Short Term Electricity Price Forecasting Using Hybrid Machine Learning And Feature Selection Techniques

Authors: Manjesh kumar, Dr. Jaya Shukla, Dr. Rajnish Bhasker

Abstract: Short-term electricity price forecasting (STEPF) plays a crucial role in modern deregulated electricity markets by enabling effective energy trading, demand-side management, grid stability, and operational planning. However, the highly volatile and nonlinear behavior of electricity prices, influenced by factors such as electricity demand, renewable energy integration, fuel prices, weather conditions, and market uncertainties, makes accurate forecasting a challenging task. To address these challenges, this paper proposes a novel hybrid Machine Learning framework that integrates advanced data preprocessing, adaptive feature selection, multi-stage deep neural learning, residual error correction, and explainable artificial intelligence (XAI) into a unified forecasting architecture. Initially, missing values are estimated, outliers are removed, time-series decomposition is performed, and normalized sliding-window sequences are generated to improve data quality. A hybrid feature selection strategy combining Mutual Information, ReliefF, and Adaptive Grey Wolf Optimization identifies the most informative features while reducing redundancy and computational complexity. The selected features are then processed through a hybrid CNN–BiLSTM–Temporal Attention–Transformer network to capture local spatial characteristics, long-term temporal dependencies, and global contextual relationships. Furthermore, a lightweight Extreme Learning Machine-based residual correction module refines the initial predictions by minimizing residual forecasting errors. To enhance model transparency, SHAP and Integrated Gradients are employed to interpret feature contributions and prediction behavior. Experimental analysis conducted on a realistic synthetic electricity market dataset demonstrates that the proposed framework achieves superior forecasting performance with an MAE of 0.48, RMSE of 0.67, MAPE of 1.21%, and an R² score of 0.995, outperforming conventional machine learning and Machine Learning benchmark models while maintaining robust generalization and computational efficiency.