. iitt and denote that
that the
the
Energies 2023, 16, x FOR PEER REVIEW 9 o

Energies 2023, 16, 7635 9 of 17

current states add degree and output degrees, respectively. x t is the input of the cur
and Ct denote
moment.  the and memory
tanh state of the last
represent theand current moments,
sigmoid activation Zfunctions,
and tanhrespectively. t is the resp
Intermediate
tively. state of the network. i t and o t denote that the current states add degree and
output degrees, respectively. xt is the input of the current moment. σ and tanh represent
The LSTM structure gathers feature information only from the current input and p
the sigmoid and tanh activation functions, respectively.
timeTheseries
LSTMat each timegathers
structure whilefeature
disregarding
information feature
only information
from the current from future
input time serie
and past
this
timepaper,
series atbidirectional
each time whileLSTM is used feature
disregarding as the back-end
information mapping
from futurenetwork of the
time series. In schedu
decision
this paper,network to improve
bidirectional the as
LSTM is used accuracy
the back-endof the decision
mapping results
network andscheduling
of the the performanc
decision network
temporal featuretoextraction.
improve theTheaccuracy
Bi LSTM of the
is decision
a variant results and the
structure of performance of
LSTM that includes b
temporal feature extraction. The Bi LSTM is a variant structure of LSTM that includes both
forward LSTM and backward LSTM layers [31]. The Bi LSTM structure enables it to ga
forward LSTM and backward LSTM layers [31]. The Bi LSTM structure enables it to gather
information from both forward and backward directions, enabling the network to c
information from both forward and backward directions, enabling the network to consider
sider past and
past and future data. future
Thisdata. This
enhances theenhances the model’s
model’s feature extractionfeature extraction
ability without ability with
requiring
requiring additional
additional data. data. The
The structure of Bistructure of Bi LSTM
LSTM is illustrated is illustrated
in Figure 6. in Figure 6.

Figure 6.6.The
Figure Thestructure of the
structure Bi LSTM.
of the Bi LSTM.
3.4. The CNN-Bi LSTM Intra-Day Scheduling Decision Network
3.4. The CNN-Bi
Trained LSTM
by a large Intra-Day
amount of real Scheduling Decision
operation data, Network
the CNN-Bi LSTM intra-day schedul-
Trained
ing decision by a large
network amount
can learn of real operation
the regularity between thedata, thestate
system CNN-Bi LSTM
and the intra-day sch
scheduling
decision result. Once the parameters are fixed in the network, it can provide the
uling decision network can learn the regularity between the system state and the schoptimal
scheduling plan extremely fast under any operating scenario.
uling decision result. Once the parameters are fixed in the network, it can provide
optimal scheduling
3.4.1. Input and Outputplan extremely
of the fast under any operating scenario.
CNN-Bi LSTM
The CNN-Bi LSTM scheduling decision network imitates the idea of MPC for intra-day
3.4.1.
rollingInput and Output
optimization, of the
with the CNN-Bi
prediction LSTM
domain set to 2 h and the control domain set to
15 min. ToCNN-Bi
The deeply mine the implicit
LSTM value information
scheduling in the system
decision network operating
imitates data,of
the idea weMPC
set for in
the input X of this network in the form of a 2-D time series grayscale graph. The output Y
day rolling optimization, with the prediction domain set to 2 h and the control domain
of the network is the optimal scheduling plan. The specific expression is as follows:
to 15 min. To deeply mine the implicit value information in the system operating data
set the input X of this network in the form of a 2-D time series grayscale graph.
X = [PLoad (t, t + τoutput
), PWind (t,Yt +ofτ )the (t, t + τ ), Pisahead
, PPVnetwork t + τ ), Pahead
SBthe( t,optimal UG (t, t + τ ), P
scheduling ahead ( t, t + τ )]0
plan.
MT (27)
The specific expression i
follows: ∗ ∗ ∗
Y = [PSB (t, t + τ ), PUG (t, t + τ ), PMT (t, t + τ )]0 (28)
X = [ PLoad (t , t +Xis),aP9Wind t +1) ),
∗ ((τt ,+ PPV (t, t +  ), P
matrix consisting
ahead
(t, t +  ), P
SB of the
ahead
intra-day (t, t +  ), P
UG state
ahead
vectorMT
of the t +  )]' (power
(t, system
load, WT, and PV) in the period t to t + τ and the day-ahead operating plan vector of
controllable devices (SB, UG, and *MTs) in the corresponding time. The number 9 indicates
the number ofYinput= [ PSB
*
(t , t +  Y), isPUG
features. a 1(t∗, t6(+τ+
*
),1P)MT
vector  )]'
(t , t +consisting of the controllable devices’
intra-day optimal operating plan in the period t to t + τ. The number 6 indicates the
number a 9*( + 1)devices
X ofiscontrollable matrix consisting
in output of the
features. Pahead
 intra-day state vector
is the day-ahead of theplan,
operating system (po
∗
load, WT, and PV) in the period t to t +  and the day-ahead operating plan
and P is the intra-day optimal operating plan. Since the MPC prediction domain is set to vecto
2 h, the τ is set as 7 in this paper and all the above variables are real.
controllable devices (SB, UG, and MTs) in the corresponding time. The number 9 indic
the number of input features. Y is a 1*6( + 1) vector consisting of the controllable
vices’ intra-day optimal operating plan in the period t to t +  . The number 6 indic
ahead
the number of controllable devices in output features. P is the day-ahead opera
*
plan, and P is the intra-day optimal operating plan. Since the MPC prediction dom
Energies2023,
Energies 2023,16,
16,7635
x FOR PEER REVIEW 10
10 of 17
of 17

3.4.2. Structure
3.4.2. Structure ofof the
the CNN-Bi
CNN-Bi LSTMLSTM
Since the
Since the mapping
mapping relationship
relationship between
between the the system
system operating
operating state
state and
and scheduling
scheduling
decision is
decision is complex,
complex, this
this paper
paper uses
uses aa multilayer
multilayer CNN-Bi
CNN-Bi LSTMLSTM network
network for for deep
deep mining
mining
of the
of the data.
data. This network
network is is mainly
mainly constituted
constituted by by aa three-layer
three-layer CNNCNN and and aa three-layer
three-layer Bi
Bi
LSTM, and
LSTM, and linked
linked byby aa Flatten
Flatten layer.
layer. The
The CNN
CNN primarily
primarily extracts
extracts thethe power
power correlation
correlation
feature, while
feature, while the
the Bi
Bi LSTM
LSTM focuses
focuses onon extracting
extracting thethe power
power time
time series
series feature.
feature. TheThe batch
batch
normalization
normalization (BN) (BN) layer
layer can
can solve
solve the
the problem
problem of of numerical
numerical instability
instability inin DNNs,
DNNs, making
making
the
the distribution
distributionof of individual
individual features
featuresin in the
the same
same batch
batch similar.
similar. In
In this
this paper,
paper, the
the BN
BN layer
layer
is
is inserted
insertedbetween
betweeneacheachconvolutional
convolutionallayerlayerand andpooling
pooling layer
layerto to
normalize
normalize thethe
features in
features
the network
in the networkandand
accelerate training.
accelerate The dropout
training. The dropoutlayer layer
is the is
layer
the used
layerafter
usedeachafterBieach
LSTMBi
to
LSTMenhance the generalization
to enhance performance
the generalization of the network.
performance Finally, Finally,
of the network. the datathearedata
adjusted to
are ad-
ajusted
vectortooutput
a vectorinoutput
the specified size through
in the specified size athrough
fully connected (Dense) layer.
a fully connected (Dense) The specific
layer. The
structure of the proposed CNN-Bi LSTM in this paper is shown
specific structure of the proposed CNN-Bi LSTM in this paper is shown in Figure 7. in Figure 7.

Figure 7. The
Figure 7. The structure
structure of
of the
the CNN-Bi
CNN-Bi LSTM.
LSTM.
3.4.3. Settings of the CNN-Bi LSTM
3.4.3. Settings of the CNN-Bi LSTM
To better extract and abstract the input feature, the number of convolutional kernels is
To better extract and abstract the input feature, the number of convolutional kernels
set to 64, 128, and 256, and the size of convolutional kernels is set to 7 × 7, 5 × 5, and 3 × 3.
is set to 64, 128, and 256, and the size of convolutional kernels is set to 7 × 7, 5 × 5, and 3 ×
The number of neurons of Bi LSTM is set to 256, 128, and 64, respectively, and the drop rate
3. The number of neurons of Bi LSTM is set to 256, 128, and 64, respectively, and the drop
of the dropout layer is set uniformly to 0.25.
rate of the dropout
Normalize the layer is set
training uniformly
data to 0.25.to between 0 and 1 using the maximum–
of the network
Normalize the training data of the
minimum normalization method. The network is network to trained
betweenusing 0 andthe
1 using
Adam the maximum–
optimization
minimum normalization method. The network is trained using the Adam
algorithm [32] and the root mean square error (RMSE) is set as the loss function of the optimization
algorithmwhich
network, [32] and the root
is defined as mean square error (RMSE) is set as the loss function of the
follows:
network, which is defined as follows:
v
uM T
 
u M T
RMSE==t ∑ ∑ (y∗ ( y−i*,t y−i,ty)i2,t /T
RMSE i =1 t =i,t
1
)2 / T (29)
(29)
i =1 t =1
*
where ∗yi , t and yi , t are the true and predicted scheduling plans for i th device at time
where yi,t and y are the true and predicted scheduling plans for ith device at time t,
t , respectively. i,t
M is the number of controllable devices in MG.
respectively. M is the number of controllable devices in MG.
3.5. The
3.5. The Power
Power Balance
Balance Correction
Correction Algorithm
Algorithm
Like load
Like load forecasting,
forecasting, the
the DNN-based
DNN-based scheduling
scheduling method
method is is fundamentally
fundamentally aa process
process
of nonlinear
of nonlinearregression.
regression.Consequently,
Consequently, the
the output
output inevitably
inevitably does
does notnot meet
meet certain
certain prac-
practical
tical constraints.
constraints. To address
To address this issue,
this issue, we ause
we use a power
power balance
balance correction
correction algorithm
algorithm (PBC)
(PBC) to
to adjust the output, making it practical
adjust the output, making it practical for use.for use.
Inspiredby
Inspired by the
the average
average consistency
consistency algorithm,
algorithm, wewe utilize
utilize the
the difference
difference between
between total
total
power demand
power t
demand and total generation at time t asasthe theconsistency
consistencyindicator.
indicator. The
The outputs
outputs
from DNN
from DNN areare updated
updated byby iteration
iteration (Equations
(Equations (30)
(30) and
and (31)).
(31)). Any
Any updated
updated results
results that
that
Energies 2023, 16, 7635 11 of 17

violate the operating constraints of the device require additional correction (Equation (32)).
This algorithm is denoted as follows:
K
− ∑i=1 Pi,t )/K
[n] net [n]
δt = ( PLoad,t (30)

[ n +1] [n] [n]

Pi,t = Pi,t + δt (31)

[ n +1]



 Pimax , Pi,t > Pimax
[ n +1]
Pimin , Pi,t < Pimin





[ n +1] [ n +1] up [ n +1] [ n +1] up
Pi,t = Pi,t + Pi , Pi,t − Pi,t−1 > Pi (32)
[ n +1] [ n +1] [ n +1]



 Pi,t down
− Pi , Pi,t − Pi,t−1 < Pidown


 [ n +1]
Pi,t , otherwise


[n]
where n is the number of iterations. Pi,t and K are the power of ith power generator at
time t and the total number of generators in MG, respectively.
So far, MG’s intra-day optimal scheduling model based on CNN-Bi LSTM-PBC is
completely constructed.

4. Simulation Analysis and Comparison

4.1. Introduction of Example Parameters and Test System
To verify the effectiveness of the proposed method, a typical grid-connected
MG system is used for simulation testing. This MG comprises one WT, PV, SB, UG power
line, and MT. The parameter settings are shown in Appendix A.
The wind, solar, and load data in this paper were taken from a power station within a
region of Guizhou for a total of 356 days. The scheduling decision data for network training
were constructed and solved using Yalmip with Cplex solver. The DNN was built, trained,
and evaluated based on the Matlab R2020a platform. Setting for 300 rounds of training with
variable learning rate training and an initial learning rate of 0.01. Simulation tests were
conducted under 11th Gen Intel(R) Core (TM) i5-11300H @ 3.10 GHz 3.11 GHz processor.

4.2. Effectiveness Analysis of the Proposed Method

The optimal number of K-means clusters was determined by the ‘elbow method’ to be
3. A randomly selected day in each data category is used as a test scenario.
The 3 types of scenarios are tested as Table 1. The operating cost and RMSE are used
as the effectiveness evaluation indicators of the proposed method in this paper. The smaller
these two indicators are, the better the scheduling plan is.

Table 1. Results of 3 types of DNN-based scheduling method.

Training
Full-Day
Scene Type Sample Training Time/s RMSE
Decision Time/s
Size/Day
1 165 278.18 5.7683 0.3599
2 91 171.54 8.9152 0.3761
3 106 196.42 7.3725 0.3698

For illustrative purposes, scenario 1 is analyzed as an example scenario. The day-

ahead optimal operating cost of MG based on MILP is CNY 12,621. MG’s intra-day optimal
operating cost based on CNN-Bi LSTM-PBC is CNY 15,320, and the full-day MPC rolling
optimization solving time is 0.4636 s. The optimal scheduling plan can be seen in Figure 8.
 reasonable and effective scheduling decisions in a very short time. The DNN-based sched-
uling decision method utilizes extensive training with historical decision data to establish
a direct mapping relationship between known inputs and decision outcomes. Once the
sample
Energies capacity and quality can be guaranteed, the method can fit any kind of scheduling
2023, 16, 7635 12 of 17

decision model with high applicability.

Figure 8. MG two-layer optimal

Figure scheduling
8. MG two-layer results.
optimal scheduling results.

The shapes of the day-ahead scheduling curves and the intra-day scheduling curves
4.3. Comparative Analysis of Different
are roughly Methods
similar, indicating that the CNN-Bi LSTM-PBC can complete the output of rea-
sonable and effective scheduling decisions in a very short time. The DNN-based scheduling
4.3.1. Comparison with Traditional Methods
decision method utilizes extensive training with historical decision data to establish a direct
In the intra-day rolling
mapping optimization
relationship betweenpart,
knowntheinputs
traditional mathematical
and decision outcomes. Once model-based
the sample
capacity and quality can be guaranteed, the method can fit any kind of scheduling decision
MPC, CNN-Bi SLTM, and CNN-Bi SLTM-PBC are compared. The results are shown in
model with high applicability.
Table 2.
4.3. Comparative Analysis of Different Methods
Table 2. Effectiveness4.3.1. Comparison
comparison of with
eachTraditional
intra-dayMethods
method.
In the intra-day rolling optimization part, the traditional mathematical model-based
Calculation MPC, CNN-Bi SLTM,
Time/s Operating and CNN-BiRMSE
Cost/CNY SLTM-PBC Doesare compared. The results
it Violate are shown in
the Constraints?
Table 2.
PC 125.4531 14,973 0 No
M 0.3599 Table 2. Effectiveness
15,779 comparison of each intra-day method.
5.7683 Yes
PBC 0.4636 Method
15,320Calculation 3.6411
Operating
RMSE
No Does it Violate
Time/s Cost/CNY the Constraints?
Traditional MPC 125.4531 14,973 0 No
The operating costs
CNN-Bicalculated
LSTM by CNN-Bi LSTM
0.3599 15,779and CNN-Bi 5.7683LSMT-PBCYes are only
5.38% and 2.32% higherCNN-Bi
than the traditional
0.4636 MPC method,15,320 respectively,
3.6411 but the calculation
No
LSTM-PBC
efficiency is improved by about 300 times. This indicates that the DNN-based scheduling
decision network will Theimitate the actual optimal scheduling operation plan and greatly
operating costs calculated by CNN-Bi LSTM and CNN-Bi LSMT-PBC are only
reduce the difficulty
5.38%ofand
solving the optimal
2.32% higher scheduling
than the traditional problem
MPC method, through
respectively, but training and
the calculation
high-dimensional nonlinear mapping.
efficiency is improved by about 300 times. This indicates that the DNN-based scheduling
decision network will imitate the actual optimal scheduling operation plan and greatly
reduce the difficulty of solving the optimal scheduling problem through training and
4.3.2. Influence of the Training Dataset
high-dimensional Capacity
nonlinear mapping.
Changing the 4.3.2.
number of of
Influence samples in Dataset
the Training the training
Capacitydata for the training of the data-
driven scheduling decision model,
Changing the ofresulting
the number samples in model is data
the training tested using
for the the
training same
of the typical
data-driven
day test samples. The resultsdecision
scheduling are shown
model,intheTable 3. model is tested using the same typical day test
resulting
samples. The results are shown in Table 3.
Table 3. Test results of different dataset capacities.

Capacity/Day Training Time/s RMSE

30 79.71 16.5679
Energies 2023, 16, 7635 13 of 17

Table 3. Test results of different dataset capacities.

Capacity/Day Training Time/s RMSE

30 79.71 16.5679
60 133.47 11.4329
90 184.55 8.9458
120 223.78 7.3893
Energies 2023, 16, x FOR PEER REVIEW 150 251.72 6.0256
13 of 17

The increasing training time and decreasing RMSE indicate that the performance
and The increasing
decision training
accuracy oftime
the and decreasingmodel
data-driven RMSE indicate
increase that the performance
with the increaseand in sample
decision accuracy
Energies 2023, 16, x FOR PEER REVIEW
capacity, i.e., the of
DNNthe data-driven
is constantly model increase
evolving and with the increase in
self-correcting assample
13 of 17 capacity,
the number of samples
i.e., the DNN isFigure
accumulates. constantly evolving
9 shows theand self-correcting
comparison as thethe
between number of samples
full-day running accu-
cost of the
mulates. Figure 9 shows the comparison between the full-day running cost of the modified
modified data-driven
The increasing
model
training time
and the realRMSE
andfull-day
decreasing
full-day running cost. With the increase in sample
data-driven model and the real running indicate thatthe
cost. With the increase
performance and
in sample capac-
capacity,
decision the operating
accuracy of the cost of
data-driven data-driven
model increase withdecision-making
the increase in sampleresults
capacity,is approaching the
ity, the operating cost of data-driven decision-making results is approaching the optimal
optimal
scheduling scheduling
i.e., the DNN
operating operating
is constantly
driven cost
evolving and
cost driven by the
self-correcting
by the traditional traditional
as the
MPC number of MPC
method. samples method.
accu-
mulates. Figure 9 shows the comparison between the full-day running cost of the modified
data-driven model and the real full-day running cost. With the increase in sample capac-
ity, the operating cost of data-driven decision-making results is approaching the optimal
scheduling operating cost driven by the traditional MPC method.

Figure 9.
Figure 9.Operating
Operatingcosts comparison
costs under
comparison different
under training
different datasetdataset
training capacities.
capacities.
Figure 9. Operating costs comparison under different training dataset capacities.
4.3.3. InfluenceofofData
4.3.3. Influence Data Clustering
Clustering
Comparing
4.3.3. Comparing
Influence of Data
thethe two scenarios
Clustering
two scenarios of clustered
of clustered and unclustered,
and unclustered, with the with
other thesettings
other settings
being
being thesame,
the
Comparingsame, theunclustered
thethe
two unclustered
scenarios of one one randomly
randomly
clustered selectsselects
and unclustered,160with160 non-repeated
non-repeated
the other days from
settings daysthefrom the
being
wholethe
whole same,data
year’s
year’s the
dataunclustered
as as
thethe one samples.
training randomly
training selects
To
samples. 160
To non-repeated
illustrate the impact
illustrate days
the from the
ofimpact
clustering of more in-
clustering more
whole year’s
tuitively on data
the as thescheduling
final training samples. To illustrate
decision’s the impact
accuracy, the DNN of clustering
output ismore
not in-
corrected us-corrected
intuitively on the final scheduling decision’s accuracy, the DNN output is not
tuitively
ing PBC. onThe
the power
final scheduling
balance decision’s
for both accuracy, the DNN output is10.not corrected us-
using
ing PBC. PBC. The power
The power balance balance forcases
for both cases both iscases
is shown
shown in Figure
is shown
in Figure 10. in Figure 10.

Figure 10.10.
Figure Comparison of clustered
Comparison training data
of clustered with unclustered
training data withtraining data. training data.
unclustered
Figure
The10. Comparison
output of clusteredscheduling
of the DNN-based training data with network
decision unclustered training
trained withdata.
unclustered
data deviates more from the real electrical load demand, i.e., it indicates that its decision
The output of the DNN-based scheduling decision network trained with unclustered
data deviates more from the real electrical load demand, i.e., it indicates that its decision
Energies 2023, 16, 7635 14 of 17

The output of the DNN-based scheduling decision network trained with unclustered
data deviates more from the real electrical load demand, i.e., it indicates that its decision
accuracy is low. This is because the generalization of the current deep learning model
cannot cope with such large scenario differences. If a DNN is used for training, it will
generate a unique compromise mapping model during the training process in the face of
very different historical sample data, which makes it difficult to guarantee decision-making
accuracy.

4.3.4. Influence of Different Backend Mapping DNN Models

Different backend mapping networks were trained and then tested for comparison on
the same samples, with the rest of the settings being the same. The results are shown in
Table 4.

Table 4. Test results of different DNNs.

Network Type Training Time/s RMSE

Single RNN 172.92 16.8767
Single LSTM 186.33 13.0241
Single Bi LSTM 213.45 10.4537
Double RNN 207.78 13.8254
Double LSTM 212.22 9.3126
Double Bi LSTM 231.89 7.8772
Triple RNN 247.55 10.6133
Triple LSTM 253.34 7.4122
Triple Bi LSTM 278.18 6.0256

Among them, the decision accuracy of the RNN always lags behind that of LSTM and
Bi LSTM, due to its simple structure, which leads to its inability to discard unimportant
information, and its tendency to suffer from the gradient explosion problem during the
training process. The decision accuracy of Bi LSTM is always higher than that of LSTM,
which is because Bi LSTM has both a forward LSTM and an inverse LSTM at the same time
node in the implicit layer, which has two more parameters and both before and after 2-time
nodes affect its output results, so it has more energy to analyze the information.

5. Discussion
This paper proposes a two-layer and DNN-based optimal scheduling decision-making
method for MG that addresses the limitations of traditional mathematical model-based
methods. Instead of studying the intrinsic mechanism of the optimization problem, the
method is based on the DNN network, which uses massive historical decision data training
to directly construct the mapping relationship between known inputs and decision results.
This approach breaks through traditional optimal scheduling solution thinking and pro-
vides a new way of MG optimal scheduling. The analysis of computational examples leads
to the following conclusions:
(1) Using the classified data to train different DNN models separately can effectively
improve the scheduling decision accuracy of CNN-Bi LSTM and prevent the models
from converging to a compromise solution with lower accuracy.
(2) The DNN-based scheduling method achieves the optimal scheduling decision by
mapping, thereby reducing the complexity and improving the efficiency of solving
the optimal scheduling problem. Moreover, as the training dataset capacity increases,
the decision accuracy of the method continues to improve.
(3) To address the issue of DNN method output results not meeting practical constraints,
the PBC model effectively rectifies the output results, which greatly enhances the
practical applicability of the DNN-based scheduling method.
In conclusion, as a novel and efficient solution algorithm, the method proposed
in this paper can provide a practical and reliable reference for MG scheduling centers
Energies 2023, 16, 7635 15 of 17

to assist decision making. This will greatly improve the reliability of operations in the
uncertain environment of MG scheduling and the economics of scheduling decisions.
In future research, we will further explore the relationship between system operation
state and scheduling decisions and try to construct complex DNN models with ‘attention
mechanisms’ to improve the decision-making accuracy of data-driven methods. In addition,
solving the optimal scheduling problem for multiple interconnected MGs will also be
considered.

Author Contributions: F.C.: Conceptualization, Methodology, Analysis, Writing-Original Draft.

Z.W.: Investigation, Data Curation, Model Analysis, Writing. Y.H.: Writing—Review and Editing,
Resources, Supervision. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported in part by the Science and Technology Foundation of Guizhou
Province ((2022) general 014).
Data Availability Statement: Third Party Data. Restrictions apply to the availability of these data.
Data were obtained from China Southern Power Grid Guizhou Power Grid and are available from
Fei Chen with the permission of China Southern Power Grid.
Conflicts of Interest: Author Fei Chen was employed by the company China Southern Power Grid
Guizhou Power Grid Co., Ltd. The remaining authors declare that the research was conducted
in the absence of any commercial or financial relationships that could be construed as a potential
conflict of interest. The funders had no role in the design of the study; in the collection, analysis, or
interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Appendix A

Table A1. Technical and cost parameters of MT.

Parameters MT 1 MT 2 MT 3 MT 4
Capacity/kW 100 150 200 250
Cost factor a/(CNY·kWh−2 ) 0.00275 0.00202 0.0016 0.0013
Cost factor b/(CNY·kWh−1 ) 0.575 0.475 0.455 0.375
Cost factor c/CNY 50 75 100 125
Start-up cost/CNY 50 75 100 125
Ramping power/kWh−1 25 40 50 65
Minimum output/kW 10 15 20 25

Table A2. Technical and cost parameters of SB and UG.

Parameters SB Parameters UG
Maximum purchased
Capacity/kW 500 300
power/kW
Cost factor/CNY·kWh−1 0.06 Maximum sales power/kW −300
Maximum
charging/discharging 100
power/kW
Ramping power/kWh−1 50
Maximum state of charge/% 0.95
Minimum state of charge/% 0.05
Charging/discharging
0.9
efficiency
Initial state of charge/% 0.5
Energies 2023, 16, 7635 16 of 17

Table A3. Time-of-use power price.

Time Period Type Time Period Powe Purchase Price Powe Sales Price
Peak 17:00–23:00 0.70 0.55
Mean 8:00–17:00 0.61 0.42
Valley 23:00–8:00 0.45 0.28

References
1. Zheng, B.; Wu, X. Integrated capacity configuration and control optimization of off-grid multiple energy system for transient
performance improvement. Appl. Energy 2022, 311, 118638. [CrossRef]
2. Yang, P.; Peng, S.; Benani, N.; Dong, L.; Li, X.; Liu, R.; Mao, G. An integrated evaluation on China’s provincial carbon peak and
carbon neutrality. J. Clean. Prod. 2022, 377, 134497. [CrossRef]
3. Chang, L.U.; Li, G.; Yixin, L.; Shuang, G.; Bin, X.U. Distributed Optimal Dispatching Method for Independent Microgrids Based
on Flexible Interconnection. Power Syst. Technol. 2019, 43, 1512–1519.
4. Ma, Z.; Xiao, M.; Xiao, Y.; Pang, Z.; Poor, H.V.; Vucetic, B. High-Reliability and Low-Latency Wireless Communication for Internet
of Things: Challenges, Fundamentals, and Enabling Technologies. IEEE Internet Things J. 2019, 6, 7946–7970. [CrossRef]
5. Li, B.; Roche, R.; Miraoui, A. Microgrid sizing with combined evolutionary algorithm and MILP unit commitment. Appl. Energy
2017, 188, 547–562. [CrossRef]
6. Zhu, J.; Mo, X.; Zhu, T.; Guo, Y.; Liu, M. Real-time stochastic operation strategy of a microgrid using approximate dynamic
programming-based spatiotemporal decomposition approach. IET Renew. Power Gener. 2019, 13, 3061–3070. [CrossRef]
7. Duan, Y.; Zhao, Y.; Hu, J. An initialization-free distributed algorithm for dynamic economic dispatch problems in microgrid:
Modeling, optimization and analysis. Sustain. Energy Grids Netw. 2023, 34, 101004. [CrossRef]
8. Li, P.; Xu, D.; Zhou, Z.; Lee, W.J.; Zhao, B. Stochastic Optimal Operation of Microgrid Based on Chaotic Binary Particle Swarm
Optimization. IEEE Trans. Smart Grid 2017, 7, 66–73. [CrossRef]
9. Zhang, F.; Shen, Z.; Xu, W.; Wang, G.; Yi, B. Optimal Power Flow Algorithm Based on Second-Order Cone Relaxation Method for
Electricity-Gas Integrated Energy Microgrid. Complexity 2021, 2021, 2073332. [CrossRef]
10. Lu, J.; Liu, T.; He, C.; Nan, L.; Hu, X. Robust day-ahead coordinated scheduling of multi-energy systems with integrated
heat-electricity demand response and high penetration of renewable energy. Renew. Energy 2021, 178, 466–482. [CrossRef]
11. Guo, L.; Liu, W.; Jiao, B.; Hong, B.; Wang, C. Multi-objective stochastic optimal planning method for stand-alone microgrid
system. IET Gener. Transm. Distrib. 2014, 8, 1263–1273. [CrossRef]
12. Ding, Y.F.; Morstyn, T.; Mcculloch, M.D. Distributionally Robust Joint Chance-Constrained Optimization for Networked Micro-
grids Considering Contingencies and Renewable Uncertainty. IEEE Trans. Smart Grid 2022, 13, 2467–2478. [CrossRef]
13. Melhem, F.Y.; Grunder, O.; Hammoudan, Z.; Moubayed, N. Energy Management in Electrical Smart Grid Environment Using
Robust Optimization Algorithm. IEEE Trans. Ind. Appl. 2018, 54, 2714–2726. [CrossRef]
14. Hosseini, S.M.; Carli, R.; Dotoli, M. Robust Day-ahead Energy Scheduling of a Smart Residential User under Uncertainty. In
Proceedings of the 2019 18th European Control Conference (ECC): 18th European Control Conference (ECC), Naples, Italy, 25–28
June 2019; IEEE: Piscataway, NJ, USA; pp. 935–940.
15. Pourmousavi, S.A.; Nehrir, M.H.; Sharma, R.K. Multi-Timescale Power Management for Islanded Microgrids Including Storage
and Demand Response. IEEE Trans. Smart Grid 2015, 6, 1185–1195. [CrossRef]
16. Acevedo-Arenas, C.Y.; Correcher, A.; Sanchez-Diaz, C.; Ariza, E.; Alfonso-Solar, D.; Vargas-Salgado, C.; Petit-Suarez, J.F. MPC
for optimal dispatch of an AC-linked hybrid PV/wind/biomass/H_2 system incorporating demand response. Energy Convers.
Manag. 2019, 186, 241–257. [CrossRef]
17. Peng, Y.H.; Jolfaei, A.; Yu, K.P. A Novel Real-Time Deterministic Scheduling Mechanism in Industrial Cyber-Physical Systems for
Energy Internet. IEEE Trans. Ind. Inform. 2022, 18, 5670–5680. [CrossRef]
18. Wang, W.; Yang, D.; Huang, N.; Lyu, C.; Zhang, G.; Han, X. Irradiance-to-power conversion based on physical model chain: An
application on the optimal configuration of multi-energy microgrid in cold climate. Renew. Sustain. Energy Rev. 2022, 161, 112356.
[CrossRef]
19. Dong, G.Z.; Chen, Z.H. Data-Driven Energy Management in a Home Microgrid Based on Bayesian Optimal Algorithm. IEEE
Trans. Ind. Inform. 2019, 15, 869–877. [CrossRef]
20. Khalil, R.A.; Saeed, N.; Masood, M.; Fard, Y.M.; Alouini, M.S.; Al-Naffouri, T.Y. Deep Learning in the Industrial Internet of Things:
Potentials, Challenges, and Emerging Applications. IEEE Internet Things J. 2021, 8, 11016–11040. [CrossRef]
21. Liu, H.Z.; Shen, X.W.; Guo, Q.L.; Sun, H.B. A data-driven approach towards fast economic dispatch in electricity-gas coupled
systems based on artificial neural network. Appl. Energy 2021, 286, 116480. [CrossRef]
22. Tushar, M.; Zeineddine, A.W.; Assi, C. Demand-Side Management by Regulating Charging and Discharging of the EV, ESS, and
Utilizing Renewable Energy. IEEE Trans. Ind. Inform. 2018, 14, 117–126. [CrossRef]
23. Mignoni, N.; Carli, R.; Dotoli, M. Distributed Noncooperative MPC for Energy Scheduling of Charging and Trading Electric
Vehicles in Energy Communities. IEEE Trans. Control Syst. Technol. 2023, 31, 2159–2172. [CrossRef]
Energies 2023, 16, 7635 17 of 17

24. Yang, N.; Yang, C.; Xing, C.; Ye, D.; Jia, J.J.; Chen, D.J.; Shen, X.; Huang, Y.H.; Zhang, L.; Zhu, B.X. Deep learning-based SCUC
decision-making: An intelligent data-driven approach with self-learning capabilities. IET Gener. Transm. Distrib. 2022, 16, 629–640.
[CrossRef]
25. Guo, F.H.; Xu, B.W.; Xing, L.T.; Zhang, W.A.; Wen, C.Y.; Yu, L. An Alternative Learning-Based Approach for Economic Dispatch
in Smart Grid. IEEE Internet Things J. 2021, 8, 15024–15036. [CrossRef]
26. Kim, M.J.; Kim, T.S.; Flores, R.J.; Brouwer, J. Neural-network-based optimization for economic dispatch of combined heat and
power systems. Appl. Energy 2020, 265, 114785. [CrossRef]
27. Hu, H.; Sotirov, R. The linearization problem of a binary quadratic problem and its applications. Ann. Oper. Res. 2021, 307,
229–249. [CrossRef]
28. Dong, S.; Wang, P.; Abbas, K. A survey on deep learning and its applications. Comput. Sci. Rev. 2021, 40, 100379. [CrossRef]
29. Li, M.; Xu, D.C.; Zhang, D.M.; Zou, J. The seeding algorithms for spherical k-means clustering. J. Glob. Optim. 2020, 76, 695–708.
[CrossRef]
30. Gu, J.X.; Wang, Z.H.; Kuen, J.; Ma, L.Y.; Shahroudy, A.; Shuai, B.; Liu, T.; Wang, X.X.; Wang, G.; Cai, J.F.; et al. Recent advances in
convolutional neural networks. Pattern Recognit. 2018, 77, 354–377. [CrossRef]
31. Shahid, F.; Zameer, A.; Muneeb, M. Predictions for COVID-19 with deep learning models of LSTM, GRU and Bi-LSTM. Chaos
Solitons Fractals 2020, 140, 110212. [CrossRef]
32. Barakat, A.; Bianchi, P. Convergence and Dynamical Behavior of the Adam Algorithm for Nonconvex Stochastic Optimization.
SIAM J. Optim. 2021, 31, 244–274. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.