Short-Term Electricity Load Forecasting Using Bayesian Regularization-Based Neural Network: A Case Study in Langsa City

Abstract

This study focuses on enhancing the accuracy of artificial neural network (ANN) methods in electricity load prediction for intelligent energy systems. Various optimization techniques, such as Bayesian regularization, have been introduced to improve model performance and generalization capability. A major challenge in ANN-based prediction models is overfitting, which occurs when the network topology fails to generalize input–output relationships, leading to poor prediction accuracy on unseen data. The research aims to develop an improved electricity load prediction model for Langsa City by applying a Bayesian regularization algorithm to minimize overfitting in the neural network topology. A quantitative experimental approach was used, which multiple ANN architectures with historical electricity load datasets. The Bayesian regularization algorithm optimized weight adjustments and minimized mean squared error during training. Results indicate that the proposed model effectively reduces overfitting and enhances predictive accuracy, achieving a Mean Absolute Percentage Error (MAPE) of 2.45%. These findings demonstrate that Bayesian regularization significantly enhances ANN reliability, stability, and forecasting capability for future intelligent energy management applications