IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO.

17, SEPTEMBER 1, 2021 13703

Machine-Learning-Based Real-Time Economic

Dispatch in Islanding Microgrids in a Cloud-Edge
Computing Environment
Wei Dong , Student Member, IEEE, Qiang Yang , Senior Member, IEEE,
Wei Li , Senior Member, IEEE, and Albert Y. Zomaya , Fellow, IEEE

Abstract—The paradigm of the Internet of Things (IoT) and construction cost [2]. The power generation intermittency of
cloud-edge computing plays a significant role in future smart massively penetrated distributed renewable power sources and
grids. The data-driven solution integrating the artificial intel- demand-side load randomness brings about technical chal-
ligence functionalities brings novel methods to address the
nontrivial task of economic dispatch in microgrids in the presence lenges to the stable and efficient operation for the energy
of uncertainties of renewable generations and loads. This arti- management system (EMS) of microgrids.
cle proposes a learning-based decision-making framework for the Islanding microgrids are often deployed to provide power
economic energy dispatch of an islanding microgrid based on the supply in remote locations (e.g., offshore islands and geo-
cloud-edge computing architecture. Cloud resources are utilized graphically remote areas) that lack electricity supply from the
to solve the optimal dispatch decision sequences over historical
operating patterns. It can be considered as a sample labeling pro- power utilities. The islanding operation mode of microgrids
cess for the supervised training that can implement the complex imposes a stringent requirement to maintain the dynamic
mapping of input–output space through an advanced machine balance of power generation and demand to ensure system
learning model. Then, the well-trained model can be adopted stability in the presence of various operational uncertainties.
locally at edge computing devices keeping the long-term param- The real-time energy dispatch aims to meet this require-
eters unchanged for implement the real-time microgrid energy
dispatch. The key benefit of the proposed solution is that it effec- ment through appropriate management of the energy storage
tively avoids the prediction of multiple stochastic variables and system (ESS) and controllable generators, e.g., diesel sources,
the design of sophisticated regulation strategies or reward pol- considering the uncertainties introduced by the intermittent
icy functions for real-time dispatch. The solution is extensively renewable sources and demands. Due to the limited avail-
assessed through simulation experiments by the use of real data ability of energy resources in the islanding microgrid, the
measurements for a set of operational scenarios and the numer-
ical results validate the effectiveness and benefit of the proposed economic dispatch is of paramount importance to minimize
algorithmic solution. the operational cost by improving the global energy utilization
efficiency while meeting the system operational constraints.
Index Terms—Data-driven control, economic dispatch,
machine learning, optimal dispatch. However, the optimal/economic energy dispatch in an
islanding microgrid is considered a nontrivial task due to
the system operational uncertainties. The power generations
I. I NTRODUCTION of renewable sources, e.g., WTs and PVs, are stochastic and
nondispatchable. Also, the power demand in such small-scale
ICROGRID is a small-scale power system consisting of
M a collection of power generation sources, storage units
and different loads that can operate in either grid-connected
microgrid often exhibits a higher degree of uncertainty com-
pared with that of the utility grid. The optimal energy dispatch
is often formulated as a multistage stochastic optimization
mode or islanding mode [1]. It enables efficient integration
problem [1]. Two key issues need to be addressed for an
of different forms of renewable distributed generators, e.g.,
economic dispatch solution: the computational complexity
micro wind turbines (WTs) and solar photovoltaic units (PVs),
of the real-time dispatch and the coordinated scheduling in
to reduce carbon emission, power loss, and infrastructure
an optimal way over multiple time periods. To address the
Manuscript received November 13, 2020; revised January 27, 2021 challenge of energy economic dispatch with operational uncer-
and February 28, 2021; accepted March 17, 2021. Date of publication tainties (e.g., intermittent generation and fluctuant demand),
March 22, 2021; date of current version August 24, 2021. This work was the existing solutions are mainly proposed based on two frame-
supported in part by the Natural Science Foundation of China under Grant
51777183. (Corresponding author: Qiang Yang.) works: 1) the model predictive control (MPC)-based energy
Wei Dong and Qiang Yang are with the College of Electrical Engineering, dispatch and 2) the Markov decision process (MDP)-based
Zhejiang Lab, Zhejiang University, Hangzhou 310027, China (e-mail: energy dispatch.
dong_w@zju.edu.cn; qyang@zju.edu.cn).
Wei Li and Albert Y. Zomaya are with the Australia–China Joint Research Based on stochastic optimization approaches, the MPC-
Centre for Energy Informatics and Demand Response Technologies, Centre for based framework can be adopted to implement the rolling
Distributed and High Performance Computing, School of Computer Science, optimization in the forthcoming transient horizon [3]. Also,
The University of Sydney, Camperdown, NSW 2008, Australia (e-mail:
weiwilson.li@sydney.edu.au; albert.zomaya@sydney.edu.au). considering the advantages of both long-term horizon of
Digital Object Identifier 10.1109/JIOT.2021.3067951 optimization and low uncertainty of real-time dispatch, a
2327-4662 
c 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
13704 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 17, SEPTEMBER 1, 2021

TABLE I
A NALYSIS OF THE R ELATED MPC AND MDP-BASED S OLUTIONS

two-stage model consisting of the day-ahead optimal schedul- Table I briefly summarizes the pros and cons of MPC and
ing using the prediction information and real-time operation MDP-based solutions.
with elaborate regulation strategies has been exploited and In recent years, along with the advances of smart grid
developed. In [4], a two-layer coordinated energy management development, the adoption of Internet-of-Things (IoT) devices
approach was proposed. The solution carried out an economic has brought tremendous data during the system operation.
operation scheduling in the day-head optimization layer and Due to the limited capacity of computation and storage, the
reduced the power regulation cost in the real-time layer to remote cloud computing facilities are often used to process
resolve the prediction errors. The study in [5] proposed a the data introduced from the field IoT devices. However, such
double-layer coordination approach in which the commitment a centralized paradigm results in significant communication
of generations and the prediction of reserve boundaries were delay and overhead. To overcome this issue, edge comput-
obtained in the day-ahead stage. Then these results can be ing brings the cloud in proximity to the IoT devices [12].
regulated through the successive approximation model in real- Unlike the conventional centralized cloud computing archi-
time. In [6], the upper layer optimized the look-ahead dispatch tecture, the cloud-edge paradigm enables the edge computing
for individual generators, and a Lagrangian formulation was devices to carry out the decision-making tasks for timely
implemented as the corrective strategy considering the oppor- energy management with significantly reduced communication
tunity cost of discharging and the marginal cost of charging. delay and congestion [13] [14]. Thus, the cloud-edge orches-
However, it should be noted that the MPC-based solutions tration/cooperation is considered a promising and efficient
often require accurate prediction of renewable generation and architecture to improve the energy management performance
power demand as well as the sophisticated designs of real-time in microgrids.
regulation strategies. The rolling optimization approach also Fig. 1 illustrates the proposed cloud-edge computing archi-
requires higher computing capacity for the local controller. tecture for the energy management of islanding microgrids.
To address the randomness and uncertainties induced by the Such an EMS architecture with ubiquitous sensing and com-
multiple stochastic variables, economic dispatch in microgrids puting devices effectively drives the exploitation and deploy-
is often considered and modeled in the MDP framework [7]. ment of more advanced operation and control functionali-
In most recent years, reinforcement learning (RL) has been ties for microgrids. The edge computing units can obtain
widely adopted as an effective solution for solving MDP the system operational states through device-to-edge com-
without any a priori knowledge or effort of stochastic munication with field devices, and make available to the
modeling [1]. In [8], a Q-learning-based algorithmic solu- cloud when necessary through edge-to-cloud communication.
tion was developed to implement the economic dispatch that Artificial intelligence (AI)-based solutions can be adopted in
relaxed the global information requirement and balanced the the intelligent orchestration in such a cloud-edge computing
exploration and exploitation in a distributed way. The study architecture [15]. Sufficient computation and storage resources
in [9] proposed a consensus Q-learning for decentralized gen- are available on cloud service to support the exhaustive anal-
eration commitment by exchanging the Q-value matrices and ysis of the massive historical data through a model training
exploiting the prior knowledge to accelerate the algorithm con- process. The well-trained model can be further made avail-
vergence rate. In [10], an RL-based economic dispatch solution able to the edge computing units in the field to fulfill the local
was proposed that combined the value-function approximation management functionalities in the microgrids.
using the distributed optimization based on multiplier splitting. The machine learning techniques are widely considered
However, the RL-based solutions that use the value function efficient for identifying the underlying patterns and knowl-
or policy function from the Bellman equation to achieve long- edge from the large volume of historical data. Our previous
term optimization can often fail into the local optimum. Also, work [16] comprehensively reviewed the state-of-the-art
a set of sophisticated learning policies and useful features are machine learning techniques and applications in smart grid,
generally needed to avoid high-dimensional state space [11]. e.g., forecasting of renewable generation forecasting [17] and

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
DONG et al.: MACHINE-LEARNING-BASED REAL-TIME ECONOMIC DISPATCH IN ISLANDING MICROGRIDS 13705

Fig. 2. Illustration of a smart islanding microgrid system.

3) The proposed solution is validated through simulations

for a range of scenarios, e.g., the ideal dispatch (without
operational uncertainties), and the results indicate that
the proposed solution outperforms the existing schemes,
Fig. 1. Cloud-edge computing environment for microgrid energy e.g., MPC-based solutions.
management. The remainder of this article is organized as follows. Section
II formulates the problem of the economic dispatch of island-
power demand [18], demand-side management [19], and fault ing microgrids. Section III presents the proposed solution in
diagnosis [20]. details. Section IV evaluates the proposed solution and pro-
Unlike the existing MPC and MDP-based energy man- vides the numerical results. Finally, the conclusive remarks are
agement solutions, this article proposes a learning-based given in Section V.
decision-making framework for microgrids energy dispatch in
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
a cloud-edge computing environment. The main idea behind
this work can be described as follows: the large number of data This work considers a multitype energy microgrid oper-
measurements collected by the IoT devices at the cloud can ated in an islanding mode with various kinds of components,
represent the real historical operating patterns. Optimization including WTs, PVs, and thermal power units (TUs), e.g.,
algorithms and tools can be adopted to derive a collection onsite diesel generator, and the lithium-ion battery-based ESS,
of optimized energy dispatch decisions over a multiperiod as illustrated in Fig. 2. In the microgrid, the EMS controls the
horizon with determined operational variables. The supervised power flows to meet certain operational objectives (e.g., cost
machine learning model is further developed to address the minimization) through managing the dispatchable and control-
complex mapping of the nonlinear input–output space between lable units [21]. The primary control functions are designed
the operational dynamic patterns and corresponding optimal to provide a timely and dynamic response to ensure voltage
dispatch decisions. The trained model at the cloud then can and frequency stability. As the islanding microgrid is gener-
be deployed in the field to make local energy management ally spanning over a small geographical area and the power
decisions in a real-time fashion. generators are installed locally to the power demand side, the
The proposed solution provides a novel perspective of power loss can be negligible.
integrating the training and inference process of the machine- In this work, economic energy dispatch in the islanding
learning-based energy management in the cloud-edge comput- microgrid aims to utilize the energy sources (including renew-
ing architecture. To the best of our knowledge, the data-driven able generation and onsite thermal units) to minimize the
energy dispatch solution that directly learns the optimized imbalances of generation and power demand while reducing
energy dispatch sequences and carries out the real-time infer- the operating costs and improving the lifetime protection of
ence decisions has not been reported yet. installed storage units. It can be formulated as a long-term
The main technical contributions made in this work can be horizon optimization problem with the predefined operational
summarized as follows. objectives subject to a set of constraints.
1) A learning-based data-driven decision-making frame- A. Objectives
work for energy dispatch of islanding microgrids is
proposed that effectively avoids the conventional effort The economic energy dispatch aims to minimize the long-
of accurate prediction of system variables and design term operational cost of a microgrid. The operation cost of
of sophisticated regulation strategies under operational a microgrid in islanding mode includes two components: the
uncertainties. cost of generation of TUs (e.g., diesel generator) Cd (·), and the
2) The proposed energy dispatch solution is developed degradation cost of ESS Cb (·) due to charging and discharging
based on a cloud-edge architecture that enables real-time actions. Thus, the total objective of economic energy dispatch
decision-making through local reasoning with signif- in each optimization cycle is
 
icantly reduced computing resource requirement and min Cd (pd (t)) + Cb (pb (t), SOH(t)) (1)
improved deployment flexibility. t∈T

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
13706 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 17, SEPTEMBER 1, 2021

where pd (t) and pb (t) are the dispatched power of the diesel where pmin
b and pmax
b are the lower and upper bounds of the
generator and the output power of the battery in time slot power output of the ESS, respectively.
t, respectively. When it is discharging, pb (t) is positive, and The dispatch constraint represents the dispatch ranges of
when it is charging, pb (t) is negative. SOH(t) is the state of decision variables. For the diesel generators, the generation
health of the battery in time slot t. Set T is the set of time limit constraint and minimum up/down time constraints are
slots for the long-term objective in global optimization. These given, respectively, as follows:
costs are, respectively, given as
pmin < pd (t) < pmax (9)
d d
Cd (pd (t)) = ad · (pd (t)) + bd · pd (t) + cd + · · ·
2
(rrt−1 − Ton )(I(pd (t − 1)) · I(pd (t))) ≥ 0
(10)
+ sd · I(pd (t)) · (1 − I(pd (t − 1))) (2) (zzt−1 − Toff )(I(pd (t)) · I(pd (t − 1))) ≥ 0
Cb (pb (t), SOH(t)) = ρb · (SOH(t) − SOH(t + 1)) (3) where pmin and pmax are the lower and upper bounds of the
d d
where ad , and cd are the generation cost coefficients of the output generation of the TUs, respectively. rrt is continuously
diesel generator [22]. sd is the unit startup cost coefficient. working time of unit at time t, zzt is continuously shutdown
I(·) is the indicator function to indicate if the diesel unit is time of unit at time t; Ton is the minimum up time of the unit,
turned on in each time slot. ρb is the degradation cost coeffi- Toff is the minimum down time of the unit.
cient for ESS. Furthermore, ESS life degradation due to battery
charging and discharging cycles can be calculated using (4) as III. P ROPOSED DATA - DRIVEN D ISPATCH S OLUTION
given in [23]: In this work, a data-driven machine-learning-based solu-
tion in a cloud-edge computing architecture is developed for
SOH(t + 1) = SOH(t) − h(t) · SOH(t) (4)
real-time economic dispatch in an islanding microgrid. The
where h(t) is the degradation factor, which is related to the main idea behind it is as follows. First, a large number of
change of battery state of charge (SOC), given as historical data measurements are adopted to simulate real-
 −1 word operation patterns. A collection of data samples are
h(t) = αh · (SOC(t))βh + ηh (5) selected to construct an offline optimization problem accord-
where αh , βh , and ηh are the degradation parameters related to ing to the system model presented in Section II. Through the
the battery specifications and the values can be obtained from optimization solver tool, the optimal decision sequence with
the empirical tests. SOC(t) is calculated by the charging or the long-term optimization horizon (e.g., 1 day) is obtained,
discharging power to label the learning targets of corresponding training sam-
  ples. The selected features containing operating information
 Eb , pb (t) ≤ 0
−cb pb (t) · t (e.g., power generations of renewable energy sources, SOC of
SOC(t) = (6)
−pb (t) · t db Eb , pb (t) > 0 ESS, and demand of loads) are used as the input variables.
where cb and db are the charging and discharging coefficients A random forest regression model is applied to fit the labeled
determined by battery charging/discharging efficiency charac- samples in a supervised learning fashion to realize the complex
teristics. t is the time interval. Eb is the battery capacity of mapping between operating pattern space and dispatch deci-
the ESS. In this work, the values of SOC and SOH are both sion space. Finally, the well-trained machine learning model
normalized to the interval [0, 1], and SOC is limited to the can address the real-time decision-making inference of the
range of [0.2, 0.8] in operation to avoid the deterioration effect economic dispatch in a new operating pattern. The proposed
due to deep charging and discharging. data-driven solution is designed to be self-adaptive through
carrying out rolling training to update the inference model
B. Constraints once new operational patterns are available. The generaliza-
tion ability and effectiveness of the model can be enhanced
The operation constraints of optimal energy management through such a rolling learning process at the cloud during
include power balance constraints, capacity constraints, and the long-term microgrid operation. The proposed machine-
dispatch constraints, as given in (7)–(10). The power balance learning-based solution in cloud-edge computing architecture
constraint requires that the generated power by all kinds of and the rolling learning process is illustrated in Fig. 3.
source generations minus the load demands equal the power
absorbed or released by the ESS, i.e.,
A. PSO-Based Optimization Process
pw (t) + pp (t) + pd (t) − pl (t) = −pb (t) (7) The data-driven optimal decision sequence labeling pro-
cess can be reduced to the typical optimal dispatch of a unit
where pw (t), pp (t), and pl (t) are the power of WTs, PVs, and
commitment (UC) problem under certain conditions. Many
loads, respectively, in time slot t which are nondispatchable
optimization solvers are available in the literature to effectively
variables. For the economic dispatch of islanding microgrid,
address the optimization problem, e.g., dynamic programming
the power dispatch of the diesel generator pd (t) is an active
(DP) [24], mixed-integer linear programming [25], Lagrangian
decision variable, and the power dispatch of the battery pb (t)
relaxation [26], and some other commercial solvers such
is passively determined by (7).
as CPLEX. In addition to them, the meta-heuristic algo-
The capacity constraint of the batteries are satisfied to
rithms are considered effective tools to solve such noncon-
b < pb (t) < pb
pmin max
(8) vex optimization problems (e.g., [27]). The particle swarm

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
DONG et al.: MACHINE-LEARNING-BASED REAL-TIME ECONOMIC DISPATCH IN ISLANDING MICROGRIDS 13707

Fig. 4. Renewable generation and load profiles of the training data set.

data measurements compared with other sophisticated deep-

learning models. Random forest uses decision trees, namely,
classification and regression tree (CART) [32], as its basic
learners [33] that contain multiple trees trained by the bag-
ging ensemble learning method. As a nonparametric model,
it does not restrict the tree structure nor set any prior density
assumptions.

IV. P ERFORMANCE A SSESSMENT AND N UMERICAL

R ESULT
A. Simulation Setup
The islanding microgrid illustrated in Fig. 2 is adopted for
the performance evaluation of the proposed data-driven eco-
nomic dispatch solution. The rated capacities of WTs, PVs,
and loads are 800 kW, 800 kW, and 1200 kW, respectively.
The capacity of ESS is 5000 kWh. The available range of SOC
is from 20% to 80%. Its charging and discharging efficiencies
Fig. 3. Proposed data-driven energy dispatch solution of islanding microgrid. are both 90%. In reality, the battery degradation parameters
(a) Machine learning framework. (b) Rolling learning process at the cloud.
can be hardly obtained, but can be estimated base on the
curve-fitting method through experimental tests [34]. In this
optimization (PSO) is a heuristic method based on popula- work, the specific values of the degradation parameters are
tion search with low implementation complexity [28], [29] set as: ρb = 100, αh = 0.001, βh = −2, and ηh = 0, as
and is widely used to address various engineering optimization suggested in [1]. For the diesel generator, the output power
d , pd ] = [800, 1200] kW. The cost coefficients
range is [pmin max
problems, including the UC problems (e.g., [30]). For the for- −4
mulated economic dispatch model given in Section II, the PSO are ad = 1.50 $/kW2 , bd = 0.025 $/kW, and cd = 0.04 $.
is adopted to solve the optimization problem with the power The unit startup cost is sd = 20 $. The minimum continuous
dispatch of the diesel generator as a decision variable. on Ton and off Toff time duration is 15 min. To reduce the
In PSO, each particle with a high dimension represents a complexity of optimization, the unit output is the same value
set of decision variables in the solution space. Individual par- in each time slot for 15 min.
ticles will update and maintain the historical optimal solution In this work, the profiles of renewable generations and
they have searched during iteration, i.e., personal optimal solu- power demand for simulations are adopted from a realistic
tions (Pbest), and collectively determine the global optimal islanding microgrid1 . The real-time data with the 5-min resolu-
solution (Gbest) through social behaviors with information tion are simulated for WTs, PVs, and loads. The collected data
sharing [31]. during three summer months from June 1, 2015 to August 31,
2015 (3 months) are constructed as the training samples for the
proposed machine-learning-based solution. Fig. 4 shows the
B. Random Forest-Based Regression Model
data distribution of power generation and load in the training
Random forest algorithm is considered efficient and set, covering a wide range of microgrid operational patterns.
lightweight to address various forms of regression problems The data from September 1, 2015, to September 15, 2015, are
with low computational complexity. Thus, it is suitable to be used as test days for performance evaluation and comparison
adopted at the edge computing units in the cloud-edge environ- with benchmark methods.
ment. Due to the fact that the random forest model is based on The proposed data-driven energy dispatch solution is imple-
bagging sampling and ensemble learning, it can make an excel- mented using Python 3.7 and simulated using a computing
lent trade-off between the stability (reduced variance) and the platform with a 3.00-GHz Intel Core i5-7400U CPU and an
accuracy (reduced bias). Therefore, it can well resist overfitting
and provide satisfactory performance in the case of insufficient 1 The data set of this work is available upon request.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
13708 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 17, SEPTEMBER 1, 2021

Fig. 6. Performance of proposed energy dispatch solution over a day.

renewable energy generations and load demands with a long-

term scope (i.e., one day). However, accurate prediction can be
hardly achieved in practice, especially at a high temporal reso-
lution (e.g., 5 min). In this simulation, the prediction errors of
power loads and generation of WTs are assumed to follow the
normal distributions [37], and the errors of the PV generation
are supposed to follow the beta distributions [38]. Here, two
different error level scenarios (i.e., small and large) are consid-
ered in the performance evaluation. The considered predicted
profiles of the generation of WTs and PVs, as well as loads
with small and large errors, are presented in Fig. 5 (a)–(c),
respectively.

B. Numerical Results and Discussion

Fig. 5. Real-profile and predictions with different error levels of (a) WTs; The proposed energy dispatch solution is evaluated through
(b) PVs; and (c) loads.
simulation experiments. Fig. 6 illustrated the performance of
the solution, including the power dispatch of the diesel gener-
ator, battery charging/discharging with SOC dynamics as well
8.00-GB RAM. The computational time of the overall model as the renewable generation and power demand over a test day
training process is about 1000 s (about 15 min), and the model (e.g., September 10). The results indicate that the renewable
execution time is in the order of milliseconds. For random for- power generation in the islanding microgrid has been fully uti-
est model, the number of trees in the forest and the minimum lized and the power generation and start-up times of the diesel
number of samples on leaf nodes are set to n estimators = 500 generators have been minimized to meet the power demand
and min samples leaf = 2, respectively. For PSO, the individ- requirement. Due to the limited range of the power generation
ual learning coefficient c1 = 2, the global learning coefficient of diesel generators, the mismatch between the power supply
c2 = 2, the numbers of particles npop = 30, the maximum and demand can be addressed by the charging/discharging of
number of iterations nite = 1000. the ESS. Meanwhile, the result shows that the deep charg-
The performance of the proposed solution is assessed ing and discharging of the battery are avoided (SOC varies
through a comparative study against the available solutions between 0.2 and 0.4) during the day to protect the battery
in the MPC framework (e.g., [35]). The benchmark meth- lifetime.
ods perform day-ahead scheduling using PSO based on the The results of energy dispatch performance over 5 consec-
prediction information that is formulated as a stochastic utive test days (from September 5 to September 9) of the
optimization problem with uncertainties. When the prediction proposed solution and the ideal dispatch (i.e., optimal dispatch
errors are reduced to 0, the ideal scheduling/dispatch can be without uncertainties) are shown in Fig. 7(a) and (b), respec-
obtained under certainty (although it is impossible to achieve tively. It clearly shows that the proposed solution can provide
in actual practice). In the real-time dispatch stage, to ensure the a satisfactory economic energy dispatch in a real-time man-
safety of the islanding microgrid system operation, all solu- ner over the test days. The SOC of the battery system can be
tions are required to apply the hard logic regulation, e.g., if well maintained within an appropriate range during the con-
SOC(t − 1) < 0.2, then pd (t) = pmax d ; if SOC(t − 1) > 0.8, secutive energy dispatch process. Also, for each test day, the
then pd (t) = 0. In this benchmark MPC-based solution to curves of the diesel generator output generation and the SOC
solve day-ahead scheduling for EMS, the available prediction variation of battery of the proposed solution exhibit the similar
approaches (e.g., [17] [36]) can be adopted to forecast the pattern to the ideal energy dispatch. This also indicates that

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
DONG et al.: MACHINE-LEARNING-BASED REAL-TIME ECONOMIC DISPATCH IN ISLANDING MICROGRIDS 13709

Fig. 8. Comparison of the operational and cumulative costs for 15 test days.

TABLE II
C OMPARISON OF THE AVERAGE C OSTS ($) OF D IFFERENT
S OLUTIONS OVER T EST DAYS

Fig. 7. Energy dispatch of diesel generator and battery SOC variation over
five consecutive days. (a) Proposed solution. (b) Ideal dispatch solution.
slightly lower cost. For the benchmark solutions, i.e., day-
ahead scheduling with two levels of prediction errors, the
the machine learning model obtained in this work has excel- efficient coordination between the battery-based ESS and dis-
lent generalization ability and robustness that can make the patchable diesel generators can be hardly achieved, and the
inference close to the optimal energy dispatch decision for the prediction errors can significantly degrade the energy dispatch
test days. performance, as confirmed by the results given in Fig. 8 and
To validate the machine-learning-based solution, the com- Table II.
parison of the operational costs in each day and their cumu- The average costs of different dispatch solutions for 15 test
lative curves obtained from the day-ahead scheduling with days are compared and presented in Table II. It shows that
different prediction errors, ideal dispatch without prediction the proposed solution is lower than the benchmark solutions
errors, and our proposed solution are shown in Fig. 8. It can be in each cost, and performance close to the ideal solution.
observed that all the evaluated solutions present a similar ten- The machine-learning-based solution can achieve the eco-
dency on operational and cumulative costs. It implies that for nomic improvement of 5.8% and 11.0% on average over the
the same operating patterns with the same state information, day-ahead scheduling with small and large prediction errors,
these solutions achieve similar basic goals for many situations. respectively. It is demonstrated that the proposed solution can
For instance, they are capable of absorbing renewable energy achieve efficient and real-time economic in the multienergy
generation and ensure the power supply reliability for local microgrid in the presence of different forms of uncertainties.
demand. Their similar tendencies indicate that these solutions
all attempt to realize the optimal strategy.
The results show that the proposed solution can achieve V. C ONCLUSION
a cost close to the ideal dispatch, and even slightly lower In this article, a learning-based framework in the cloud-edge
than that optimized without uncertainties in some days. This is computing environment is proposed to address the challenge
because, for nonconvex optimization problems, it is still diffi- of economic dispatch in islanding microgrids by considering
cult to achieve an absolute global optimal solution even under the stochastic nature of renewable generations and local loads.
certain conditions. The proposed machine-learning-based solu- The proposed data-driven solution leverages an external cloud
tion makes appropriate decisions in the real-time stage based service to label and generates the optimal decision samples
on the knowledge and experience learned in the histori- of a dispatch sequence through the heuristic algorithm. These
cal pattern, to realize other locally optimal solutions with a samples containing the different operating patterns are used in

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
13710 IEEE INTERNET OF THINGS JOURNAL, VOL. 8, NO. 17, SEPTEMBER 1, 2021

the supervised training process of a machine learning model. [16] M. S. Ibrahim, W. Dong, and Q. Yang, “Machine learning driven smart
The well-trained model can be deployed in the edge devices electric power systems: Current trends and new perspectives,” Appl.
Energy, vol. 272, Aug. 2020, Art. no. 115237.
to carry out the real-time dispatch task locally in the island- [17] W. Dong, Q. Yang, and X. Fang, “Multi-step ahead wind power genera-
ing microgrid. The numerical results obtained from extensive tion prediction based on hybrid machine learning techniques,” Energies,
simulation experiments confirm that the proposed solution out- vol. 11, no. 8, p. 1975, 2018.
performs the existing stochastic optimization methods with a [18] M. Rafiei, T. Niknam, J. Aghaei, M. Shafie-Khah, and J. P. S. Catalao,
“Probabilistic load forecasting using an improved wavelet neural
similar cost to the ideal dispatch scenario. network trained by generalized extreme learning machine,” IEEE Trans.
For future work, two research directions are considered Smart Grid, vol. 9, no. 6, pp. 6961–6971, Nov. 2018.
worth further research effort. The proposed learning-based [19] I. Dusparic, A. Taylor, A. Marinescu, F. Golpayegani, and S. Clarke,
“Residential demand response: Experimental evaluation and compari-
framework needs to be further assessed in the multi- son of self-organizing techniques,” Renew. Sustain. Energy Rev., vol. 80,
energy microgrids consisting of additional energy sources and pp. 1528–1536, Dec. 2017.
demands (e.g., cooling and heating) through a trail study. [20] Y. Zhao, T. Li, X. Zhang, and C. Zhang, “Artificial intelligence-based
Additional effort is needed to develop more advanced and fault detection and diagnosis methods for building energy systems:
Advantages, challenges and the future,” Renew. Sustain. Energy Rev.,
interpretable machine learning models, e.g., graph neural vol. 109, pp. 85–101, Jul. 2019.
network, through combining the domain knowledge. [21] W. Dong, Q. Yang, X. Fang, and W. Rua, “Adaptive optimal fuzzy
logic based energy management in multi-energy microgrid consider-
ing operational uncertainties,” Appl. Soft Comput., vol. 98, Jan. 2021,
Art. no. 106882.
R EFERENCES [22] T. A. Nguyen and M. L. Crow, “Stochastic optimization of renewable
based microgrid operation incorporating battery operating cost,” IEEE
[1] W. Liu, P. Zhuang, H. Liang, J. Peng, and Z. Huang, “Distributed Trans. Power Syst., vol. 31, no. 3, pp. 2289–2296, May 2016.
economic dispatch in microgrids based on cooperative reinforcement [23] A. Hoke, A. Brissette, K. Smith, A. Pratt, and D. Maksimovic,
learning,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 6, “Accounting for lithium-ion battery degradation in electric vehicle charg-
pp. 2192–2203, Jun. 2018. ing optimization,” IEEE J. Emerg. Sel. Topics Power Electron., vol. 2,
[2] L. Mariam, M. Basu, and M. F. Conlon, “Microgrid: Architecture, policy no. 3, pp. 691–700, Sep. 2014.
and future trends,” Renew. Sustain. Energy Rev., vol. 64, pp. 477–489, [24] A. Rong, H. Hakonen, and R. Lahdelma, “A dynamic regrouping based
Oct. 2016. sequential dynamic programming algorithm for unit commitment of
[3] G. Hug, S. Kar, and C. Wu, “Consensus+innovations approach for dis- combined heat and power systems,” Energy Convers. Manag., vol. 50,
tributed multiagent coordination in a microgrid,” IEEE Trans. Smart no. 4, pp. 1108–1115, 2009.
Grid, vol. 6, no. 4, pp. 1893–1903, Jul. 2015. [25] L. Yang, J. Jian, and Y. Wang, “Projected mixed integer programming
[4] Q. Jiang, M. Xue, and G. Geng, “Energy management of microgrid formulations for unit commitment problem,” Int. J. Elect. Power Energy
in grid-connected and stand-alone modes,” IEEE Trans. Power Syst., Syst., vol. 68, no. 68, pp. 195–202, 2015.
vol. 28, no. 3, pp. 3380–3389, Aug. 2013. [26] Q. Jiang, B. Zhou, and M. Zhang, “Parallel augment Lagrangian relax-
[5] X. Lei, T. Huang, Y. Yang, Y. Fang, and P. Wang, “A bi-layer multi- ation method for transient stability constrained unit commitment,” IEEE
time coordination method for optimal generation and reserve schedule Trans. Power Syst., vol. 28, no. 2, pp. 1140–1148, May 2013.
and dispatch of a grid-connected microgrid,” IEEE Access, vol. 7, [27] J. Zhao, S. Liu, M. Zhou, X. Guo, and L. Qi, “Modified cuckoo search
pp. 44010–44020, 2019. algorithm to solve economic power dispatch optimization problems,”
[6] Z. Zhang, J. Wang, T. Ding, and X. Wang, “A two-layer model for IEEE/CAA J. Automatica Sinca, vol. 5, no. 4, pp. 794–806, Jul. 2018.
microgrid real-time dispatch based on energy storage system charg- [28] J. G. Vlachogiannis and K. Y. Lee, “A comparative study on parti-
ing/discharging hidden costs,” IEEE Trans. Sustain. Energy, vol. 8, cle swarm optimization for optimal steady-state performance of power
no. 1, pp. 33–42, Jan. 2017. systems,” IEEE Trans. Power Syst., vol. 21, no. 4, pp. 1718–1728,
[7] J. Wu and X. Guan, “Coordinated multi-microgrids optimal control algo- Nov. 2006.
rithm for smart distribution management system,” IEEE Trans. Smart [29] W. Dong and M. Zhou, “A supervised learning and control method to
Grid, vol. 4, no. 4, pp. 2174–2181, Dec. 2013. improve particle swarm optimization algorithms,” IEEE Trans. Syst.,
[8] F. Li, J. Qin, and W. X. Zheng, “Distributed Q-learning-based Man, Cybern., Syst., vol. 47, no. 7, pp. 1135–1148, Jul. 2017.
online optimization algorithm for unit commitment and dispatch in
[30] W. Dong and Q. Yang, “Data-driven solution for optimal pumping units
smart grid,” IEEE Trans. Cybern., vol. 50, no. 9, pp. 4146–4156,
scheduling of smart water conservancy,” IEEE Internet Things J., vol. 7,
Sep. 2020.
no. 3, pp. 1919–1926, Mar. 2020.
[9] X. S. Zhang, Q. Li, T. Yu, and B. Yang, “Consensus transfer Q-learning
[31] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc.
for decentralized generation command dispatch based on virtual gen-
IEEE Int. Conf. Neural Netw., vol. 4. Perth, WA, Australia, 1995,
eration tribe,” IEEE Trans. Smart Grid, vol. 9, no. 3, pp. 2152–2165,
pp. 1942–1948.
May 2018.
[10] P. Dai, W. Yu, G. Wen, and S. Baldi, “Distributed reinforcement learning [32] L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen, Classification
algorithm for dynamic economic dispatch with unknown generation cost and Regression Trees. Hoboken, NJ USA: CRC Press, 1984.
functions,” IEEE Trans. Ind. Informat., vol. 16, no. 4, pp. 2258–2267, [33] L. Breiman, “Random forest,” Mach. Learn., vol. 45, pp. 5–32,
Apr. 2020. Jan. 2001.
[11] V. Mnih et al., “Human-level control through deep reinforcement [34] A. Millner, “Modeling lithium ion battery degradation in electric
learning,” Nature, vol. 518, pp. 529–533, Feb. 2015. vehicles,” in Proc. IEEE CITRES, Sep. 2010, pp. 349–356.
[12] C. Gong, F. Lin, X. Gong, and Y. Lu, “Intelligent cooperative edge [35] P. Li, D. Xu, Z. Zhou, W.-J. Lee, and B. Zhao, “Stochastic optimal oper-
computing in Internet of Things,” IEEE Internet Things J., vol. 7, no. 10, ation of microgrid based on chaotic binary particle swarm optimization,”
pp. 9372–9382, Oct. 2020. IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 66–73, Jan. 2016.
[13] F. Liang, W. Yu, X. Liu, D. Griffith, and N. Golmie, “Toward edge-based [36] M. Gao, J. Li, F. Hong, and D. Long, “Day-ahead power forecasting
deep learning in industrial Internet of Things,” IEEE Internet Things J., in a large-scale photovoltaic plant based on weather classification using
vol. 7, no. 5, pp. 4329–4341, May 2020. LSTM,” Energy, vol. 187, Nov. 2019, Art. no. 115838.
[14] S. Kulkarni et al., “Enabling a decentralized smart grid using [37] R. Doherty and M. O’Malley, “A newapproach to quantify reserve
autonomous edge control devices,” IEEE Internet Things J., vol. 6, demand in systems with significant installed wind capacity,” IEEE Trans.
no. 5, pp. 7406–7419, Oct. 2019. Power Syst., vol. 20, no. 2, pp. 587–595, May 2005.
[15] Y. Wu, “Cloud-edge orchestration for the Internet-of-Things: [38] S. H. Karaki, R. B. Chedid, and R. Ramadan, “Probabilistic performance
Architecture and AI-powered data processing,” IEEE Internet Things assessment of autonomous solar-wind energy conversion systems,” IEEE
J., early access, Aug. 7, 2020, doi: 10.1109/JIOT.2020.3014845. Trans. Energy Convers., vol. 14, no. 3, pp. 766–772, Sep. 1999.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.
DONG et al.: MACHINE-LEARNING-BASED REAL-TIME ECONOMIC DISPATCH IN ISLANDING MICROGRIDS 13711

Wei Dong (Student Member, IEEE) received the Wei Li (Senior Member, IEEE) received the Ph.D.
B.S. degree from the Huazhong University of degree from the School of Information Technologies,
Science and Technology, Wuhan, China, in 2017. The University of Sydney, Camperdown, NSW,
He is currently pursuing the Ph.D. degree with Australia, in 2012.
the College of Electrical Engineering, Zhejiang He is currently a Research Fellow with the Centre
University, Hangzhou, China. for Distributed and High Performance Computing,
His research interests include complex system School of Computer Science, The University of
modeling, machine learning and optimization, and Sydney. His research interests include Internet of
data-driven-based control. Things, edge computing, sustainable computing, task
scheduling, and energy efficiency and optimization.
Dr. Li is the recipient of four IEEE or ACM con-
ference best paper awards. He received the IEEE TCSC Award for Excellence
in Scalable Computing for Early Career Researchers in 2018 and the IEEE
Outstanding Leadership Award in 2018. He is a Senior Member of the IEEE
Computer Society and a member of ACM.

Albert Y. Zomaya (Fellow, IEEE) received the

Ph.D. degree from the Department of Automatic
Control and Systems Engineering, Sheffield
University, Sheffield, U.K., in 1990.
He is currently the Chair Professor of high
performance computing and networking with the
School of Information Technologies, The University
of Sydney, Camperdown, NSW, Australia. He is
Qiang Yang (Senior Member, IEEE) received also the Director of the Centre for Distributed
the Ph.D. degree in electronic engineering and and High Performance Computing which was
computer science from the Queen Mary University established in late 2009. He has published more
of London, London, U.K., in 2007. than 500 scientific papers and articles and is an author, coauthor, or editor
He has worked with the Department of Electrical for more than 20 books. His research interests are in the areas of parallel
and Electronic Engineering, Imperial College and distributed computing and complex systems.
London, London, U.K., from 2007 to 2010. He Prof. Zomaya is the recipient of the IEEE Technical Committee
visited the University of British Columbia and the on Parallel Processing Outstanding Service Award in 2011, the IEEE
University of Victoria Canada as a Visiting Scholar Technical Committee on Scalable Computing Medal for Excellence in
in 2015 and 2016. He is currently a Full Professor Scalable Computing in 2011, and the IEEE Computer Society Technical
with the College of Electrical Engineering, Zhejiang Achievement Award in 2014. He served as the Editor-in-Chief for the
University, Hangzhou, China. He has published more than 200 technical IEEE T RANSACTIONS ON C OMPUTERS from 2011 to 2014. He currently
papers, applied 60 national patents, coauthored two books, and edited two serves as an Editor-in-Chief for Scalable Computing (Springer) and an
books and several book chapters. His research interests over the years include Associate Editor for 22 leading journals, such as the ACM Computing
smart energy systems, large-scale complex network modeling, control and Surveys, IEEE T RANSACTIONS ON C OMPUTATIONAL S OCIAL S YSTEMS,
optimization, and learning-based optimization and control. IEEE T RANSACTIONS ON C LOUD C OMPUTING, and Journal of Parallel
Prof. Yang is a Fellow of the British Computer Society and a Senior and Distributed Computing. He is a Chartered Engineer and a Fellow of
Member of IET and China Computer Federation. AAAS and IET, U.K.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 17,2022 at 11:39:21 UTC from IEEE Xplore. Restrictions apply.