Short-term electricity price forecasting (STEP) plays a pivotal role in ensuring the stability and economic efficiency of modern power markets, especially under conditions of high volatility and increasing renewable penetration. However, existing models fail to adequately capture extreme price spikes, hierarchical periodic structures, and the dynamic impact of exogenous variables. To overcome these challenges, we propose Fourier-EPNet, a novel deep learning framework that integrates frequency-domain attention with multimodal variable fusion. Specifically, it features: (i) a Fourier Softmax attention mechanism, which extracts dominant periodic signals while suppressing high-frequency noise; and (ii) an exogenous-endogenous cross-attention module, which dynamically aligns historical price trends with forward-looking external drivers such as load and wind forecasts. Extensive experiments on five benchmark datasets (NP, PJM, BE, FR, DE) from the EPF corpus show that Fourier-EPNet consistently surpasses state-of-the-art baselines, achieving 49.5% lower MSE and 40.4% lower MAE on average. Ablation studies and theoretical visualization further validate the contribution and interpretability of each component. Overall, Fourier-EPNet offers a robust, interpretable, and generalizable solution for real-world electricity price forecasting, setting a strong foundation for intelligent energy market decision-making.
Sun et al. (Sun,) studied this question.