Multi-distribution fusion based Bayesian deep neural network for short-term probabilistic electricity price forecasting

Abstract

Probabilistic electricity price forecasting (EPF) is paramount for stakeholder scheduling and trading in deregulated energy markets. However, during the process of establishing a probability prediction model, the consequential aleatoric and epistemic uncertainties are indistinguishable. This indistinguishability undermines the credibility and interpretability of forecasts, especially in complex and volatile electricity markets. This article proposes a hybrid approach under a Bayesian probability framework for performing reliable and accurate probabilistic forecasting. Specifically, it applies Bayesian deep learning model to convey the latent probability distribution of electricity prices, which represents the desirable epistemic uncertainty and inevitable aleatoric uncertainty. The embedded architecture employs a multi-head attention mechanism (MHAM) to allocate previous hidden states of long short-term memory (LSTM) cells, detecting both the local and global price patterns simultaneously. Considering the forecasting bias arising from aleatoric uncertainty in the process of establishing Bayesian networks, a new Wasserstein distance-based multi-distribution fusion mechanism driven by evidence theory is proposed and employed to identify and mitigate the aleatoric uncertainty of credible probabilistic EPF problems. The effectiveness of the proposed approach is evaluated on two real-world electricity price datasets and compared with fifteen state-of-the-art models. The proposed approach achieves CRPS scores of 6.50 and 10.21 on the respective datasets, outperforming other benchmark models and demonstrating superior predictive accuracy and robustness across diverse market conditions.