TVNet: A times vision network for electrical load forecasting by temporal 2d-variation

Baichun Wang, Ao Sun, Zhikai Xu, Qinglun Zhang, Yan Shi, Hong Men

Abstract: Accurate power load forecasting is crucial for the safety and stability of all power system sectors. However, load data’s high volatility and uncertainty are significant challenges to load forecasting. This paper proposes the times vision network (TVNet) that combines sequence decomposition, two-dimensional convolution, and self-attention mechanisms to address these issues. An adaptive sequence decomposition method is introduced, where the decomposed components are transformed into a two-dimensional form and fed into a vision backbone network for feature extraction. Incorporating a horizontal self-attention mechanism, the model’s capability to effectively capture long-range dependencies in long sequence inputs is enhanced. Through empirical evaluations on three datasets with different meteorological and load characteristics, the results demonstrate that TVNet significantly outperforms the baseline models across all evaluation metrics. The MSE on the three datasets is reduced by 2.23 %, 3.22 %, and 0.27 % compared to the best-performing model, respectively. Additionally, for long sequence inputs, the MSE is reduced by 6.87 %, 2.68 %, and 1.87 % compared to conventional input lengths. Furthermore, compared to the Transformer model, the number of parameters is reduced by up to 10.7504 M, maintaining a low computational overhead.

External IDs:dblp:journals/eswa/WangSXZSM25