Real-time efficient short-term peak load and day-ahead electricity load forecasting system using machine learning approach

Accurate short-term electricity demand forecasting is essential for grid reliability, operational efficiency, and energy resource planning. Traditional models often struggle with the highly non-linear and dynamic nature of load patterns, especially for high-resolution forecasts. To address this challenge, we propose a novel hybrid deep learning framework for forecasting 30-min interval peak electricity demand. The model utilises real-time demand data collected from January 1 to December 31, 2023. The input data undergo zero-mean normalisation and dimensionality reduction using Principal Component Analysis Network (PCANet) to enhance feature relevance. Temporal dependencies are captured via a Deep Walk Gated Recurrent Unit (DWGRU) model, which integrates spatial, temporal, and semantic information through a dynamic attention mechanism. A Hybrid Sampling and Self-Attention Deep Neural Network (HSSA-DNN), optimised using an Improved Moth Flame Optimisation (IMFO) algorithm, further refines the forecasting process. Experimental results demonstrate superior predictive performance, with a Mean Absolute Percentage Error (MAPE) of 0.85 %, Mean Absolute Error (MAE) of 14.3 MW, and Root Mean Square Error (RMSE) of 19.8 MW. The proposed model achieves a day-ahead forecasting accuracy of 99.15 %, significantly outperforming benchmark models such as LSTM, GRU, and standard DNNs by margins of 4.2 %–7.8 % in MAPE. These findings underscore the effectiveness of incorporating advanced feature extraction, temporal modelling, and optimisation techniques for high-precision load forecasting. The model's robust performance supports its potential application in real-time energy management systems and operational planning.