A novel electricity price forecasting approach based on multi-attention feature fusion model optimized by variational mode decomposition

Abstract

Electricity price forecasting (EPF) is crucial for the optimal dispatch of energy markets. The increasing penetration of renewable energy for electricity generation has added more influencing variables to the electricity price curve, making the EPF more challenging. Therefore, this paper addresses electricity price data in energy markets with renewable energy generation and proposes an innovative Variational Mode Decomposition (VMD)-based multi-attention mechanism feature fusion model (V-MAF) for EPF. First, VMD processing reduces noise and captures multi-scale features in price and load sequences. Next, by integrating Gated Recurrent Units (GRU), Temporal Convolutional Networks (TCN), and Squeeze-and-Excitation Networks (SENet), a parallel network architecture combining SE-TCN and SE-GRU is constructed. This architecture captures local fluctuations and periodic patterns in VMD-separated multi-scale data, enhancing feature exploration and improving the model’s ability to fit price variations. Finally, the output features from both networks are combined and fed into a Multi-Head Attention (MHA) along with the original features, allowing the model to focus on different parts of the input features from multiple perspectives. The innovative architecture enhances the ability to capture multi-scale features in time series and further focuses on key features through adaptive weight allocation of the attention mechanism. Experiments on the Singapore dataset and ablation studies demonstrated the effectiveness of VMD, SENet, and MHA in enhancing network performance. Multi-model comparisons showed that the V-MAF model outperformed others, providing more stable and accurate predictions. On Dataset 1, the V-MAF model achieved the Root Mean Square Error (RMSE) of 1.3168, reduced errors by 11.09% to 59.13% compared to other models such as XGBoost, ATT-CNN-LSTM, BiGRU, and VMD-Transformer.