A Deep Learning and Metaheuristic Optimization Framework for
Short-Term Electricity Consumption Forecasting Using
High-Resolution SCADA Data
Wei Hong Lim1,*, Amel Ali Alhussan2
1Faculty of Engineering, Technology and Built Environment, UCSI University, Kuala Lumpur 56000,
Malaysia
2Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint
Abdulrahman University, P.O. Box 84428, Riyadh 11671, Saudi Arabia
Emails: limwh@ucsiuniversity.edu.my; aaalhussan@pnu.edu.sa
Abstract
Accurate prediction of electricity consumption is a critical requirement for improving operational efficiency,
enhancing grid reliability, and supporting sustainability objectives in urban power distribution systems,
particularly in regions experiencing steady population growth and increasing demand pressure. Motivated
by the limitations of conventional statistical and physics-inspired forecasting approaches, as well as the
strong sensitivity of deep learning architectures to hyperparameter configuration, t his s tudy p roposes a
robust data-driven framework that integrates deep learning with advanced metaheuristic optimization for
high-precision short-term electricity consumption forecasting. The main contribution of this work lies in
the systematic development and evaluation of hybrid metaheuristic–Bidirectional Long Short-Term Memory
(BiLSTM) models, in which multiple state-of-the-art optimization algorithms are employed to tune model
hyperparameters. Particular emphasis is placed on the integration of the Ninja Optimization Algorithm
with BiLSTM (NijOA + BiLSTM), which is designed to effectively navigate complex, high-dimensional
hyperparameter search spaces encountered in deep learning–based load forecasting tasks. Baseline
experiments demonstrate that BiLSTM outperforms other deep learning models, including Artificial Neural
Network (ANN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent
Unit (GRU), achieving a baseline Root Mean Squared Error (RMSE) of 0.0964 and a coefficient of
determination (R2) of 0.854. These results confirm t he a dvantage o f b idirectional t emporal l earning in
capturing the nonlinear and time-dependent characteristics of electricity consumption recorded at high
temporal resolution from SCADA systems. Following metaheuristic optimization, the NijOA + BiLSTM
model delivers a substantial improvement in predictive performance. The optimized configuration reduces
RMSE to 0.0038, Mean Squared Error (MSE) to 1.45 × 10−5, and Mean Absolute Error (MAE) to 0.00019,
while increasing the correlation strength to r = 0.973 and the explanatory power to R2 = 0.97. Comparative
analysis across different optimization strategies further confirms t he s uperiority o f t he NijOA + BiLSTM
hybrid model over alternative configurations, including WAO + BiLSTM, BBO + BiLSTM, GA + BiLSTM,
SFS + BiLSTM, DE + BiLSTM, and JAYA + BiLSTM. The implications of these findings are significant for
real-world urban electricity distribution applications. The proposed framework enables highly accurate and
reliable short-term electricity consumption forecasting, making it well suited for deployment within smart
grid and distribution management systems. Such predictive capability can support informed operational
decision-making, improve demand-side management strategies, reduce uncertainty in short-term planning,
and contribute to the long-term sustainability and resilience of urban power distribution networks.
Keywords: Short-term electricity consumption forecasting; Deep learning; BiLSTM; Metaheuristic
optimization; SCADA data