Electric Power Systems Research 238 (2025) 111112

Contents lists available at ScienceDirect

Electric Power Systems Research

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

Forecasting data-driven system strength level for inverter-based

resources-integrated weak grid systems using multi-objective machine
learning algorithms
Md Ohirul Qays a,*, Iftekhar Ahmad a , Daryoush Habibi a, Mohammad A.S. Masoum b
a
School of Engineering, Edith Cowan University, Joondalup, WA, 6027, Australia
b
Department of Engineering, Utah Valley University, Orem, UT, 84058, USA

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

Keywords: Shortage of grid-fault level, known as system strength inadequacy, impacts on grid instability and can lead to
Forecasting blackouts. System strength is generally measured by short circuit ratio index at point of coupling (POC) of
Hedge-backpropagation inverter-based resources (IBRs) and the grid system. Nowadays, accurate knowledge of system strength fore-
Inverter-based resource
casting for ‘next day’ to ‘next week’ duration is essential to power system operators, owing to the higher-growth
Machine learning
Synchronous condenser
of IBRs. However, releavant publications about this subject remain limited when compared with load demand,
System strength active and reactive power prediction. Therefore, a data-driven system strength forecasting scheme is presented in
this paper to surmount these issues. Multi-objective machine learning (MOML) algorithms are used to obtain the
best result. The designed model uses energy management system (EMS) to collect historical online data and
complete the training and testing procedures via learning frameworks such as Hedge-backpropagation neural
network-based tangent function (Hedge-BPNNT), support vector machine (SVM) and long short-term memory
(LSTM). The methodology is developed to predict up to seven days of system strength forecasting levels by using
the last thirty-days data status. The designed model is tested on both simulated and experimented cases, con-
firming higher accuracy performance with reduced computational time when compared to existing literature.

1. Introduction Accordingly, system strength prediction is essential to support grid-

strength and prevent unexpected contingencies. To date, some publi-
In accordance with the 2023 DNV report, global electricity genera- cations have focused on power generation and energy demand fore-
tion forecast from different energy sources is demonstrated in Fig. 1 [1]. casting. For example, various kinds of machine learning (ML) algorithms
In each year, applications of IBRs for global electricity generation are have been reviewed in [4] to forecast energy consumption, with rec-
estimated to grow by approximately 15 PWh. By 2033, above 50 % of ommendations of data normalization to increase model accuracy and
electricity generation is predicted to use IBRs, which will increase to 80 reduce computation time. In [5], an extreme learning machine (ELM)
% by 2050. Since IBRs are generally installed in remote areas or far from was implemented to forecast wind energy systems in China with regard
synchronous generators (SGs), fault level/system strength shortfall is to uncertainty analysis. The designed model presented a feasible solu-
experienced. To evaluate system strength level, short circuit ratio (SCR) tion with 1.0581 % error reduction as compared to variational mode
index is widely implemented. Low SCR value represents weak-grid decomposition and multi-objective salp swarm algorithms. Deep recur-
system that experience post-fault grid instability issues, failure of con- rent neural networks (DRNN) was pursued to forecast long-term power
tingencies, and other features. To overcome these difficulties, system system state estimations [6]. Data was collected from an IEEE 118-bus
strength supply can be provided using different kinds of devices, for system for training purposes, where a case study revealed 0.012 p.u.
example, STATCOM, SVC, capacitors bank, synchronous condenser voltage magnitude error at Bus 100.
(SynCon) and others [2]. Among these devices, SynCon is a feasible Herewith, forecasting has been classified into four types in [7] for
solution because of its faster post-fault recovery response and power power system engineering, including: ultra short-term, short-term,
flow quality [3]. mid-term and long-term, which are further explained in Table 1 [8,9].

* Corresponding author at: School of Engineering, Edith Cowan University, 270 Joondalup Drive, Joondalup, WA, 6027, Australia.
E-mail address: [email protected] (M.O. Qays).

https://doi.org/10.1016/j.epsr.2024.111112
Received 19 March 2024; Received in revised form 10 July 2024; Accepted 21 September 2024
Available online 30 September 2024
0378-7796/© 2024 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

sequence-to sequence network (TSeq2Seq Net) to perform hour-ahead

and day-ahead forecasts up to seven days. However, high errors, up to
12.42 % were found in the applied algorithm. In [11], reactive power
forecasting was studied for ‘next-day’ dynamic plan with collected data
from EDP Distribuição company using SCADA device. Their model
reduced power loss by approximately 12 % with zero penalties. As a
result of this, it is recommended to apply machine learning algorithms to
fix poor data issues and to assess large-scale situations.
Before the learning process occurs in forecasting, training data
batches are essential, and this requires memory for data storage, where
there is a lack of knowledge about the convergence difficulty from dis-
appearing the gradient and deteriorating the feature reuse. To address
these issues, Hedge-backpropagation (HBP) has been proposed in [12]
to update network parameters in online at data screaming. Accordingly,
Higgs and Susy datasets were implemented and the method was
Fig. 1. Global electricity generation forecasting. compared with Linear and Kerner learning methods. The HBP method
showed the least online cumulative error rate (0.3739). Further to this,
the long-short-term-memory (LSTM) method was implemented in [13]
Table 1
to precisely forecast an online 10 MW wind power system. 100,000
Forecasting classification based on prediction period.
NREL data sets were collected in a five minute sampling system that used
Forecasting Period range Applications the HBP learning approach. Results showed 0.0580 MSE, with a
type
recommendation to apply the designed method in PV and load fore-
Ultra short- Few seconds to half an hour • Electricity market casting systems. In [14], the SVM learning algorithm was used to fore-
term ahead • Regulation activities cast ‘next day’ reactive load prediction. Data from two 220 kV lines of
Short-term Half an hour to six hours • Load incremental or
ahead decremental decision
Shanxi Power Grid were extracted using EMS, showing poor results with
• Load dispatching plan a 2 MVAr relative error.
• Economic analysis Additionally, energy operator industry AEMO reported on October
Mid-term Six hours to one day ahead • Electricity market 2022 about system strength shortfalls (around 500 MVA) for July 2026
• Decisions on offline or online
within the Victoria province in Australia. It has also suggested to
generator
• Operational security establish a proper system strength forecasting technique to perform net-
Long-term One day to one week or more • Energy generation zero carbon emission target in renewable energy zones [15]. In [16],
duration ahead • Load calculation system strength forecasting has been recommended for system strength
• Cost analysis dynamic plan up to the next five years, which is not possible due to
• Maintenance scheduling
changing operation regimes and exact information of IBRs.
On the basis of the discussion above, although researchers and power
The forecasting classification was established according to its prediction system operators have suggested about the system strength forecasting
period. Based on this, short-term reactive power forecasting was studied for both ‘next day’ and ‘next week’ dynamic plans, publications on this
in [10] by collecting data from Multifamily Residential Electricity topic remain limited. Thus, this paper aims to study system strength
Dataset (MFRED). The study customized attention-based Transformer forecasting plans up to the next seven days. The proposed model ensures

Fig. 2. ML algorithms [17].

2
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

Fig. 3. Schematic diagram of Hedge-BPNNT.

error minimization with less computation time using MOML algorithms. accurate results are obtained. For example, K-Means, Gaussian
The key contributions of this research are as follows: Mixture, Hidden Markov etc.
iii. Semi-supervised learning: It is a hybrid method of supervised and
• A day-ahead system strength forecasting technique is proposed to unsupervised learning algorithms. This learning method uses
predict up to ‘next week’ plan. both labeled and unlabeled datasets as a combined dataset. In
• System strength at POCs is measured over a minimum threshold real-world application, since the unlabeled datasets are found in
using SCR index. maximum compared to the labeled datasets, semi-supervised
• MOML algorithms such as Hedge-BPNNT, SVM and LSTM are learning algorithms are useful in this perspective. However, it is
employed for training and testing datasets. mildly complex and less accurate when compared to the super-
• Error determination and quality improvement functions are used to vised learning method. For example, LSTM, Graph-based Method
verify the performance accuracy of learning approaches and are then etc.
compared to existing publications. iv. Reinforcement learning: This method improves the efficiency of a
• Experimental and simulation models such as SMIBPS and IEEE 39 particular context and its environments. It is primarily an optimal
bus modified system are investigated to assess larger power systems. decision-making algorithm-based on the reward or penalty pur-
poses. One of its fundamental objectives is to enhance the reward
Remining parts of this paper are organized as follows: Section 2 or reduce any associated risks by analysing the environmental
describes the classification of ML algorithms and the selected learning activities. This learning algorithm is also highly complex and less
schemes. The proposed strength forecasting method is detailed in Sec- accurate for forecasting purposes. For example, Q-Learning, R-
tion 3. In Section 4, obtained results of different case studies are dis- Learning, Monte Carlo etc.
cussed. Section 5 concludes the research paper.
As per the above learning approaches, to obtain higher accurateApplied
re-
2. Machine learning algorithms sults, a few commonly applied supervised and semi-supervised learning
ML
algorithms are applied in this paper such as BPNNT, SVM and LSTM as
algorithm
Multi-objective Machine Learning (MOML) algorithms highly described below: s to in
contribute substantial role to build an effective forecasting model this
depending on their learning abilities and the data sets characteristics. As 2.1. Hedge-BPNNT
paper
demonstrated in Fig. 2, four kinds of ML algorithms are widely used:
supervised learning, unsupervised learning, semi-supervised learning, Widely applied BPNN algorithm [18,19] needs data samples prior to
and reinforcement learning [17].The learning algorithms are detailed as the training process and fails to update parameterization during online
follows: data streaming. Vanishing gradient descent and feature refusal make
further learning convergence issue. On the contrary, tangent hyperbolic
i. Supervised learning: It yields a learning function to map between activated Hedge-BPNN (BPNNT) is an online-based algorithm because of
input and output data from the provided samples. A labeled/ its self-learning characteristic in non-linear systems and feedback fea-
identified training data is used in supervised learning algorithms tures at data streaming condition [12,13]. As shown in Fig. 3 [20], on-
to imply the function. This learning method is further categorized line historical data is collected from EMS and start processing.
into classification and regression schemes that use target vari- Afterwards, data is pre-processed with normalization by (1) since input
ables as discrete and continuous forms respectively. Although this data with different magnitudes may distort prediction results.
method is slightly expensive due to data labeling, it presents low xi − xmin
complexity with higher accuracy results. For example, back- x∗i = (1)
xmax − xmin
propagation neural network (BPNN), support vector machine
(SVM), Naive Bayes etc. Here, x∗i = [0, 1], xi , xmin and xmax are normalized data, raw input
ii. Unsupervised learning: In this learning method, two different data, minimum input data and maximum input data respectively. With
kinds of data such as clustering, and association are employed. this, data are sorted as training and testing data sets that lead to variable
Both of these schemes use unlabeled data for prediction purposes. combinations as data subsets which are forwarded to input layer. Net
Since data are not labeled, this method is highly complex and less input of the hidden-layer neuron (nethi ) is determined by (2),

3
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

⎧ ( )
⎪ f = softmax h ω ; ∀l = 0, …, L
⎪
⎪ (l) (l) (l)
⎪
⎪
⎪
⎨ ( )
h(l) = τ h(l− 1) wi (l) ; ∀l = 0, …, L (8)
⎪
⎪
⎪
⎪
⎪ 1
⎪
⎩ λ(l) = ; l = 0, …, L
L+1
Here, weighted parameter λ(l) needs to be additionally learnt during
online parameterization. It is involved with the initial forecasting data. l
is the number of hidden layers. With this, loss function L (y(x)) of the
forecasting data is used from (9) to update the other weighted param-
eters (10) to (12).
L
∑ ( )
L (y(x)) = λ(l) L f (l) (xl , yl ) (9)
l=0

λ(l+1) ←λ(l) βL (f (xl , yl )) ; β ∈ (0, 1) (10)

(l)

( )
ω(l+1) ←ω(l) − ηλ(l) ∇ω(l) L f (l) (xl , yl ) (11)

L
∑ ( )
wi (l+1) ←wi (l) − η λ(l) ∇ω(l) L f (l) (xl , yl ) (12)
l=0

Here, β and ω(l) are discount rate and lth hidden layer weighted
parameter. Further, the accuracy is checked. If the performance is
satisfied, the built network is saved, otherwise weights are updated to
enhance the accuracy. Lastly, results are predicted when no subsets are
remained.

2.2. SVM

Using SVM models, forecasting is achieved by finding optimal clas-

Fig. 4. Schematic diagram of SVM.
sification surfaces. It minimizes the error performance from training
data samples obtained by optimal classification surfaces [21]. As illus-
n
∑ trated in Fig. 4, the collected data samples are scaled with normalization
nethi = xi wih + hbi (2)
i=1
by (1) and sorted for training and testing data sets (xi , yi ). Here, xi ∈ Rn ,
yi ∈ R, i = 1, 2,…,n. SVM regression function (r(x)) is calculated by (13)
Here, wih and hbi are weights between input to hidden layer neurons [22].
and bias of hidden-layer neurons. Herewith, the activation function (τ)
of hidden layer neurons (hi ) is calculated by (3) r(x) = wφ(x) + d (13)

1 − exp (− 2nethi ) Here, w, φ(x) and d are weight-vector, non-linear mapping and offset
hi = τ(nethi ) = (3) degree respectively. The SVM regression model is solved by trans-
1 + exp (− 2nethi )
forming it into an optimization problem using loss minimization func-
Net input of output layer neurons (netyi ) is calculated by (4) to (5) tion by (14) to (15),
n
∑ n
netfi = fin (4) 1 ∑ ( )
min ‖ w‖2 + C ρi + ρ∗i (14)
i=1 2 i

fin = hi wiy + ybi (5) ⎧

⎨ yp − wφ(xi ) − d ≤ s + ρi
⎪
⎪
Here, wiy and ybi are weights between hidden layer to output neurons s.t. yp + wφ(xi ) + d ≤ s + ρ∗i (15)
and bias of output-layer neurons. Herewith, the activation function (τ) of ⎪
⎪
⎩ 0 ≤ ρ , 0 ≤ ρ∗
i i
output-layer neurons (fi ) is calculated by (6)
1 − exp (− 2netyi ) Here, C, s, and (ρi , ρ∗i ) are penalty vector, precise estimation and
fi = τ(nethi ) = (6) relaxation vectors respectively. To precisely define the non-linear
1 + exp (− 2netyi ) ( )
mapping, Kernel function (k xi , xj )-based Radial Basis Function (RBF)
To work with data stream, online-based Hedge-BPNNT is proceeded is implemented as follows (16) to (17)
from (7) to (12). The forecasting data y(x) is calculated by (7) to (8). ( ) ( )
k xi , xj = φ(xi )φ xj (16)
L
∑
y(x) = λ(l) f (l) ; i, n ∈ l = 0, …, L (7) ( ) ( )
l=0 k xi , xj = exp − q ‖ xi − xj ‖2 (17)

Here, q represents the kernel function variances.

2.3. LSTM

LSTM model supports to learn the relationship between long-term

4
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

∑m ⃒⃒ ⃒
i=1 yi − yi
( ) ∗⃒
MAE yi , y∗i = (23)
m
m ⃒ ⃒
( ) 1 ∑ ⃒yi − y∗i ⃒
MAPE yi , y∗i = ⃒ ⃒ × 100% (24)
m i=1 ⃒ yi ⃒

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑m ( )
∗ 2
i=1 yi − yi
( )
RMSE yi , y∗i = (25)
m
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅/(√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ )
m m m
( ) 1 ∑ ( )2 1 ∑ 1 ∑
U yi , y∗i = yi − y∗i yi 2 + y∗ 2 (26)
m i=1 m i=1 m i=1 i

Here, m represents number of the applied data points, yi and y∗i

represent actual value and predicted value of ith data.

3.2. Grey relational (GR) analysis

Dynamic GR analysis is considered in this paper that measures the

similarity between actual value and predicted value. The reference (Y0 ),
comparison (Yi ) and standard (yi (t)) series are prepared by (27) to (29)
respectively [26].
Y0 = Y0 (1), …, Y0 (m); m ∈ R (27)

Fig. 5. Schematic diagram of (a) LSTM layer (b) Operation block.

Yi = Yi (1), …, Yi (m); i = 1, 2, …, m ∈ R (28)
∑m ∑m
i=1 Yi (t) − m i=1 Yi (t)
1
dependency of data sequences (x) and time steps (t). As presented in yi (m) = √
√ )̅; m ∈ i
̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( (29)
Fig. 5(a), it analyses data through LSTM operation block, memory cell c, √ ∑m ∑ m
i=1 Yi (t) − m i=1 Yi (t)
√ 1 1
and hidden cell h. The LSTM operation block is, furthermore, illustrated m− 1

in Fig. 5(b) that consists of input node (gt ), input gate (it ), forget gate (ft ),
( )
and output gate (ot ). Output response ot at time t is predicted applying Here, GR coefficient μ(s) and its grade G yi , y∗i are calculated by (30)
( )
(18) to (22) by the interaction among ct− 1 , ht− 1 and xt [23]. to (31) respectively. Higher value of G yi , y∗i ∈ [0, 1] ensures the
(
ft = σ wfx xt + wfh ht− 1 + bf
)
(18) distinguished similarity.
⃒ ⃒ ⃒ ⃒
mini mins ⃒y0 (s) − yi (s)⃒ + ρmaxi maxs ⃒y0 (s) − yi (s)⃒
it = σ (wix xt + wih ht− 1 + bi ) (19) μ(s) = ⃒ ⃒ ⃒
⃒y0 (s) − yi (s)⃒ + ρmaxi maxs ⃒y0 (s) − yi (s)⃒
⃒ ; s ∈ i, ρ ∈ (0, 1]
(
gt = τ wgx xt + wgh ht− 1 + bg
)
(20) (30)

m
( ) 1 ∑
ct = ct− 1 ⊙ ft + gt ⊙ it (21) G yi , y∗i = μ(i) ∈ [0, 1] (31)
m i=1
ht = τ(ct ) ⊙ ot (22)
Here, w and b are weighted vectors and bias of the particular gates, ⊙ 3.3. Pearson analysis
presents elementwise/dot multiplication; σ and τ are sigmoid and
tangent activation functions respectively. To assess the correlation between actual data and forecasting data,
Pearson coefficient analysis is performed [27]. The coefficient value
varies from -1 to 1. The higher positive value is obtained, the stronger
3. Model accuracy ( )
linear relationship is ensured. The Pearson coefficient P yi , y∗i is
In power system forecasting approaches, accuracy is one of the calculated by (32) to (33) [27].
topmost additional important factors to reduce the loss functions values. ∑m ( )( )
i=1 yi − yi ŷi − ŷ∗i
( ) ∗
Because, the accuracy improvement increases substantial profits in in- P yi , y∗i = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
) ∈ [− 1, 1] (32)
∑m ( ) ∑ (
dustrial applications. Thus, this section comprehends error analysis and i=1 yi − yi
∗ 2
× m i=1 yi − yi
̂ ̂∗ 2
quality of forecasting results analysis as follows: ⎧
⎨ 1; Strong relationship
P= 0; No relationship (33)
3.1. Multiple error analysis ⎩
− 1; Opposite direction

When evaluating the performance of ML models, relying on a single Here, ŷi and ŷ∗i are mean values of the actual data and forecasting
error metric often provides an incomplete assessment of the model’s data respectively.
accuracy and reliability [24,25]. Accordingly, to measure the perfor-
mance accuracy of the selected MOML algorithms, the following widely 3.4. Forecasting quality analysis
used error minimization functions are considered such as mean absolute
( ) ( )
error MAE yi , y∗i , mean absolute percentage error MAPE yi , y∗i , root The production of IBRs are weather dependent and in turn, the sys-
( ) ( )
mean square error RMSE yi , y∗i and Theil’s statistic U yi , y∗i by (23) to tem strength prediction gets uncertain. Therefore, along with the above
(26) correspondingly. mentioned deterministic forecasting method, probabilistic forecasting

5
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

n
( ) 1∑
FICP yi , y∗i = bi × 100%; (34)
n i=1
{
1; yi ∈ [Li , Ui ]
bi = (35)
0; Otherwise

n
( ) 1 ∑
FINAW yi , y∗i = (Ui − Li ) ∈ [0, 1] (36)
nr i=1
⎧
⎪ L i − yi
⎪
⎪ ; Ai < Li
⎨ U i − Li
⎪
⎪
⎪
AWDi = 0; yi ∈ [Li , Ui ] (37)
⎪
⎪
⎪ L i − yi
⎪
⎪
⎪
⎩ ; Ai > Ui
U i − Li

n
( ) 1 ∑
AWD yi , y∗i = AWDi ∈ [0, 1] (38)
nr i=1

Here, bi ∈ {0, 1} is a Boolean index that is determined by (29). Li and

Ui are lower limit and upper limit of the forecasting interval respec-
tively, r = ±3 % is the variation range of the predicted value.

4. Proposed system strength forecasting

System strength considers reactive power supply by SynCon in MVAr

which is further determined by SCR index. Basic SCR (BSCR) is princi-
pally applied to measure system strength that considers the interaction
between short-circuit capacity (SCCi ) at ith POC and ratings of the
installed IBRs (PIBRi ) on that bus by (39) to (40) [28,29].
√̅̅̅
SCCi = 3Vi If ; for i = 1, …, N (39)

SSCCi
BSCRi = (40)
PIBRi
Here, Vi is instant voltage before the fault occurrence at ith bus and If
is the fault current of that bus.
As illustrated in Fig. 6, the proposed study is performed by collecting
data from experimented and simulated systems [30,31]. The experi-
mental part was conducted in the laboratory environment as detailed in
the sub-Section 4.1. Moreover, to confirm the effectiveness of the
designed methodology, two larger systems were simulated in MAT-
LAB/SIMULINK environment which are described in sub-Sections 4.2
and 4.3 correspondingly. Pseudo code of the applied MOML algorithms
are also explained in Algorithm 1 [32]. The selected MOML algorithms
collect datasets of power systems voltage (Vi ), current (Ii ), and angle
Fig. 6. Flowchart of the proposed methodology.
difference (θi ) from each ith POC. Data are processed and normalized to
avoid accuracy distraction and sorted as training and testing data sets.
strategy is further implemented in this research to analyse the fore- Afterwards, iteration is run until the error is minimized using MAE,
casting quality. Accordingly, two broadly applied probabilistic metrics MAPE, RMSE, U, G and P calculations. Meanwhile, different parameters,
such as, forecast interval coverage probability (FICP) and forecast in- for instance, weights, biases etc. are updated for the best accuracy. When
terval normalized average width (FINAW) are employed to evaluate the the network-model is established, system strength level in VAr and SCR
forecasting quality [5]. Probability of predicted data within the range of results are achieved which are stored in a data storage system for further
forecasting interval is assessed by the FICP model. FICP reflects the analysis. Accordingly, the results are compared from where the best
probability that the actual value falls within the predicted fluctuation MOML model is selected that produce system strength forecasting.
interval and allows the reliability of the prediction model to be assessed.
Larger FICP percentage value from (34) to (35) ensures that high
number of predicted data is found within the forecasting interval. 4.1. Experimental
Additionally, FINAW reflects the information of the interval width by
(36). In contrast, when the interval width is too high and predicted data The projected experimental configuration for system strength
is outside the range of the forecasting interval, FICP and FINAW results compensation is explained in Fig. 7. A three-phase AC IBR-integrated
can’t be reasonably granted. For this reason, accumulated width devi- power system was established along with a 2.8 MW variable load in
ation (AWD) metric is further used by (37) to (38) to fix the uncertainty the laboratory environment. A 105 kVAr rated SynCon was installed to
issue. Both FINAW and AWD metrics are varied between 0 and 1. The compensate system strength and load fluctuation. In all cases of this
less value presents the less deviation of the predicted data. paper, the employed SynCons received excitation current from a DC
source through the field winding and runs with the help of stator

6
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

Algorithm 1
Pseudo code for moml algorithms.

Fig. 7. System strength compensation diagram.

7
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

machine infinite bus power system (SMIBPS) model to verify the effec-
tiveness. From Fig. 8 [33], the SMIBPS model includes a 275 MW rated
doubly fed induction generator (DFIG) converters-based wind farm
system. The energy farm comprises 108 inverters each rated 2.65 MW
capacity. Additionally, two individual transformers each for 265 MVA,
TF1 and TF1–2 are installed with two buses B1 and B2 respectively. TF1
transforms generated voltage 1.5 kV into 67 kV medium voltage level
which is further transformed into 227 kV high voltage level. A SynCon
rated 27 MVA is further connected in bus B1 along with a variable load
Fig. 8. Implemented SMIBPS model.
to support system strength. The infinite bus is modeled with 227 kV, 50
Hz, 187 Ω impedance, X/R = 1.95, and B1 bus ratings SCR = 1.33, SCC =
265 MVA. The lumped π-transmission line 250 km is modeled for posi-
winding. Parameters such as instantaneous voltage (V), current (I) and
tive sequence values, for example admittance, conductance and sus-
their angle differences (θd ) were stored in an energy management system
ceptance are valued 42.85 S 18.59 S 8.25 S correspondingly. The power
(EMS) device. The EMS analysed data and calculated instantaneous
flow data are stored in EMS for the system strength forecasting analysis.
active/load power (P), reactive power (Q), and power factor (PF), by
(41) to (44). Herewith, system strength was measured in terms of kVAr
4.3. IEEE 39-bus system
and BSCR indexes. The projected data was collected from May 19 to
June 26, 2022 using EMS. The training and testing datasets were chosen
The proposed approach is implemented in a modified IEEE 39-bus
in terms of the first four weeks and the last week datasets respectively.
larger power system as presented in Fig. 9 for the system strength
Sampling frequency of the datasets was set to 10 min.
forecasting analysis. Four IBRs are installed in this network; battery farm
θ d = θV − θ I (41) 225 MW at Bus 19, a solar farm 245 MW at Bus 25 and two wind farms
(435 MW and 455 MW) are installed at Bus 20, and Bus 23, corre-
PF = cosθd (42) spondingly. For case purposes, data Bus 20 is collected and analysed in
this paper with SCR = 1.65, X/R = 1.67 and SCC = 720 MVA. For the
P = VIcosθd (43) grid stability, a 50 MVA SynCon is connected at Bus 20 along with a
variable load demand (Loadavg = 431 MW, Loadpeak = 443 MW). An EMS
Q = VIsinθd (44) is installed in Bus 20 to store power flow data which is analysed for the
forecasting purposes.
4.2. SMIBPS

The proposed model is tested on a wind farm integrated single

Fig. 9. IEEE 39-bus test-bed system.

8
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

2.05 % respectively) when compared to LSTM method (2.19 %). Even,

forecasting quality enhanced about 2.20 % in Hedge-BPNNT learning
scheme determined by high FICP (97.7483 %) and reduced FINAW
(0.0521) and AWD (0.00013) indexes. The worst condition is resulted in
LSTM learning method because of the Day 4 to Day 5 proper prediction
failure. Hedge-BPNNT method, nevertheless, well-performed with for
computation time shrank by 0.048 s.

5.3. IEEE 39-bus system

As demonstrated in Fig. 12(a) and (b) for the fluctuated 435 MW load
power at Bus 20 of the modified IEEE 39-bus system, system strength
value (70 ± 10 MVAr) and its associated BSCR index for a week are
presented respectively. From Table 6, Hedge-BPNNT and SVM algo-
rithms performed 2.18 % and 2.35 % error with 0.467 s and 0.499 s to
achieve the target value. In contrast, LSTM needed 0.523 s with 2.97 %
error which is relatively high. Moreover, FICP value for Hedge-BPNNT
and SVM learning methods are almost same although 0.08 % less
value is noticed for LSTM method. FINAW (0.0437) and AWD (0.00005)
values are constantly less in Hedge-BPNNT method than the rest
methods. It also presented worse condition on Day 4 at the system
strength fluctuation period. Hedge-BPNNT learning method, however,

Table 3
Forecasting quality analysis for experimental case.
Fig. 10. Forecasting (a) system strength (b) BSCR index for experimental data. Forecasting method FICP FINAW AWD Computation Time (sec)

BPNNT 96.6847 0.0687 0.00017 0.785

5. Results and discussion SVM 95.5261 0.0723 0.00025 0.841
LSTM 94.4669 0.0745 0.00039 0.873
System strength forecasting results are discussed in the section for
both experimental and simulation cases. The comparative ML algo-
rithms are scrutinized in terms of considering the target value of system
strength level as described below:

5.1. Experimental

System strength level and the BSCR index of the experimental test-
bed system is presented in Fig. 10(a) and (b) respectively. Due to the
variable load power (around 2.8 MW), system strength altered from 92
kVAr to 112 kVAr approximately in each day. At the end of June 19,
2022, the MOML algorithms predicted almost similar value of the target
value. After five days, however, only Hedge-BPNNT predicted the best
results during fluctuation compared to SVM and LSTM algorithms.
Herewith, the forecasting error, forecasting quality, calculation time etc.
are summarized in Tables 2 and 3, where Hedge-BPNNT method illus-
trated the least forecasting error such as MAE (0.00237), MAPE (2.17
%), RMSE (0.00291), U (0.00163), G (0.97253), P (0.99957), fore-
casting quality increment 1.16 % and less computation time 0.785 s.

5.2. SMIBPS

From Fig. 11(a) in SMIBPS model, system strength variations of the

whole day long for a week are seen along with the approximate load
power 275 MW. It varied between 9 MVAr to 29 MVAr. Fig. 11(b) also
depicts the BSCR values of that week. In both figures, Hedge-BPNNT and
SVM algorithms showed the predicted values close to the target value.
Fig. 11. Forecasting (a) system strength (b) BSCR index for SMIBPS data.
From Table 4 to Table 5, the error was also achieved nearby (2.01 % and

Table 2
Forecasting accuracy for experimental case.
Forecasting method MAE MAPE RMSE U G P

Hedge-BPNNT 0.00237 2.17 0.00291 0.00163 0.97253 0.99957

SVM 0.00249 2.23 0.00299 0.00172 0.96671 0.99949
LSTM 0.00265 2.31 0.00307 0.00185 0.95287 0.99933

9
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

Table 4
Forecasting accuracy for SMIBPS case.
Forecasting method MAE MAPE RMSE U G P

Hedge-BPNNT 0.00112 2.01 0.00233 0.00159 0.98647 0.99989

SVM 0.00127 2.05 0.00245 0.00167 0.97423 0.99983
LSTM 0.00185 2.19 0.00261 0.00193 0.96869 0.99975

improvement 2.107 % that confirmed the superiority of the designed

Table 5 system strength forecasting approach.
Forecasting quality analysis for SMIBPS case.
Forecasting method FICP FINAW AWD Computation time (sec) 6. Conclusion
Hedge-BPNNT 97.7483 0.0521 0.00013 0.597
SVM 96.6645 0.0543 0.00021 0.621 A data-driven system strength forecasting method is proposed in this
LSTM 95.5749 0.0559 0.00031 0.645 paper in order to address demand profile by following industrial rec-
ommendations. SynCon has been implemented to compensate system
strength level measured by BSCR index. Supervised and semi-supervised
multi-objective ML algorithms have been implemented in this research
to determine a feasible forecasting solution. Online historical data was
collected from experimental and simulated models to explore this study.
The experiment was conducted in a laboratory environment from May
2022 to June 2022. Additionally, two larger power systems, SMIBPS and
the IEEE 39-bus test-bed, were designed for data collection. Based on the
provided training and testing datasets, the ablative model predicted
results up to next seven days. In all cases, Hedge-BPNNT outperformed
high accuracy than SVM and LSTM learning techniques. The forecasting
quality comparatively increased above 2.11 % in Hedge-BPNNT
learning scheme. The proposed method was further compared with
the existing benchmark publications, showing a minimum error of only
1.93 % and reduced 2.50 % above computation time. From this point,
the precise solution solely contributed by designing an accurate fore-
casting model of the grid system connected to higher-penetration of
IBRs. Future studies are recommended for the application of advanced
ML algorithms, its control systems and cost analysis of IBRs-integrated
grid systems.

CRediT authorship contribution statement

Md Ohirul Qays: Writing – original draft, Visualization, Validation,

Software, Resources, Project administration, Methodology, Investiga-
tion, Funding acquisition, Formal analysis, Data curation, Conceptuali-
zation. Iftekhar Ahmad: Validation, Supervision. Daryoush Habibi:
Fig. 12. Forecasting (a) system strength (b) BSCR index for IEEE 39-
Validation, Supervision. Mohammad A.S. Masoum: Supervision.
bus system.

Declaration of competing interest

performed better results to obtain the target value rather than the other
methods (Table 7 [36]. The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
5.4. Comparison the work reported in this paper.

This study accomplished a day-ahead forecasting system to predict

system strength demand in industrial applications. The proposed
research work is, in addition, compared with the existing publications in
Table 7
Table 8 to present its effectiveness. For prediction evaluation, from
Forecasting quality analysis for IEEE 39-bus case.
Table, a few learning methods are already utilized in different litera-
Forecasting method FICP FINAW AWD Computation time (sec)
tures, such as ELM, DRNN, backpropagation neural network-based sig-
moid (BPNNS) function, TSeq2Seq, LSTM, BPNNT etc. Amongst them, Hedge-BPNNT 99.8847 0.0437 0.00005 0.467
Hedge-BPNNT-based proposed study showed the least error perfor- SVM 99.4381 0.0465 0.00011 0.499
LSTM 98.9865 0.0479 0.00019 0.523
mance 1.93 %, determination time 0.469 s and forecasting quality

Table 6
Forecasting accuracy for IEEE 39-bus case.
Forecasting method MAE MAPE RMSE U G P

Hedge-BPNNT 0.00103 2.18 0.00233 0.00137 0.98883 0.99995

SVM 0.00117 2.35 0.00241 0.00145 0.98875 0.99987
LSTM 0.00135 2.97 0.00267 0.00151 0.96869 0.99979

10
M.O. Qays et al. Electric Power Systems Research 238 (2025) 111112

Table 8
Comparison with existing publications.
Ref Year Data collection Method Computation time (sec) Error (%) Quality improvement (%)

[5] 2019 Simulated ELM 0.849 2.074 1.037

[6] 2019 Simulated DRNN 1.422 3.153 1.009
[34] 2020 Experimental LSTM 1.583 13.287 0.647
[35] 2022 Industrial building BPNNS 0.937 5.145 0.941
[10] 2023 Residential building TSeq2Seq 1.246 12.421 0.725
[36] 2024 Simulated grid system Classifier-Regression Mapping 1.460 9.529 0.501
Proposed 2024 Experimental, Simulated Hedge-BPNNT 0.469 1.932 2.107

Data availability [17] I.H. Sarker, Machine learning: algorithms, real-world applications and research
directions, SN Comput. Sci. 2 (2021) 160, https://doi.org/10.1007/s42979-021-
00592-x.
Data will be made available on request. [18] O. Qays, Y. Buswig, L. Hossain, S.M. Ieee, M. Rahman, A. Abu-siada, S.M. Ieee, A.
C. Modeling, Active cell balancing control strategy for parallel connected LiFePO4
batteries, CSEE J. Power Energy Syst. (2020) 1–8, https://doi.org/10.17775/
CSEEJPES.2020.00740.
Acknowledgement [19] O. Qays, Y. Buswig, M.L.H. Hossain, A. Abu-Siada, Recent progress and future
trends on state of charge estimation methods to improve battery-storage efficiency:
Funding of the research publication is supported by School of Engi- a review, CSEE J. Power Energy Syst. 1 (2020) 1–12, https://doi.org/10.17775/
CSEEJPES.2019.03060.
neering, Edith Cowan University, Australia. [20] S. Hien, J. Hansen, J.E. Drewes, K. Koch, BioTOOL—a readily and flexible biogas
rate prediction tool for end-users, Environ. Model. Assess. 24 (2019) 87–94,
References https://doi.org/10.1007/s10666-018-9609-3.
[21] J. Li, Y. Lei, S. Yang, Mid-long term load forecasting model based on support vector
machine optimized by improved sparrow search algorithm, Energy Reports 8
[1] DNV, Energy transition outlook 2023, 2021. https://eto.dnv.com/2021.
(2022) 491–497, https://doi.org/10.1016/j.egyr.2022.02.188.
[2] Australian Energy Market Operator, System Strength Impact Assessment
[22] Z. Liu, D. Wu, Y. Liu, Z. Han, L. Lun, J. Gao, G. Jin, G. Cao, Accuracy analyses and
Guidelines, 2018. https://www.aemo.com.au/-/media/Files/Electricity/NEM/Se
model comparison of machine learning adopted in building energy consumption
curity_and_Reliability/System-Security-Market-Frameworks-Review/2018/Syste
prediction, Energy Explor. Exploit. 37 (2019) 1426–1451, https://doi.org/
m_Strength_Impact_Assessment_Guidelines_PUBLISHED.pdf.
10.1177/0144598718822400.
[3] S. Hadavi, M.Z. Mansour, B. Bahrani, Optimal allocation and sizing of synchronous
[23] X. Liu, S. Fan, J. Qin, Y. Liu, W. Wang, Long-term prediction method of reactive
condensers in weak grids with increased penetration of wind and solar farms, IEEE
load based on LSTM neural network, IOP Conf. Ser. Earth Environ. Sci. 467 (2020),
J. Emerg. Sel. Top. Circuits Syst. 11 (2021) 199–209, https://doi.org/10.1109/
https://doi.org/10.1088/1755-1315/467/1/012041.
JETCAS.2021.3053554.
[24] N. Mehdiyev, D. Enke, P. Fettke, P. Loos, Evaluating forecasting methods by
[4] S.S. Arnob, A.I.M.S. Arefin, A.Y. Saber, K.A. Mamun, Energy demand forecasting
considering different accuracy measures, Procedia Comput. Sci. 95 (2016)
and optimizing electric systems for developing countries, IEEE Access 11 (2023)
264–271, https://doi.org/10.1016/j.procs.2016.09.332.
39751–39775, https://doi.org/10.1109/ACCESS.2023.3250110.
[25] R.J. Hyndman, A.B. Koehler, Another look at measures of forecast accuracy, Int. J.
[5] W. Yang, J. Wang, H. Lu, T. Niu, P. Du, Hybrid wind energy forecasting and
Forecast. 22 (2006) 679–688, https://doi.org/10.1016/j.ijforecast.2006.03.001.
analysis system based on divide and conquer scheme: a case study in China,
[26] J. Wang, W. Yang, P. Du, T. Niu, A novel hybrid forecasting system of wind speed
J. Clean. Prod. 222 (2019) 942–959, https://doi.org/10.1016/j.
based on a newly developed multi-objective sine cosine algorithm, Energy Convers.
jclepro.2019.03.036.
Manag. 163 (2018) 134–150, https://doi.org/10.1016/j.enconman.2018.02.012.
[6] L. Zhang, G. Wang, G.B. Giannakis, Real-time power system state estimation and
[27] H. Liu, Z. Duan, F. Han, Y. Li, Big multi-step wind speed forecasting model based
forecasting via deep unrolled neural networks, IEEE Trans. Signal Process. 67
on secondary decomposition, ensemble method and error correction algorithm,
(2019) 4069–4077, https://doi.org/10.1109/TSP.2019.2926023.
Energy Convers. Manag. 156 (2018) 525–541, https://doi.org/10.1016/j.
[7] S.S. Soman, H. Zareipour, O. Malik, P. Mandal, A review of wind power and wind
enconman.2017.11.049.
speed forecasting methods with different time horizons, in: North Am. Power
[28] M.O. Qays, I. Ahmad, D. Habibi, A. Aziz, T. Mahmoud, System strength shortfall
Symp., IEEE, 2010, pp. 1–8, https://doi.org/10.1109/NAPS.2010.5619586, 2010.
challenges for renewable energy-based power systems: a review, Renew. Sustain.
[8] H. Mansoor, M.S. Gull, H. Rauf, I. ul H. Shaikh, M. Khalid, N. Arshad, Graph
Energy Rev. 183 (2023) 113447, https://doi.org/10.1016/j.rser.2023.113447.
convolutional networks based short-term load forecasting: leveraging spatial
[29] A. Aziz, M.O. Qays, A. Jawad, M.F. Ishraq Faruqui, M.L. Hossain, A. Kobeissi,
information for improved accuracy, Electr. Power Syst. Res. 230 (2024) 110263,
Optimal placement and sizing of synchronous condenser to enhance system
https://doi.org/10.1016/j.epsr.2024.110263.
strength of a renewable energy integrated weak grids, in: 2023 6th Int. Conf.
[9] Ç.B. Bozlak, C.F. Yaşar, An optimized deep learning approach for forecasting day-
Electr. Inf. Commun. Technol., IEEE, 2023, pp. 1–6, https://doi.org/10.1109/
ahead electricity prices, Electr. Power Syst. Res. 229 (2024) 110129, https://doi.
EICT61409.2023.10427908.
org/10.1016/j.epsr.2024.110129.
[30] R. Bayindir, S. Sagiroglu, I. Colak, An intelligent power factor corrector for power
[10] H. Dong, J. Zhu, S. Li, W. Wu, H. Zhu, J. Fan, Short-term residential household
system using artificial neural networks, Electr. Power Syst. Res. 79 (2009)
reactive power forecasting considering active power demand via deep transformer
152–160, https://doi.org/10.1016/j.epsr.2008.05.009.
sequence-to-sequence networks, Appl. Energy 329 (2023) 120281, https://doi.org/
[31] AEMO, AEMO power system dashboard, 2023. https://aemo.com.au/en/energy-sy
10.1016/j.apenergy.2022.120281.
stems/electricity/national-electricity-market-nem/data-nem/data-dashboard-nem.
[11] R. Gonçalves, T. Soares, M. Louro, L.A.F.M. Ferreira, P.M.S. Carvalho, F. Carvalho,
[32] A. Shrestha, A. Mahmood, Review of deep learning algorithms and architectures,
J. Machado, Dynamic reactive power management based on forecasts and
IEEE Access 7 (2019) 53040–53065, https://doi.org/10.1109/
chronological power flow, CIRED - Open Access Proc. J. 2020 (2020) 305–308,
ACCESS.2019.2912200.
https://doi.org/10.1049/oap-cired.2021.0042.
[33] B. Baadji, H. Bentarzi, A. Bouaoud, SMIB power system model with PSS for
[12] D. Sahoo, Q. Pham, J. Lu, S.C.H. Hoi, Online deep learning: learning deep neural
transient stability studies, in: 2017 5th Int. Conf. Electr. Eng. - Boumerdes, IEEE,
networks on the fly, in: Proc. Twenty-Seventh Int. Jt. Conf. Artif. Intell.,
2017, pp. 1–5, https://doi.org/10.1109/ICEE-B.2017.8191996.
International Joint Conferences on Artificial Intelligence Organization, California,
[34] X. Zhang, Y. Wang, Y. Zheng, R. Ding, Y. Chen, Y. Wang, X. Cheng, S. Yue, Reactive
2018, pp. 2660–2666, https://doi.org/10.24963/ijcai.2018/369.
load prediction based on a long short-term memory neural network, IEEE Access 8
[13] C. Pan, S. Wen, M. Zhu, H. Ye, J. Ma, S. Jiang, Hedge backpropagation based online
(2020) 90969–90977, https://doi.org/10.1109/ACCESS.2020.2991739.
LSTM architecture for ultra-short-term wind power forecasting, IEEE Trans. Power
[35] D. Błaszczok, T. Trawiński, M. Szczygieł, M. Rybarz, Forecasting of reactive power
Syst. (2023) 1–14, https://doi.org/10.1109/TPWRS.2023.3304898.
consumption with the use of artificial neural networks, Electronics (Basel) 11
[14] L. Bo, X. Cheng, X. Liu, H. Zheng, J. Hao, The forecasting model of reactive power
(2022) 2005, https://doi.org/10.3390/electronics11132005.
based on SVM, in: 2020 IEEE 4th Inf. Technol. Networking, Electron. Autom.
[36] B.V. Surya Vardhan, M. Khedkar, I. Srivastava, S.K. Patro, Impact of integrated
Control Conf., IEEE, 2020, pp. 2136–2140, https://doi.org/10.1109/
classifier — regression mapped short term load forecasting on power system
ITNEC48623.2020.9084912.
management in a grid connected multi energy systems, Electr. Power Syst. Res. 230
[15] AEMO, AEMO Annual Planning Report, Australia, 2022. http://www.aemo.com.
(2024) 110222, https://doi.org/10.1016/j.epsr.2024.110222.
au/Electricity/Planning/Victorian-Annual-Planning-Report.
[16] AEMO, System strength in the NEM explained, March 2020, 2020. https://aemo.
com.au/-/media/files/electricity/nem/system-strength-explained.pdf.

11