Electric Power Systems Research 225 (2023) 109782

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Distributed control strategy for secondary frequency regulation with EV

demand aggregation and delay compensation in AC unbalanced microgrid
Amrutha Raju Battula *, Sandeep Vuddanti
Department of Electrical Engineering, National Institute of Technology Andhra Pradesh, Tadeplalligudem, India

A R T I C L E I N F O A B S T R A C T

Keywords: Effective dispatch of ancillary services is a significant challenge for microgrids that incorporate renewable energy
Microgrid energy management (RE), distributed generators (DG), and electric vehicles (EVs), due to uncertainties and communication delays
Secondary frequency regulation during their operation. The secondary frequency regulation becomes a key task to restore the frequency to the set
Communication delay
point through a load control strategy. This paper presents a cost-effective two-stage distributed energy man­
Demand aggregation
Phase balancing
agement system (EMS) for microgrid operation to reduce reliance on battery storage systems and diesel gener­
ators, prolonging battery life and minimizing the use of expensive and polluting generators for secondary
frequency regulation. Communication delays are addressed using an adjusted consensus ADMM algorithm,
ensuring convergence. Additionally, a Stackelberg game framework is employed to design pricing at EV charging
stations and incentivize customers to participate in secondary frequency control. An adaptive phase switching
algorithm for phase switches is also proposed to ensure voltage phase balancing. Simulation results demonstrate
an 8.98% reduction in cost deviation across stages, with a $234.46 profit by the EV aggregator during the
intraday stage, considering secondary frequency regulation. The proposed approach offers a promising solution
for efficient microgrid operation and effective integration of renewable energy and electric vehicles.

and mitigating impacts of variable loads, intermittent renewable energy

sources, and changes in grid conditions. Load frequency control (LFC) is
1. Introducion classified into primary and secondary frequency control. Primary fre­
quency control is fast but has limited control capabilities, primarily for
In many regions around the world, the dependence on traditional immediate response to sudden frequency deviations [6]. However, the
energy sources and centralized power grids has resulted in issues such as primary control can lead to voltage and frequency deviations. To restore
power outages, voltage fluctuations, and environmental pollution. As a the voltage and frequency secondary control is applied secondary fre­
result of the growing energy demand with population and the urge for quency regulation (SFR) is slower but more sophisticated, accounting for
sustainable energy supply, microgrids have emerged as a promising system and load characteristics to regulate the system frequency and
resilient alternative in which energy is generated, stored, and consumed maintain power balance. In microgrid energy management, SFR is
locally, reducing reliance on fossil fuels and lowering the carbon foot­ crucial, especially with intermittent renewable energy sources, variable
print. Microgrids are integrated with renewable energy sources (RESs), loads, and distributed energy storage devices such as batteries and EVs
distributed generators (DGs), Electric vehicles (EVs) and energy storages [7]. These devices provide bidirectional energy flow capabilities, which
systems (ESS) that are optimized for energy utilization, to reduce energy can dynamically regulate the frequency and power balance [8–9].
costs, and enhance overall system efficiency [1]. Advanced technologies Secondary frequency control has been extensively studied with
such as smart grid communication and control systems, energy fore­ various control structures proposed, including both centralized [10–15]
casting, and optimization algorithms can be employed for effective en­ and distributed control approaches [16–21]. These methodologies have
ergy management in microgrids [2–4]. been explored to regulate and stabilize the frequency under different
Frequency is a critical ancillary parameter in microgrid operation, operating conditions. For instance, in [10], a state-machine-based co­
affecting stability, reliability, and performance of connected loads and ordinated control strategy has been developed to provide reliable fre­
distributed energy resources (DER) [5]. Maintaining a stable frequency quency ancillary service through battery storage. In [11], a self-adaptive
within an acceptable range is crucial for reliable microgrid operation

* Corresponding author.
E-mail address: [email protected] (A.R. Battula).

https://doi.org/10.1016/j.epsr.2023.109782
Received 19 May 2023; Received in revised form 20 July 2023; Accepted 21 August 2023
Available online 1 September 2023
0378-7796/© 2023 Elsevier B.V. All rights reserved.
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Nomenclature EESS Energy of ESS

ζESS
dis Discharge rate of ESS
ACADMM Adjusted Consensus ADMM ηESS / η ESS
Efficiency of EVA
c d
ADMM Alternating Direction Method of Multipliers
CESS Capacity of ESS
DER Distributed Energy Resources
iq Charge quotient of battery (Am)
DG Distributed generators
ηEVA EVA
G2V / ηV2G Efficiency of EVA
EVs Electric vehicles
EMS Energy Management System ηEVCS
st Efficiency of station ‘st’
ESS Energy Storages Systems NCS Number of charging stations
EVA EV aggregator Ωg Governor position
UT Grid Utility mfuel Fuel input rate
MPC Model Predictive Controller Kfuel Fuel flow coefficient
MCS Monte Carlo simulation Kg Governor gain coefficient
SFC Secondary frequency Regulation Ng Number of generators
SOC State of charge an,bn,cn Cost coefficients of nth DG
ns Total number of data points
Sϕi Apparent power for phase ɸ at bus i
xd(k) Desired state at time k
Viϕ Voltage of the phase ɸ at bus i ud(k) Desired input at time k
Iiϕ Voltage of the phase ɸ at bus i N Prediction horizon
Yijϕ Admittance matrix Q, R Weight matrices for the cost function
ζ Penalty coefficient of SFR signal
Δf˙ State of the frequency deviation
φ Frequency regulation signal
M Inertia constant
λ Lagrange multipliers
D Damping coefficients
ρ Penalty parameter.
ΔP Deviation in power
εp, εd Primal and dual residuals
NDT Number of dispatchable generators
N The total number of agents
m Agent
K Kalman gain matrix
τgm time constants of agent m
ϑ The step size for power modulation
ΔX Governor position
V imb Maximum voltage imbalance among the phases.
Rm Speed regulation coefficient of m.
pm Total power demand of charging stations.
Δu Deviation in control input
RESξ ΔPv Power required to balance the voltage unbalance
Pm Renewable source with uncertainty td Time delay
δ Logical operator (ON/OFF)

SFR strategy based on virtual synchronous generator with intermittent distribution systems [20]. The existing research provides on using
renewables is proposed. The droop control, which mimics the behavior Electric Vehicles (EVs) for frequency regulation services has shown
of a synchronous generator in the power system for the reliable power promising results with both centralized and decentralized control
supply and proper load sharing in inverters [12]. In [13], authors structure schemes [21].
developed an IoT-enabled framework for multi-energy microgrid market EVs are more cost-effective than Battery Energy Storage Systems
management, while their study [14] proposed a risk-based bidding (BESS) and other sources for providing ancillary services. Additionally,
system for interconnected microgrids. Additionally, a demand since EVs remain idle for almost 96% of their time, they offer a wider
response-based SFC has been proposed to calculate the magnitude of scope for efficiently utilizing them for frequency support [22]. Several
disturbances and simultaneously determine the area where disturbances studies have been carried out to regulate the storage concern of EVs with
with uncertainties, a control schemes using a Model Predictive load frequency control (LFC) models by controlling the number of
Controller (MPC) have been proposed [15]. conversions between charging and discharging states [23]. In [24], an
While microgrid control center (MGCC) with centralized control is a EV aggregator provided ancillary services while considering uncertainty
common technique that employs low bandwidth communication chan­ in prices and drivers’ patterns, capacity constraints, and state-of-charge
nels to control DERs in a microgrid. However, centralized control is not constraints. An optimized fuzzy controller for Vehicle-to-Grid (V2G)
the ideal choice when dealing with a large number of small-scale DERs. based secondary frequency control using EVs was proposed in [25] to
Additionally, communication link failures between the central control compensate for the variable imbalances between demand and genera­
and DER nodes can disrupt frequency regulation. In contrast, distributed tion. A proportional–integral-proportional-derivative (PI-PD) cascade
secondary control utilizes local control at each DER node, independent controller was developed in [26] to analyze the effect of storage units
of central or peer coordination. A Distributed Averaging Integration and EVs during SFC through mitigating active power imbalances. A
technique based MPC is proposed [16], to solve a local optimization distributed load frequency control (DLFC) strategy was designed in [27]
problem with the cost function penalizing the deviation of each agent. In to integrate the credibility-based resilient distributed frequency esti­
[17], a consensus algorithm was proposed for secondary control of mation scheme. These studies have demonstrated the potential of EVs to
voltage and frequency, based on the idea of adjusting the reference provide significant economic benefits. However, the existing literature
values when the consensus is realized [18]. To study the distributed reveals that the development of control systems for EV charging stations
control strategy, Arduino microcontrollers-based approach was pro­ to actively participate in providing ancillary services is still in its nascent
posed to stabilize the secondary frequency control scheme [19]. A stage. Achieving feasibility for a profit-efficient model in the microgrid
two-stage reactive power optimization method based on the alternating domain remains a critical area of exploration. This article is dedicated to
direction method of multipliers (ADMM) algorithm was proposed to addressing the research gap concerning the seamless integration of
achieve optimal reactive power dispatch in wind farm-integrated electric vehicles and intermittent renewable energy sources (RES) into

2
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Table 1 • A two-stage cost-effective approach to minimize penalties due to

Literature comparison on SFR in microgrids. deviation from prediction
Ref. SFR Distributed Pricing based Communication • Developed a parameter adjusting based distributed secondary fre­
through EV control schedule delay quency regulation approach, utilizing consensus-based ADMM.
[20] ✓ ✓ • Utilizing rolling horizon MPC-based approach to unlock the potential
[22] ✓ ✓ flexibility of various controllable devices with uncertainties.
[23] ✓ ✓ ✓ • An effective aggregation and coordination of electric vehicles (EVs)
[24] ✓ ✓ with charging stations, taking into account profit considerations, is
[26]
achieved through the application of the Stackelberg game theory.
✓ ✓ ✓
[27] ✓ ✓ ✓
Proposed ✓ ✓ ✓ ✓ • An adaptive phase switching algorithm is utilized to balance voltages
across different phases in the microgrid by introducing phase
switches.
microgrids. Specifically, our study focuses on the essential task of con­
ducting effective analysis of secondary frequency regulation, consid­ The structure of the article is organized as follows. Section 2 provides
ering the dynamic nature of load demand and RES fluctuations. In order a description of the microgrid architecture, which includes network
to gain a comprehensive understanding of the existing literature structure, components, and the objectives to be achieved. The proposed
alongside our proposed approach, a comparative analysis has been two-stage proposed energy management is discussed in Section 3. The
conducted. This analysis meticulously highlights the various secondary results are discussed in Section 4 while the conclusion and discussion are
frequency response control strategies and the gaps addressed in the field. presented in Section 5.
The findings of this comparative study are detailed in Table 1, offering
valuable insights into the current state of research in microgrid energy 2. Architecture of EMS structure
management.
To tackle these challenges, a novel and sustainable approach is 2.1. Network dynamics
proposed, leveraging EV aggregation for secondary frequency regulation
in microgrids. The approach aims to enhance frequency regulation This section presents a mathematical description of power network
performance, enable efficient EV demand aggregation, and maintain analysis and the imbalance voltage index in a distribution system, laying
balanced voltage levels in the microgrid by promoting effective coor­ the foundation for subsequent discussions in the manuscript. Addition­
dination and communication among distributed energy resources ally, the problem objective is discussed, along with dynamic modeling
(DERs). The contributions of this article to the advancement of micro­ for secondary frequency regulation.
grid energy management include optimization of system cost and per­
formance. The key contributions of this novel work are as follows: 2.1.1. Network
A connected graph (I, E) is used to represent the power network.

Fig. 1. Structure of the microgrid frequency regulation.

3
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Here, I = {1, 2, …, |I|} represents the set of buses in the network, and E ⊆ formulated as,
I × I represents the set of transmission lines connecting these buses. The ( )
∑
NDT
buses in the network are classified into two types: buses with inertia and ˙
Δf =
1
− DΔf + DT
ΔPm − ΔP Load
(3)
buses without inertia. Generators in power systems typically have M m=1
inertia due to the rotating masses, while other components such as loads [ DE ]
and capacitors do not have inertia. Therefore, it is reasonable to assume ΔPDT ESS
m = ΔPm ΔPm ΔPEVA
m (4)
that only buses with inertia have non-trivial generation dynamics. Here
The dynamic model of the dispatchable generators is,
B = {1, 2, …, |B|} is defined as the set of buses with inertia, and L = {|B|
+ 1, …, |I|} as the set of buses without inertia. The total number of buses 1 1
(5)
DT
ΔṖm = − ΔPDT
m + ΔXmDT
in the network is given by |B| + |L| = |I|, where |I| represents the total τgm τgm
number of buses in the network. The following assumptions are made for
the unbalanced network: DT
ΔẊ m = −
1
ΔXmDT −
1 1
Δf + g ΔuDT (6)
τgm Rm . τDT
m τm m
i Bus voltage magnitudes are assumed to be |Vj| = 1 p.u. (per unit) for
all buses j ∈ N. ΔuDT DT DT
m ≤ Δum ≤ Δum (7)
ii Reactive power flows are assumed to not affect bus voltage phase
g
angles and frequencies. where, ΔXm is the governor position; ΔuDT DT
m and Δum are the upper and
lower limits of control signals to power deviation in DTs [30].
The structure for the secondary frequency regulation of the micro­ Renewable energy sources (RESs), such as solar photovoltaic (PV)
grid energy management is shown in Fig. 1. and wind turbines, are commonly connected to the grid using inverters.
B. Power flow analysis However, due to their inherent characteristics, RESs do not actively
Conventional load flow techniques cannot converge on a microgrid contribute to frequency regulation. The output of PV generation is
distribution system due to its radial or ill-meshed structure, unbalanced influenced by temperature and irradiance, while wind generation is
load, distributed generators, and high resistance by reactance ratio dependent on wind speed. The power values for these sources can be
branches. In an unbalanced distribution system, power flow equations obtained from [22]. It is important to note that renewable generation
must take into account the unbalanced characteristics of the system, introduces uncertainty, denoted as RESξ, which is expressed in Eq. (8),
including the unbalanced voltages, currents, and power fluxes. Using the
PRESξ
= PPV ξ WTξ
m + Pm (8)
nodal or branch admittance matrix and taking into consideration un­
m

balanced load models, line and transformer models, and Kirchhoff’s where the forecasted output of solar photovoltaics (PVξ) and wind tur­
laws, these equations can be formulated. bine (WTξ) are compounded.
The power flow equations for an unbalanced three-phase distribution The dispatchable sources are modelled in the discrete state with their
system can be mathematically represented as shown in Eq. (1). These operation states with logic variable δ ϵ (0,1). The DTs can be modelled at
equations capture the overall power flow characteristics within the each time period ‘t’ as in (9).
system. For further detailed analysis and insights, do refer to the work of ∑
[28], which provides in-depth insight of the power flow modeling. PDT
t = PDT
m,t ⋅δ ∀t, ∀m, ∀δ ∈ (0, 1) (9)

( )∗ ∑ I (( )∗ ) The general dynamic model of the DTs (including DE, ESSs, EVs, etc.)
Siϕ = Viϕ Iiϕ + Vijϕ Yijϕ Vjϕ ∀ϕ, ∀i (1)
can be formulated as in [31].
i=1

The aforementioned equation models the power injection and power 1 1

(10)
DE
ΔṖm = − ΔPDE
m + ΔuDE
m
flow at each bus for individual phases, incorporating complex power, τDE
m τDE
m

voltage, and current variables, as well as admittance matrices that

describe the system’s impedance for each phase. To assess the voltage ΔuDE DE DE
m ≤ Δum ≤ Δum (11)
imbalance level and its permissible limitations, many definitions and
1 1
standards have been devised. In accordance with IEEE Std. 141–1993, (12)
ESS
ΔṖm = − ΔPESS
m + ΔuESS
m
τESS τESS
voltage phase unbalance (VU) in unbalanced distribution of 5% can m m

cause a 25% increase in motor temperature, necessitating the reduction

ΔuESS ESS ESS
m ≤ Δum ≤ Δum (13)
of voltage unbalances greater than 2% [29]. Voltage unbalance (VU) is
calculated as the ratio of the maximal deviation of phase voltage 1 1
(14)
EVA
magnitude from the average phase voltage magnitude to the average ΔṖm =− ΔPEVA
m + ΔuEVA
m
τEVA τEVA
phase voltage magnitude (2), where np is the number of phases (ϕ) m m

connected to bus i.
ΔuEVA
m ≤ ΔuEVA
m ≤ ΔuEVA
m (15)
( )
( ) 1 ∑ ϕ When intermittent renewable sources are used, energy storage de­
VU = max Viϕ − Vi × 100 (2)
np ϕ=a,b,c vices are implemented to guarantee the system’s smooth operation.
Depending on the materials, the system can either store energy as an
electrical charge or as a chemical electrolysis. Their bidirectional
2.2. Dynamic secondary frequency model operation assures optimal performance and a long service life when
utilized within their parameters. The charging and discharging dy­
This section discusses on the dynamic strategy utilized for microgrid namics of a battery are shown (16) while the State-of-charge (SOC)
components for regulating secondary frequency. The Swing equations which is used in (17) to determine the quantity of energy released during
can be utilized to depict the rate of change of frequency at the point of battery conversion.
common coupling (PCC) or the microgrid interface [30]. Controllable
components of a microgrid, such as diesel generators, batteries, and
inverters, work as the dispatchable generators (DT) to regulate the
secondary frequency and ensure stable operation. The resulting system
dynamics for the secondary frequency response of a microgrid can be

4
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

EESS (k + 1) = ζESS
dis E
ESS
(k)+ problem objective of the proposed MEM is formulated as shown in (24).
⎧
⎫
⎨ ηESS ⋅PESS (k), PESS (k) < 0
c (16) ∑Ng
⎪
/ min F(P) = fn (Pn ) ⎬
⎩ − PESS (k) ηESS (24)
ESS
d ,P (k) ≥ 0 n=1
⎪
⎭
2 2
⎧ ESS ∫ = an ⋅Pn + bn ⋅Pn + cn
⎪
⎪ ηc ESS
⎨ CESS iq , P (k) < 0
⎪
where, an be the cost of generating efficiently, bn be the fuel cost of
SOCESS (k + 1) = SOCESS (k) + (17)
⎪ ηESS ∫
⎪ generating unit n, cn be the maintenance cost of generating unit n, and Pn
⎪
⎩ − dESS iq , PESS (k) ≥ 0 be the power generated by each unit.
C
Ng
∑
where EESS and PESS are the real residual energy and power output of ESS s.t. Pn (k) = PLoad (k) (25)
during control period k. n=1

When electricity prices are high or demand is at its highest, the

Electric Vehicle Aggregator (EVA) participates in the energy exchange PDG DG DG
n ≤ Pn ≤ Pn (26)
while earning a profit [32]. The EV charging stations (EVCS) act as
batteries during their operations with the purpose to support the QDG DG DG
n ≤ Qn ≤ Qn (27)
aggregator’s objective. The energy at the aggregator and state of charge
at each station ‘st’ are calculate as described in the Eqs. (8), (9) and (10), SOC(k)ESS ≤ SOC(k)ESS ≤ SOC(k)ESS (28)
respectively.
⎧
⎪ SOC(k)EVCS ≤ SOC(k)EVCS ≤ SOC(k)EVCS (29)
⎪
⎪
⎪ ηEVA ⋅PEVA (k),
⎪
⎪
⎪ G2V
⎪
⎨ ΔPEVA (k) < 0 Viϕ ≤ Viϕ ≤ Viϕ (30)
EVA EVA EVA
E (k + 1) = ζdis E (k) + EVA
/ (18)
⎪
⎪ − P (k) ηEVA
V2G ,
⎪
⎪
⎪ VU ≤ VULimit (31)
⎪
⎪ ΔPEVA (k) ≥ 0
⎪
⎩
The objective of the proposed energy management system (EMS) for
the microgrid is a minimization function subject to both equality and
NCS (
∑ )
inequality constraints. The equality constraint is modeled to maintain
1
PEVA (k) = ηEVCS
st ⋅PEVCS
st (k) − P EVCS
st (k) (19) power balance within the microgrid (25), ensuring that the total power
ηst
EVCS

generated is equal to the total power consumed. On the other hand, the
st=1

Nst EV inequality constraints are imposed to limit the generation capacity, state
1 ∑ ( )
of charge, and voltages of the microgrid components, including PV,
SOCstEVCS (k + 1) = SOCeEV (k)
NstEV e=1 (20) wind, diesel generator, energy storage, and EV charging stations (26)-
+PEVCS
(
(k)⋅ ηch
/ d ) (30). By imposing these constraints, the system ensures that the micro­
EV − 1 ηEV
grid components operate within their safe operating limits.
st

Diesel generator (DE) participate in frequency regulation through

governor control which adjusts the fuel flow rate and hence the output
mechanical power. The equations for controlling the diesel generator 3.2. Proposed approach
output using the variables fuel input rate mfuel and governor position Ωg
is shown in (21)-(23). The proposed two-stage distributed energy management system for
( ) microgrids consists of a day-ahead and intraday EMS. It aims to optimize
ΔPDE = f mfuel , Ωg (21) energy management by minimizing costs while addressing challenges
such as secondary frequency regulation, uncertainty, phase balancing,
ṁfuel = Kfuel ⋅ufuel (22) and communication delays. The methodology utilizes electric vehicle
(EV) demand aggregation to reduce the demand on battery storage
Ω̇g = Kg ⋅ug (23) systems and diesel generators. This approach prolongs battery life and
reduces reliance on expensive and polluting generators, enhancing
where Kfuel and Kg is the fuel flow coefficient, and governor gain coef­ secondary frequency control.
ficient . ufuel and ug is the control signals for the fuel for the governor
valve position. 3.2.1. Stage 1: day ahead economic dispatch
This subsection presents the day-ahead energy scheduling approach,
3. Distribued secondary frequency regulation in microgrid which uses historical weather and load data to forecast 5-minute energy
framework scheduling intervals for PV, wind, diesel generator, battery storage, and
EV aggregator. The forecasting process takes into account the uncer­
3.1. Problem formulation tainty in the predicted data for wind, PV, and load through the appli­
cation of Monte Carlo simulation (MCS) method. By simulating multiple
In this section, problem objective is formulated for the proposed scenarios with random errors, day-ahead energy scheduling can identify
microgrid energy management. the optimal energy scheduling plan that minimizes the cost while ac­
After the primary frequency control, the secondary control mecha­ counting for the potential variability and uncertainty in the forecasted
nism is introduced to restore the frequency to its nominal level. One data.
effective approach is the integration of economic dispatch as secondary
control, which involves balancing generation and demand. The eco­ A Data Preparation and LSTM-based Forecasting
nomic dispatch technique aims to minimize the generation cost by
scheduling between the generators and loads. The cost function of each The data collection and pre-processing step is crucial in the day-
generator (fn) is represented as a convex quadratic equation and the total ahead energy scheduling process. Historical data for weather and load
cost of the individual generating units (F) is the system cost. The is collected and pre-processed to extract relevant features. The data is

5
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Algorithm 1 where ns is the total number of data points, yn is the actual value, and̂
yn
Day ahead forecasting through LSTM is the predicted value. By evaluating the RMSE and MAE after the Monte
Input: Historical data for PV, wind, and load, LSTM parameters Carlo simulation, the impact of the simulation on reducing uncertainty
Output: Forecasted values for PV, wind, and load. can be assessed.
Data Preparation:
1. Collect and pre-process historical data for PV, wind, and load.
3.2.2. Stage two: intraday economic dispatch
2. Split the data into training, validation, and test sets.
3. Normalize the data to have zero mean and unit variance. This section presents a distributed secondary frequency regulation
Model Training: approach that aims to minimize cost deviation from the day-ahead
1. Define the LSTM model with the following architecture: dispatch. The approach addresses communication delay through the
• Input layer with 3 neurons (corresponding to PV, wind, and load). use of a consensus algorithm and prioritizes the participation of the EV
• Two hidden layers with 50 neurons each and tanh activation function.
• Output layer with 1 neuron (corresponding to the forecasted value).
aggregator in secondary frequency regulation, while also achieving
2. Train the model on the training data using the Adam optimizer and mean voltage phase balancing. The objective of the intraday as shown in (35)
squared error (MSE) loss function. has three terms; first term is to minimize the overall cost, second term is
3. Validate the model on the validation set and adjust hyperparameters if to maintain the frequency at nominal, while the third term minimize the
necessary. Save the model weights.
day ahead power schedule cost deviation with weight prioritization µ1
Day-ahead Load and Renewable Forecasting:
1. Load the saved model weights. and µ2 at every step k.
2. Input the previous day’s load and renewable energy generation data.
∑
T
3. Generate day-ahead load and renewable energy generation forecasts using
min [F(P(k)) + μ1 (Δf (k)) + μ2 (Δp(k))] (35)
the trained LSTM model.
k=1

Ng
∑
then fed into a Long Short-Term Memory (LSTM) neural network to s.t . Δp(k) = |P(k) − P(k′)|, where P(k) = Pn (k) (36)
forecast future values of wind, solar, and load for each interval of 5 min n=1
for 24 h. LSTM is a type of recurrent neural network (RNN) that has been
widely used for time series forecasting due to its ability to capture long-
A Distributed Model prdictive control (DMPC)
term dependencies and handle sequential data effectively [33].
Compared to other traditional forecasting methods, LSTM has shown to
To solve the convex problem with secondary frequency regulation is
be more accurate and efficient in handling non-linear and
distributed manner, the optimization problem is formulated as a
non-stationary time series data. A pseudo algorithm is given below to
quadratic programming (QP) problem. The QP minimizes the deviation
illustrate the approach of using LSTM for forecasting. Algorithm 1
from the day-ahead dispatch while considering secondary frequency
regulation with necessary constraints. These constraints ensure that the
B Day-Ahead Energy Scheduling:
power output of the system is balanced and meets the required fre­
quency regulation. The resulting QP problem can be formulated as fol­
The generated scenarios are then used to evaluate the performance of
lows:
the day ahead economic dispatch system under different conditions,
which helps in making a more informed decision regarding energy N− 1 [
∑ ]
(37)
2 2
scheduling.The forecasted values are then used to formulate a day-ahead min J = ‖ x(k) − xd (k) ‖Q + ‖ u(k) − ud (k) ‖R + ζ‖ φ(k) ‖
energy scheduling plan for each interval of 15 min (k’). This plan takes
k=1

into account the various sources of energy such as PV, wind, diesel where, x(k): state at time kxd(k): desired state at time ku(k): input at time
generator, battery and EV aggregator consisting of EV charging stations kud(k): desired input at time kN: the prediction horizon
(EVCS). However, the forecasted data may contain uncertainties due to Q, R: weight matrices for the cost function
the unpredictable nature of weather and load. To handle the uncertainty ζ: penalty coefficient of SFR signal
in the forecasted data, monte carlo simulation (MCS) method is utilized. φ: frequency regulation signal at time k
MCS generates multiple scenarios of the uncertain parameters, such as
wind, PV, and load, based on the probability distribution of each s⋅t⋅xi (k + 1) = Axi (k) + Bui (k), ∀k (38)
parameter. The generated scenarios are then used to calculate the ex­ To solve the problem (37), the Alternating Direction Method of
pected cost function for the day ahead economic dispatch with an Multipliers (ADMM) algorithm is adopted to distribute the regulation
objective function of minimizing the total system cost (32), while among different dispatchable power sources or agentsi ∈ N . Each agent
adhering to the constraints (25)-(30). solves a local subproblem to minimize its own cost while ensuring that
( ( ′ N
∑T ∑ g
)) the states remain within the desired range at each step using the MPC
min E F fn (Pn (k)) ∀k′ (32) (38). The local solutions are coordinated through consensus update to
k′=1 n=1 achieve a global solution that satisfies the system-wide constraints [34].
The consensus ADMM algorithm solves in the form (39):
To evaluate the effectiveness of the MCS in reducing uncertainty, the
root mean square error (RMSE) and mean absolute error (MAE) are min J(xi ) + G(z) (39)
evaluated under a 95% confidence interval. These metrics are commonly
used to measure the accuracy of a forecasting model by comparing the where J, G are convex functions, variable z is introduced to enforce the
predicted values to the actual values. The equations for RMSE and MAE consensus constraint among the local primal variables x, and G(z) is a
are: penalty function that promotes consensus among the agents. The
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ consensus ADMM problem can be represented in the following
1 ∑ augmented Lagrangian form:
ns
RMSE = (yn − ̂ y n )2 (33)
ns i=1
ρ
Lρ (xi , z, λi ) = J(xi ) + G(z) + λTi (xi − z) + ‖ xi − z ‖2 (40)
2
1 ∑ ns
MAE = |yn − ̂y n | (34)
ns i=1 λ is the vector of Lagrange multipliers, and ρ is the penalty parameter
[35].
The consensus ADMM algorithm is an iterative process that updates

6
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Algorithm 2 B Delay Compensation Method (DCM)

Distributed Consensus-based ADMM.
Input: Initialize local decision variables, dual variables, and consensus variables for The communication delay (td) affects the information exchange be­
each agent. x0, z0, λ0, u0, ρ, N tween agents and can cause the following problems. The delay in
Output: u*, x*(k+1) exchanging information between agents can result in outdated and
for k= 1, 2, ..., T
inaccurate data, leading to suboptimal decision-making. Delayed in­
• minimization of the problem objective J (x,u) as in (37):
for m = 1, 2, ..., N formation exchange can impact the convergence and accuracy of the
• Each agent i updates its decision variable using the following rolling horizon MPC optimization process, leading to degraded MPC performance in terms of
optimization problem for N prediction. cost minimization and system stability.
for n =1,2.., N The ADMM communication among the agent is implemented in
• Update local decision variables of agents by solving the local optimization
asynchronous graph where each agent updates its solution at its own
problem as in (40): xi (k+1) = Axi(k)+Bui(k).
• Update the consensus variables using the ADMM rule (41): xi (k+1). time and global consensus is calculated at k. The proposed method ad­
• Agent i computes its consensus variable (42): zi (k+1), justs the ADMM step sizes and prediction horizon, by which the agents
• Agent i updates its dual variable (43): λi (k+1) can take into account the delay and improve the accuracy of their pre­
• Each agent i broadcasts its zi (k+1) to its neighbours and receives zj (k+1) from
dictions, leading to better overall convergence of the system. A Kalman
its neighbours.
end control gain K is incorporated as delay estimator and, the using the
• Check the convergence criteria through primal and dual feasibility as in (44), and estimated delay the prediction horizon, and the ADMM step sizes are
(45). adjusted [37]. The consensus ADMM algorithm step size is tuned using a
end process called ρ-tuning method. This involves adjusting the penalty
• Update the system state based on the control action.
parameter, which is denoted as ρ, to balance the primal and dual re­
• Update the demand forecast for the next time step.
• k=k+1 sidual errors.
end. The Kalman gain matrix need to define the state and measurement
equations. Let’s assume that the delay between the agents is the state
variable to be estimated, and the difference between the predicted and
the primal and dual variables, as well as the consensus variable, to received value of the control input is the measurement variable. The
converge towards a solution. The primal variable update is performed by state equation can be written as:
minimizing the augmented Lagrangian with respect to the local primal
variable xi, while holding the other variables fixed. The update rule for xi xi (k + 1) = Ad xi (k) + Bd ui (k) + wi (k) (46)
is given by (41).
⎧ ( ) ⎫ where, x(k) is the state vector at time step k, Ad is the discrete-time
system matrix, Bd is the discrete-time input matrix. The measurement
T
⎨ Ji (xi ) + λi (k) (xi − z(k)) ⎬
xi (k + 1) = argmin ρ (41) equation can be written as:
xi ⎩ + ‖ xi − z(k) ‖2 ⎭
2
yi (k) = Cd (ui (k) − ui (k − td )) + vi (k) (47)
The consensus variable update is performed by taking the average of
the local primal variables xi, and then adding the current estimate of the where y(k) is the measurement vector at time step k, Cd is the mea­
local dual variable [28]. The update rule for z is given by (42): surement matrix, u(k) is the predicted control input, u(k-td) is the
⎧ ( T ) ⎫ received control input delayed by d time steps, and v(k) is the mea­
⎨ G(z) + λi (k) (xi − z(k)) ⎬ surement noise. We can define the Kalman filter equations as follows
z(k + 1) = argmin
z ⎩ ρ
+ ‖ xi (k + 1) − z(k) ‖ ⎭
2 with Prediction step:
2
( ) x i (k + 1) = Ai xi (k) + Bi ui (k)
̂ (48)
1 ∑ N
1
= xi (k + 1) + λi (k) (42)
ρ
(49)
N i=1
P(k + 1) = Ai P(k)Ai T + H(k)

where, N is the total number of agents in the network. where ̂ x i (k +1) is the estimated state, P(k + 1) is the predicted error
The dual variable update is performed by adding the difference be­ covariance matrix, P(k) is the estimated error covariance matrix, and H
tween the local primal variable xi and the consensus variable z to the (k) is the process noise covariance matrix. The error covariance matrix
current estimate of the local dual variable λi (k). The update is as shown and estimated are updates as follows:
in (43).
( )− 1
K(k + 1) = P(k + 1)Cd T Cd P(k + 1)CdT + R (50)
λi (k + 1) = λi (k) + ρ(xi (k + 1) − z(k + 1)) (43)
Each agent i updates its Lagrange multiplier and computes its control x (k + 1) = Ad ̂x (k) + Bd u(k) + K(y(k + 1) − Cd u(k + 1))
̂ (51)
input. The convergence of the consensus ADMM algorithm can be
checked by monitoring the primal and dual residuals. The primal re­ P(k + 1) = [I − K(k + 1)Cd ]P(k + 1) (52)
sidual is defined as the norm of the difference between the local primal
variable and the consensus variable, while the dual residual is defined as where K(k + 1) is the Kalman gain matrix, R(k + 1) is the measurement
the norm of the difference between the current and previous Lagrangian noise covariance matrix, and C(k + 1) is the measurement matrix. We
multipliers [36]. The termination of the consensus ADMM algorithm are can use the Kalman filter as a delay estimator by assuming the
given by (44) and (45): communication delay td as the time difference between the predicted
state at time k + 1|k and the estimated state at time k + 1|k + 1:
εp ≤ ‖ xi (k + 1) − z(k + 1) ‖2 ∀i ∈ N (44)
td = (x(k + 1) − ̂
x (k + 1))/Ad (53)
2
εd ≤ ρ‖ λi (k + 1) − λi (k) ‖ ∀i ∈ N (45)
here A is the matrix that relates the state at time k + 1 to the state at time
The algorithm terminates when both the primal and dual residuals k.
are below a certain threshold, indicating that the primal and dual var­ Let N be the original prediction horizon of the ADMM algorithm, and
iables have converged to a consensus solution. The pseudo algorithm is td be the estimated communication delay. Then, the updated prediction
provided for implementation is given, Algorithm 2 horizon Nd, can be computed as: Nd, = N - td . The update of the penalty

7
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

parameter rho is critical for achieving convergence in the consensus communication delay td.
ADMM algorithm with communication delay td.
Proof: Combining Lemma 1 and Lemma 2, it can be shown that the
ρ(k + 1) = αρ(k) + β(td − ̂t d ) (54) consensus ADMM algorithm converges to the optimal solution with
communication delay td.
where ρ(k) is the penalty parameter at time step k, ̂t d is the estimated In summary, by incorporating the Kalman filter as a delay estimator
delay at time step k, and α and β are constants that control the rate of and updating the predictive horizon and ADMM step size using adaptive
change of the penalty parameter. penalty, the consensus ADMM MPC algorithm can effectively compen­
The value of α∈(0,1) to ensure that the penalty parameter does not sate for communication delay and converge to the optimal solution of
increase too rapidly. The value of β is typically set to a small positive the distributed MPC problem.
value to ensure that the penalty parameter is adjusted in response to
changes in the estimated delay. If the estimated delay is larger than the C EVCS pricing schedule: Stackleberg game
communication delay, the penalty parameter is increased to compensate
for the delay. Conversely, if the estimated delay is smaller than the The EVA aims to stabilize the secondary frequency at nominal by
communication delay, the penalty parameter is decreased to prevent effectively allocating power among the scheduled resources using
overcompensation distributed control. To achieve this, the EVA utilizes an optimal power
To prove the convergence of the algorithm, we need to show that the allocation strategy. Following each step of secondary control, the EVA
ADMM algorithm converges to a consensus solution under the updated calculates the optimal power allocation for each charging station
prediction horizon and adaptive penalty [38]. The convergence of the through the development of a pricing strategy, which is formulated as a
ADMM algorithm can be shown by using the following theorem: Stackelberg game. This approach ensures that the power allocation is
Theorem: Consider the distributed consensus ADMM algorithm with a optimized while considering the complex interactions between the
non-increasing sequence of penalty parameters ρ and the prediction charging stations and the grid.
horizon N updated based on the estimated delay td, with N = N+td. If the Stackelberg game is a type of game in which one player (the leader)
following conditions hold: makes a decision before the other player (the follower) and the follower
responds to the decision of the leader [39]. It is also known as a
leader-follower game or a hierarchical game.
a) The cost function is strongly convex and Lipschitz continuous in In this approach, the EVA acts as the leader, while each charging
xi, ui, and zi. The step size µ(k) is chosen such that 0 < µ(k) < 2/L, station acts as a follower. The leader’s objective is to maximize its profit
where L is the Lipschitz constant of the cost function. by setting the price of electricity to the charging stations, while each
b) The penalty parameter ρ(k) is chosen such that ρ(k) >L/ϵ where ϵ follower’s objective is to maximize its own profit by choosing the power
is the desired level of accuracy. demand that balances its revenue from EV charging against its costs of
c) The convergence of the algorithm can be proven through the purchasing and storing electricity The objective of the EVA is to find the
following lemmas: charging rates that maximizes the profit, which is the sum of the profit of
all the EV charging stations minus the cost of the secondary frequency
regulation.
Lemma 1. The estimation error e(k), converges to zero asymptotically as k
goes to infinity. ∑
Ncs
maxΠL = pm (cm − γ) (58)
Proof: The estimation error dynamics can be written as: pm
m=1

e(k + 1) = Ae(k) + w(k) (55) ∑

Ncs
s.t. pm ≤ PEVA (59)
The system is assumed to be stable, i.e., the eigenvalues of A are less m=1
than one in magnitude. Therefore, as k goes to infinity, e(k) converges to
zero. EVCS
SoCmin ≤ SoCmEVCS (k + 1) ≤ SoCmax
EVCS
(60)
Lemma 2. The ADMM algorithm converges to the optimal solution of the
pm (k) ≥ 0, ∀m (61)
optimization problem (37) with communication delay td.
Proof: The ADMM algorithm can be written as: where ΠL profit of EV Aggregator,pm power of each EV charging stations
m, cm cost of electricity at station and γsecondary frequency regulation.
xi (k + 1) = argminLi (xi , ui (k), zi (k), λi (k), ρ(k)) (56) The objective of the EV stations and customers is to find the optimal
x
charging rates that maximize their own profits. The followers’ problem
zi (k + 1) = argminLi (xi (k + 1), zi , λi (k), ρ(k)) (57) is subject to the constraint that the energy level of the storage should not
z
exceed the maximum or minimum limits, and the power demand should
where Li is the local objective function of agent i, z(k) is the consensus not exceed the maximum limit. The followers’ decision variables are the
variable, k is the penalty parameter. power demand of the station and the energy level of the storage, which
The convergence of the ADMM algorithm can be proven using the are determined by the leader’s decision variables, i.e., the secondary
following three conditions: frequency regulation price.
maxΠm = rm − pm (t)Tm (62)
1) The local objective functions are strongly convex. pm

2) The intersection of the constraint sets is non empty.

3) The penalty parameter satisfies a certain condition.
EVCS
s.t.SoCmin ≤ SoCmEVCS (t + 1) ≤ SoCmax
EVCS
(63)

Under these conditions, the ADMM algorithm converges to the where, Πm profit of follower j (EV charging station),rm utility function of
optimal solution with communication delay. follower m (based on the satisfaction level of EV customers),Tm charging
time of EV at station m.
Lemma 3. The consensus ADMM MPC algorithm with delay compensation The solution of the Stackelberg game gives the optimal power de­
converges to the optimal solution of the distributed MPC problem with mand and energy level for each EV charging station, as well as the

8
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Fig. 2. Proposed methodology flowchart.

optimal secondary frequency regulation price that maximizes the profit required to balance the voltage and determines the power modulation
of the EVA. The game is solved iteratively, with the leader solving the based on the voltage imbalance.
problem first and then the followers updating their decisions based on
V imb = max(|Va − Vb |, |Vb − Vc |, |Vc − Va |) (64)
the leader’s decision. The process is repeated until a Nash equilibrium is
reached, where no follower can improve their profit by changing their
ΔPv = ϑ ∗ V imb (65)
decision given the leader’s decision.
s.t., Pa = Pa − − ΔPv ∀a, b, c (66)
D Adaptive Phase Switching Algorithm (APSA)
where, ΔPv is the power modulation required to balance the voltage, ϑ is
Adaptive Phase Switching Algorithm (APSA) is a simple control
the step size for power modulation, V imb is the maximum voltage
strategy developed to ensure stable and efficient power supply to Elec­
imbalance among the phases.
tric Vehicle (EV) charging stations by balancing voltage levels in a
The step size for power modulation, δ, can be calculated as:
microgrid. The algorithm utilizes a feedback mechanism to continuously
monitor the voltage levels of each phase at the charging stations and ϑ = W/pm (k) (67)
trigger phase switching when the voltage imbalance exceeds the
threshold limit. where W is a constant factor that determines the sensitivity of the al­
The algorithm is initialized by setting the initial power demand for gorithm, and pm(k) is the total power demand from all charging stations.
each charging station, the target voltage for each phase, the threshold The algorithm continuously monitors voltage levels and adjusts the
voltage imbalance limit, and the step size for power modulation. If the power demand of charging stations until the voltage imbalance for all
voltage imbalance exceeds the threshold limit, the algorithm triggers phases is within the threshold limit. Once the imbalance is within the
phase switching by identifying the phase with the highest voltage limit, the algorithm terminates and maintains the power demand. If
imbalance. The algorithm then calculates the power modulation future voltage imbalance occurs, the algorithm can be re-triggered. This

9
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Table 2
Parameters of AC unbalanced microgrid system.
Equipment Description Capacity Effeciencies (%)

Solar panel DESERV Galactic Ultra 1000 kW 20.17

Wind turbine Vestas V20 1000 kW 44
Energy storage Exide Li- ion 1800 +90,− 95
kWh
EV Charging Type 2- Delta AC EV 11 kWh +95, − 95
point Charger
Diesel Mitsubhushi 500 kVA 1/4 load-11 gal/
hr
full load- 35 gal/
hr

Table 3
Parameter’s generator load data and locations.

Fig. 3. IEEE 123 node modified Unbalanced microgrid. Elements Characteristics Capacity Locations

Grid Infinite bus – 1

Load Commercial 3490+1170i kVA 1–34
Industrial 35–52
Residential 53- 123
EVPL load Amaron li-ion 600 kWh 16,52,64, 89,101

Table 4
System parameters and its values.
Parameter Value Parameter Value

R (Hz/kW) 3 pESS(¢/kWh) 13.5

D(kW/Hz) 0.02 µDE (s) 20
M(kW.s) 0.18 µESS (s) 20
aDE 0.30 µEVA(s) 20
bDE 5.5 µDE ($/kWh) 0.14
Fig. 4. Values of the Estimated and real-time data.
cDE 2.5 µESS ($/kWh) 0.13$/kWh
pRES (¢/kWh) 12.5 µEVA($/kWh) 0.11$/kWh

Fig. 5. EV demand or EV pricing.

Fig. 7. Day ahead power scheduling.

Fig. 6. (i) EV density profile of arrival and departure, (ii) Uncertainty error in
the forecasting.
Fig. 8. Intraday power scheduling.

Adaptive Phase Switching Algorithm (APSA) ensures stable and efficient

4. Results and discussion
power supply to EV charging stations.
This section presents the validation for the proposed algorithm of
two stage distributed secondary frequency control approach through the
modified IEEE 123 bus unbalanced distribution system. The results

10
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

Fig. 9. Secondary frequency regulation signals. Fig. 13. The results of the central MPC and the distributed ADMM approach in
terms of convergence and handling communication delays with pro­
posed approach.

Table 6
Comparison study on different characteristics.
C-MPC [44] D-MPC ADMM [45] D-MPC P-CADMM

SOCESS% (50, 55.7) (50,49.2) (50,53.3)

SOCEVA% (50, 31) (50,38) (50,34)
Iterations 377 182 86
Objective, Fx ($) 4438.33 4447.95 4434.23
Penalty ($) 435.45 233.17 162.94
Total cost ($) 4873.78 4681.12 4597.17

Fig. 10. Deviation of agents during the intraday dispatch.

Fig. 11. SOC of the battery and EVA during the intraday. Fig. 14. Pricing of the EVA at each time step.

Fig. 12. Objective comparison for: (i) with variable penalty parameter, (ii)
with variable delay.

Fig. 15. The generated pricing at each EVCS.

Table 5
Iteration with variable delay and penalty parameter.
ρ0.5 ρ1 ρ1.5 ρ*
td =0 524 385 168 86
td =10s 852 531 359 142
td =30s – – – 271
td = ∞ – – – –

highlight the simulation overview to the proposed scheme. In stage one,

using the forecasted data energy scheduling is performed for the day
ahead, often in 15-minute for a 24-hour consisting of 96 intervals. In the
later stage, an intraday response schedule is developed for economical
operation with ancillary services at an interval of 5 min. The analysis of
the proposed methodology is performed using an i5 processor with 8 GB
of RAM. The proposed method flowchart is shown in Fig. 2. Fig. 16. Scheduling of demand at the five EVCSs.

11
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

The EVA comprises five charging stations (EVCS) with a total of 75

charging points located throughout the distribution system, which
supports bidirectional Vehicle-to-Grid (V2G) services. The microgrid’s
structure is depicted in Fig. 3. The microgrid’s total load is composed of
residential, commercial, and industrial sectors. The residential load
consumption in the microgrid corresponds to approximately 1000
households, assuming that at least 30% of consumers own electric
vehicles.
Fig. 17. The unbalance voltage ratio for before and after APSA.
The forecasted data of the renewables and load demand is taken from
Indian Meteorological Department at Visakhapatnam location [41] and
central electricity authority of India which is modified accordingly to
test system [42]. The EV demand is modelled based on data acquired
from Quebec and modified to account for the maximum capacity. Hourly
electricity prices are obtained from an electricity pricing platform [43].
The Fig. 4, shows the forecasting and actual data of renewables (PV,
Wind) generation, load and EV demand for 24 h. The microgrid buying
and selling prices follow a time-of-use pricing scheme asshown in Fig. 5.
The uncertainties associated with forecasted data and EV arrival and
Fig. 18. Phase power shifting at each step. departure at charging stations are depicted in Fig. 6. Microgrid param­
eter specifications are listed in Table 2, while. Table 3. presents gener­
4.1. Test system ation and load data along with their corresponding locations. Table 4.
provides different values of the system parameters.
The proposed energy management control is validated through a The forecasting error is calculated as the sum of the differences be­
modified IEEE 123-node unbalanced microgrid, which serves as a tween the actual values and the predictions. Before being utilized for
representative example of a typical urban distribution system [40]. The simulation validation, the prediction error is adjusted by the open real
microgrid’s energy generation units include renewable sources such as measurement database. The horizon for scheduling is generally set for a
wind and solar PV, a fuel-based distributed generation unit powered by 24-hour.
diesel, an EV Aggregator (EVA), and a battery storage system (BESS).

Fig. 19. Voltage profile at different instance of time: (i) 04:00hr, (ii) 10:00hr, (iii) 1600hr, (iv) 2200hr.

12
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

4.2. Day-ahead economic dispatch unbounded delay (td=30 s), and failure scenarios. The proposed algo­
rithm has demonstrated its efficiency in converging with all the
This stage presents a proposed day-ahead forecasting and scheduling considered delay scenarios, with an improved convergence rate, and the
method with the objective for cost-effectiveness. Firstly, a three-year failure scenario shows unconverged values.
dataset that includes photovoltaic (PV), wind, and energy demand The results of the delay communication and penalty parameter
data is utilized to forecast through Long Short-Term Memory (LSTM) analysis indicate that the proposed method achieves optimal conver­
model. The LSTM model is fine-tuned with specific hyperparameters, gence faster. Table 5. shows the number of iterations required to reach
which include two layers, 64 hidden neurons, and a batch size of 128 for convergence for each parameter under different delay times and penalty
forecasting the energy demand and renewable energy generation for the parameter.
next 24 h at 15-minute intervals. The analysis reveals that higher penalty parameter values lead to a
Next, a Monte Carlo simulation is adopted to generate 10,000 sam­ decrease in the number of iterations, indicating faster convergence in
ples, which effectively reduce the uncertainty in the forecasting error. the optimization process. However, for the td = 30 s and td = ∞ sce­
method utilizes a Mixed-Integer Linear Programming (MILP) optimiza­ narios, indicating a failure to reach convergence. These observations
tion algorithm to minimize costs while balancing the load across the highlight the sensitivity of the iteration process to the choice of delay
available resources. The MILP algorithm is formulated to assign weights and penalty parameters, emphasizing their impact on convergence
to each resource and optimize the energy dispatch for the next 24 h. The behavior.
objective function is defined to minimize the total operating cost, which To evaluate the proposed method against existing control tech­
includes the costs of generation, as well battery storage which are sub­ niques, we implemented a central MPC-based approach and a distrib­
ject to constraints such as the maximum and minimum power output of uted ADMM approach with MPC, both having similar parameters as the
each unit and demand-supply balance. proposed method, on the tessystem and compared the results. They are
The proposed day-ahead forecasting and scheduling method is compared under no delay a bounded delay of 10 s.The findings showed
optimized with a 95% confidence level constraint, resulting in a cost that the proposed method outperformed both the central MPC [44] and
minimization objective of 4876.68$, as shown in Fig. 7. The effective­ the distributed ADMM [45] approach in terms of convergence and
ness of this approach is demonstrated through a root mean square error handling communication delays, as illustrated in Fig. 13. The results
(RMSE) of 0.982% and mean absolute error (MAE) of 0.453%, with a show, during delay the centralized MPC doesn’t converge while the
standard deviation of 12.13%. ADMM converges at suboptimal solution. Furthermore, Table 6. pro­
The battery management system plays a crucial role in balancing the vides a comparison of different characteristics among the three ap­
load and optimizing the State of Charge (SOC) to ensure efficient energy proaches considered during no communication delay of the test system.
utilization. The SOC is initially set at 50% and reaches a stable state of The above table show that the D-MPC P-CADMM approach shows
47.8% through the usage of the battery. The results indicate the effec­ significant advantages over the C-MPC and D-MPC ADMM approaches in
tiveness of the Monte Carlo method in reducing the uncertainty in terms of performance metrics. Compared to C-MPC, it achieves a 0.09%
forecasting errors and suggest that the day ahead economic dispatch reduction in the objective function value and a 62.64% decrease in the
approach offers a promising solution for cost-effectiveness. penalty. Compared to D-MPC ADMM, it achieves a 0.06% reduction in
the objective function value and a 30.20% decrease in the penalty. These
4.3. Intra day economic dispatch results indicate that the D-MPC P-CADMM approach performs better in
terms of cost optimization and adherence to constraints. Additionally, it
In the intraday scheduling, the challenge arises due to the unpre­ demonstrates faster convergence with a 53.85% reduction in the number
dictable nature of intermittent sources and consumption deviating from of iterations compared to C-MPC and a 52.75% reduction compared to D-
the estimated cost. To address this issue, a control scheme has been MPC ADMM.
developed to adopt the estimated cost using the stage 2 objective pro­ By scheduling the available resources optimally, the agent EVA can
posed in (35) at an interval of 5 min. The control approach operates as distribute itself among the charging stations to achieve optimal charging
discussed in Section 3.2-B and solves the quadratic convex problem with and discharging at each step. The EVA model employs a Stackerberg
penalties for deviations through a distributed approach using ADMM. game theory to determine optimal pricing for each charging station,
Each agent in the system solves its problem (38) using the rolling ho­ ensuring efficient allocation of resources [46]. Acting as a leader, the
rizon MPC and shares the information about the expected outcome with EVA sets pricing for profit while supporting the SFR, as outlined in
neighbouring agents at each step. The proposed intraday scheduling section 3.2.2.C. Fig. 14. depicts the pricing strategy designed for optimal
method is optimized, resulting in a cost minimization objective of profit. Pricing is adjusted during frequency fluctuations, with increases
4434.23 $. The scheduling of the resources for the intraday is shown in during negative frequency and decreases during positive frequency
Fig. 8. signals.
To maintain nominal frequency, frequency signals are generated at As followers, charging stations prepare pricing to support the cu­
each step according to Eq. (3). Additionally, the EVA agent is given a mulative profit of the EVA, with Fig. 15. showing the pricing generated
higher priority among the DT agents by assigning it a weight of (7) to for each station. EVs accept these pricing conditions, as it improves their
compensate for any frequency response deviation as shown in Fig. 9. The chances of receiving incentives during the operation.
agents aim to minimize the estimated cost and dispatch optimally to Load scheduling is performed at each station according to the
match the deviation of real-time data at each step k, while the DTs generated pricing. Fig. 16. illustrates the charging and discharging
optimize their dispatch to minimize penalties from day-ahead de­ patterns of EV stations over a 24-hour period, with different EVAs being
viations. The control states of the deviated agents are depicted in Fig. 10. dispatched at varying times to balance the load. The proposed method
During the intraday stage result in the SOCEVA (50, 34) and SOCESS (50, yields a profit of $234.46 for the intraday stage with balanced charging
53) are illustrated in Fig. 11. and discharging across all charging stations.
To validate the robustness of the distributed algorithm, two key Each charging station in the microgrid is equipped with a three-
parameters are considered - the penalty parameter, ρ, and communica­ phase transformer, but due to varying charging and discharging re­
tion delay, td. The penalty parameter plays a crucial role in the quirements of the EVs, the power demand can become unbalanced
convergence of the solution, and different values of ρ are analyzed to across the phases. To mitigate this issue, phase switches are installed at
understand the convergence behavior, as illustrated in Fig. 12. In each EV station, which can shift the power between phases to minimize
addition, the communication delay of the distributed algorithm is the unbalance voltage. In this work, a simple yet effective algorithm is
analyzed, considering no delay (td =0 s), bounded delay (td=10 s), developed in Section 3.2.2. D to maintain the phase balance of the

13
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

system. The proposed control strategy is applied at each step k of the References
intraday to maintain the unbalance. The results show a significant
reduction in the unbalance voltage ratio, as shown in Fig. 17. [1] H. Shayeghi, A. Rahnama, H.H. Alhelou, Frequency control of fully-renewable
interconnected microgrid using fuzzy cascade controller with demand response
The changes in phase power at each step are illustrated by the R-a, Y- program considering, Energy Reports 7 (2021) 6077–6094, https://doi.org/
b, and B-c phases, as shown in Fig. 18. These phase shifts occur through 10.1016/j.egyr.2021.09.027.
the phase switches present at the EV station during each time step. The [2] M.R. Jokar, S. Shahmoradi, A.H. Mohammed, L.K. Foong, B.N. Le, S. Pirouzi,
Stationary and mobile storages-based renewable off-grid system planning
impact of phase shifting on voltage phases can be analyzed by considering storage degradation cost based on information-gap decision theory
comparing the voltage profiles before and after the algorithm is applied optimization, J. Energy Storage 58 (2023), 106389.
at specific time intervals, as depicted in Fig. 19. This comparison helps to [3] E. Akbari, S.F.M. Shabestari, S. Pirouzi, M. Jadidoleslam, Network flexibility
regulation by renewable energy hubs using flexibility pricing-based energy
demonstrate the effectiveness of the algorithm in balancing the voltage management, Renew. Energy 206 (2023) 295–308.
across all phases. [4] S.A. Mansouri, E. Nematbakhsh, A.R. Jordehi, M. Tostado-Véliz, F. Jurado,
From the above analysis, the proposed approach utilizes EV demand Z. Leonowicz, Improved double-surface sliding mode observer for flux and speed
estimation of induction motors, IET Electric Power App. 14 (2020) 1002–1010,
aggregation to support for SFR through Stackelberg game framework,
https://doi.org/10.1049/iet-epa.2019.0826.
and adaptive phase switching to minimize the unbalance voltages with [5] S. Mansouri, E. Nematbakhsh, A.R. Jordehi, M. Marzband, M. Tostado-Véliz,
effectively integration of electric vehicles (EVs) and renewable energy F. Jurado, An interval-based nested optimization framework for deriving flexibility
sources (RES) into microgrids. Moreover, ACADMM enhances commu­ from smart buildings and electric vehicle fleets in the TSO-DSO coordination, Appl.
Energy 341 (2023), 121062, https://doi.org/10.1016/j.apenergy.2023.121062.
nication resilience, reducing the risk of failure during communication [6] G. Piltan, S. Pirouzi, A. Azarhooshang, A.R. Jordehi, A. Paeizi, M. Ghadamyari,
delays. This approach addresses the challenges of distributed energy Storage-integrated virtual power plants for resiliency enhancement of smart
management. distribution systems, J. Energy Storage 55 (B) (2022), 105563 pt.
[7] A. Pirouzi, J. Aghaei, S. Pirouzi, V. Vahidinasab, A.R. Jordehi, Exploring potential
storage-based flexibility gains of electric vehicles in smart distribution grids,
5. Conlusion J. Energy Storage 52 (C) (2022) pt105056.
[8] A. Dini, A. Hassankashi, S. Pirouzi, M. Lehtonen, B. Arandian, A.A. Baziar,
A flexible-reliable operation optimization model of the networked energy hubs
In conclusion, the proposed approach outperforms other methods with distributed generations, energy storage systems and demand response, Energy
regarding the objective function value and penalty, exhibiting a 0.8% 239 (A) (2022), 121923 pt.
improvement over C-MPC and a 0.4% improvement over D-MPC ADMM. [9] M. Zadehbagheri, M.J. Kiani, S. Pirouzi, M. Movahedpour, S. Mohammadi, The
impact of sustainable energy technologies and demand response programs on the
It demonstrates substantial penalty reductions of 62.7% and 30.2% hub’s planning by the practical consideration of tidal turbines as a novel option,
respectively, reinforcing the importance of its adoption for effective Energy Reports 9 (2023) 5473–5490.
microgrid energy management. Leveraging collective knowledge from [10] P. Ojaghi, M. Rahmani, LMI-based robust predictive load frequency control for
power systems with communication delays, IEEE Trans. Power Syst. 32 (5) (2017)
system agents, the approach robustly handles uncertainties and varia­
4091–4100, https://doi.org/10.1109/TPWRS.2017.2654453.
tions. Furthermore, it effectively addresses communication delays, [11] Jin Zhang, Zhenpo Wang, Eric J. Miller, Dingsong Cui, Peng Liu, Zhaosheng Zhang,
enhancing resilience and reducing failure risks. The EV aggregator Charging demand prediction in Beijing based on real-world electric vehicle data,
model, utilizing Stackelberg game theory for optimal pricing strategies, J. Energy Storage 57 (2023), 106294, https://doi.org/10.1016/j.est.2022.106294.
ISSN 2352-152X.
ensures efficient resource allocation, contributing to an intraday cu­ [12] Md Rahman, Eskinder Gemechu, Abayomi Oni, Amit Kumar, The development of a
mulative profit of $234.46. The incorporation of phase switches ad­ techno-economic model for assessment of cost of energy storage for vehicle-to-grid
dresses voltage imbalances at charging stations, enhancing system applications in a cold climate, Energy 262 (2022), 125398, https://doi.org/
10.1016/j.energy.2022.125398.
performance. The approach fosters efficient coordination, adapts to [13] Z. Li, K.W. Chan, J. Hu, J.M. Guerrero, Adaptive droop control using adaptive
demand variations, and holds potential for advancing smart grid tech­ virtual impedance for microgrids with variable PV outputs and load demands, IEEE
nology. Future research can focus on enhancing scalability and flexi­ Trans. Ind. Electron. 68 (10) (2021) 9630–9640, https://doi.org/10.1109/
TIE.2020.3022524.
bility, integrating advanced machine learning techniques for further [14] S. Mansouri, A. Rezaee Jordehi, M. Marzband, M. Tostado-Véliz, F. Jurado,
optimization. By leveraging advancements in renewable energy tech­ J. Aguado, An IoT-enabled hierarchical decentralized framework for multi-energy
nologies, this approach contributes to the development of intelligent and microgrids market management in the presence of smart prosumers using a deep
learning-based forecaster, Appl Energy 333 (2023), 120560.
resilient energy systems. Overall, adopting distributed optimization [15] S.A. Mansouri, E. Nematbakhsh, A.R. Jordehi, M. Tostado-Véliz, F. Jurado,
with ADMM and consensus improves microgrid performance and sta­ Z. Leonowicz, A risk-based Bi-level bidding system to manage day-ahead electricity
bility while addressing energy management challenges, making it a market and scheduling of interconnected microgrids in the presence of smart
homes, in: 2022 IEEE International Conference on Environment and Electrical
promising research area for the advancement of intelligent and resilient
Engineering, Prague, Czech Republic, IEEE Industrial and Commercial Power
energy systems. Systems Europe, 2022, pp. 1–6.
[16] C. Ahumada, R. Cárdenas, D. Sáez, J.M. Guerrero, Secondary control strategies for
CRediT authorship contribution statement frequency restoration in islanded microgrids with consideration of communication
delays, IEEE Trans. Smart Grid 7 (3) (2016) 1430–1441, https://doi.org/10.1109/
TSG.2015.2461190.
Amrutha Raju Battula: Investigation, Conceptualization, Method­ [17] F. Mehmood, B. Khan, S. Ali, J Rossiter, Distributed model predictive based
ology, Software, Writing – original draft. Sandeep Vuddanti: Method­ secondary control for economic production and frequency regulation of microgrid,
IET Control Theory and Applications 13 (2019), https://doi.org/10.1049/iet-
ology, Software, Writing – review & editing. cta.2018.6226.
[18] S. Hu, L. He, H. Zhao, H. Liu, X. Liu, J. Qiu, Distributed secondary control of
Declaration of Competing Interest microgrids with unknown disturbances and non-linear dynamics, Front. Energy
Res. 10 (2023).
[19] S. Nigam, O. Ajala, M. Zholbaryssov, A.D. Domínguez-García, P.W. Sauer,
The authors declare that they have no known competing financial Distributed secondary frequency control in AC microgrids with lossy electrical
interests or personal relationships that could have appeared to influence networks, in: 2020 IEEE Power & Energy Society General Meeting (PESGM); 2020
Aug 2-6; Montreal, QC, Canada, Piscataway (NJ), IEEE, 2020, pp. 1–5.
the work reported in this paper. [20] Ronghui Zhang, Na Wu, Zihan Wang, Kening Li, Zhumei Song, Zhenting Chang,
Xia Chen, Fan Yu, Constrained hybrid optimal model predictive control for
Data availability intelligent electric vehicle adaptive cruise using energy storage management
strategy, J. Energy Storage 65 (2023), 107383, https://doi.org/10.1016/j.est.
2023.107383. ISSN 2352-152X.
Data will be made available on request. [21] S. Huang, P. Li, Q. Wu, F. Li, F. Rong, ADMM-based distributed optimal reactive
power control for loss minimization of DFIG-based wind farms, Int. J. Electric.
Power & Energy Syst. 118 (2020), 105827, https://doi.org/10.1016/j.
ijepes.2020.105827.
[22] K. Kaur, R. Rana, N. Kumar, M. Singh, S. Mishra, A colored petri net based
frequency support scheme using fleet of electric vehicles in smart grid

14
A.R. Battula and S. Vuddanti Electric Power Systems Research 225 (2023) 109782

environment, IEEE Trans. Power Syst. 31 (6) (2016) 4638–4649, https://doi.org/ [35] J. Zhang, Asynchronous decentralized consensus ADMM for distributed machine
10.1109/TPWRS.2016.2518743. learning, in: 2019 International Conference on High Performance Big Data and
[23] Anatole Desreveaux, Alain Bouscayrol, Rochdi Trigui, E. Hittinger, Élodie Castex, Intelligent Systems (HPBD&IS), Shenzhen, China, 2019, pp. 22–28.
G.M. Sirbu, Accurate energy consumption for comparison of climate change impact [36] S. Wang, A. Dekorsy, Distributed consensus-based extended Kalman filtering: a
of thermal and electric vehicles, Energy 2023 (268) (2023), 126637, https://doi. bayesian perspective, in: 2019 27th European Signal Processing Conference
org/10.1016/j.energy.2023. 126637. (EUSIPCO), A Coruna, Spain, 2019, pp. 1–5, https://doi.org/10.23919/
[24] Q. Zhang, Y. Li, C. Li, C Li, Grid frequency regulation strategy considering EUSIPCO.2019.8902553.
individual driving demand of electric vehicle, Electric Power Syst. Res. 163 (A) [37] C. Chen, S. Duan, Microgrid economic operation considering plug-in hybrid electric
(2018) 38–48. Part. vehicles integration, J. Mod. Power Syst. Clean. Energy 3 (2) (2015) 221–231,
[25] Torre S de la, J.A. Aguado, E. Sauma, Optimal scheduling of ancillary services https://doi.org/10.1007/s40565-015-0116.
provided by an electric vehicle aggregator, Energy 265 (2023), 126147, https:// [38] Mingze Zhang, Weidong Li, Samson Shenglong Yu, Kerui Wen, S.M Muyeen, Day-
doi.org/10.1016/j.energy.2022.126147. ahead optimization dispatch strategy for large-scale battery energy storage
[26] S. Falahati, S.A. Taher, M. Shahidehpour, Grid secondary frequency control by considering multiple regulation and prediction failures, Energy 270 (2023),
optimized fuzzy control of electric vehicles, IEEE Trans. Smart Grid 9 (6) (2018) 126945. ISSN 0360-5442.
5613–5621, https://doi.org/10.1109/TSG.2017.2692265. [39] P. Shinde, K. Shanti Swarup, Stackelberg game-based demand response in multiple
[27] S. Sun, Q. Yang, W. Yan, Optimal temporal-spatial PEV charging scheduling in utility environments for electric vehicle charging, IET Electr. Syst. Transp. 8 (3)
active power distribution networks, Prot. Control Mod. Power Syst. 2 (2017) 34, (2018) 167–174, https://doi.org/10.1049/iet-est.2017.0046.
https://doi.org/10.1186/s41601-017-0065-x. [40] S.A. Papathanassiou, N. Hatziargyriou, K. Strunz, A benchmark low voltage
[28] Zhijian Hu, Shichao Liu, Ligang Wu, Credibility-based distributed frequency microgrid network, in: Proceedings of the CIGRE Symposium: Power Systems with
estimation for plug-in electric vehicles participating in load frequency control, Int. Dispersed Generation. Athens, 2005.
J. Electric. Power & Energy Syst. 130 (2021), 106997. ISSN 0142-0615. [41] India meteorological department, Visakhapatnam (2023) [accessedMay 12],
[29] A.R. Battula, S. Vuddanti, S.R. Salkuti, A day ahead demand schedule strategy for https://mausam.imd.gov.in/visakhapatnam/.
optimal operation of microgrid with Uncertainty, Smart Cities 6 (1) (2023) [42] J. Purdon, O. Bahn, G. Lachapelle, C. Marceau, Modeling Increased electric vehicle
491–509, https://doi.org/10.3390/smartcities6010028. charging demand in Quebec. EMI project report, Montreal (QC): Energy Modelling
[30] V. Rigoni, A. Keane, Open-DSOPF: an open-source optimal power flow formulation Initiative (2020). Available from, https://emi-ime.ca/wp-content/uploads/2020/0
integrated with OpenDSS, in: 2020 IEEE Power & Energy Society General Meeting 2/Iqcarbone_Purdon_Bahn_Modeling_Increased_Electric_Vehicle_Charging_Dema
(PESGM); 2020 Aug 2-6; Montreal, QC, Canada, Piscataway (NJ), IEEE, 2020, nd_In_Qc.pdf.
pp. 1–5. [43] Commonwealth Edison Company, Day-Ahead Prices, 2023 [accessedMay 12], htt
[31] R. Rana, M. Singh, S. Mishra, Design of modified droop controller for frequency ps://hourlypricing.comed.com/live-prices/day-ahead-prices/.
support in microgrid using fleet of electric vehicles, IEEE Trans. Power Syst. 32 (5) [44] A.R. Battula, S. Vuddanti, A chance constraint microgrid energy management with
(2017) 3627–3636, https://doi.org/10.1109/TPWRS.2017.2651906. Sept. phase balancing using electric vehicle demand aggregation, Energy Sources Part A
[32] S.J. Sultanuddin, R. Vibin, A. Rajesh Kumar, Nihar Ranjan Behera, M. Jahir Pasha, 45 (1) (2023) 111–139, https://doi.org/10.1080/15567036.2022.1964635.
K.K. Baseer, Development of improved reinforcement learning smart charging [45] Z. Yi, Y. Xu, W. Gu, Z. Fei, Distributed model predictive control based secondary
strategy for electric vehicle fleet, J. Energy Storage 64 (2023), 106987. ISSN 2352- frequency regulation for a microgrid with massive distributed resources, IEEE
152X. Trans. Sustain. Energy (2020).
[33] C. Song, S. Yoon, V. Pavlovic, Fast ADMM algorithm for distributed optimization [46] Z. Wang, D. Yue, J. Liu, Z. Xu, A Stackelberg game modelling approach for
with adaptive penalty, in: Proceedings of the Thirtieth AAAI Conference on aggregator pricing and electric vehicle charging, in: 2019 IEEE 28th International
Artificial Intelligence and the Twenty-Eighth Innovative Applications of Artificial Symposium on Industrial Electronics (ISIE); 2019 June 12-14; Vancouver, BC,
Intelligence Conference, 2016, https://doi.org/10.7282/T3RV0QM9. Published Canada, Piscataway (NJ), IEEE, 2019, pp. 2209–2213, https://doi.org/10.1109/
online. ISIE.2019.8781294.
[34] Zheng Xu, Gavin Taylor, Hao Li, Mário A.T. Figueiredo, Xiaoming Yuan, Tom
goldstein proceedings of the 34th international conference on machine learning,
PMLR 70 (2017) 3841–3850.

15