ABMF-Net: An Attentive Bayesian Multi-Stage Deep Learning Model for Robust Forecasting of Electricity Price and Demand

Abstract

This article presents a novel deep learning model, the Attentive Bayesian Multi-Stage Forecasting Network (ABMF-Net), designed for robust forecasting of electricity price (USD/MWh) and demand (MW). The model incorporates an attention-based data selection mechanism, an encoder-decoder structure with masked time-series prediction, and a Bayesian neural network to generate both point and interval forecasts. Furthermore, a multi-objective Salp Swarm Algorithm (MSSA) is used to optimize forecasting accuracy and stability. Experimental evaluation on four real-world datasets from the Australian electricity market demonstrates that ABMF-Net achieves a MAPE as low as 1.89%, MAE of 0.67, RMSE of 0.98, and FICP of 0.98, outperforming LSTM, GRU, and Transformer models. Seasonal evaluations confirm the model’s robustness across high-variability conditions. These results position ABMF-Net as a high-performing and reliable forecasting model for modern electricity markets.