EPformer: Unlocking day-ahead electricity price forecasting accuracy using the time–frequency domain feature learning strategy considering renewable energy

Accurate day-ahead price forecasting is critical for power market participants in the electricity market. However, the increasing penetration of renewable energy introduces greater complexity and intermittency into electricity price patterns, making accurate day-ahead forecasting a significant challenge. To this end, we propose a novel framework for day-ahead electricity price forecasting, EPformer. Specifically, the proposed framework begins with a two-stage data preprocessing module. Then, the proposed framework adopts an encoder–decoder architecture. It first employs Bidirectional Long Short-Term Memory (BiLSTM) as the temporal encoder to capture sequential dependencies and utilizes Temporal Convolutional Networks (TCN) as the feature encoder to extract data features, respectively. The learned representations and additional data features are then fused and fed into the decoder to generate the final prediction. In addition, the proposed framework is trained using a customized loss function that integrates time–frequency domain features. This strategy replaces the conventional MSELoss training paradigm, which enables the model to effectively capture peak and valley features in electricity price, while also mitigating the inherent autocorrelation in the label sequences under the direct forecasting (DF) paradigm. Finally, we conduct a comprehensive evaluation and validation of the proposed model on two electricity price datasets from Shanxi in China.