Accurate short-term power forecasting is crucial for enhancing the reliability and operational efficiency of modern power systems. However, the non-stationary and multi-scale fluctuating characteristics of time-series data pose significant challenges to existing forecasting models. To address this issue, this paper proposes an innovative short-term power forecasting framework. This framework incorporates a teacher knowledge fusion module based on hybrid expert gating and a student model based on the iTransformer architecture. Specifically, we leverage multiple pre-trained teacher models with distinct inductive biases to provide complementary guidance, dynamically fused through a hybrid expert gating network. The student model adopts a dual-branch encoder architecture with reverse attention to enhance temporal feature modeling capabilities. In addition, a dynamic data augmentation strategy with random segment replacement is proposed to improve model robustness. Experimental results on four public datasets—ETTh1, ETTh2, ETTm1, and ETTm2—demonstrate that our method outperforms mainstream baseline models in both prediction accuracy and generalization capability. Notably, it achieves ∼5% improvement over Informer on the mean squared error metric.
Wang et al. (2026) studied this question.