Abstract
Accurate day-ahead electricity consumption forecasting is important for electricity market trading and cost management. In this paper, a decomposition-driven dual-branch network is proposed for this task. First, the original consumption sequence is decomposed into trend and residual components, which are modeled by a linear forecasting branch and an improved ModernTCN branch, respectively. In the residual branch, gated multi-scale blocks are employed to extract temporal patterns at different scales, and the resulting features are adaptively fused using input-dependent weights. The proposed method is evaluated on real-world electricity consumption data from Ansteel Mining. It achieves the highest mean number of qualified test days among the compared methods. Across five independent runs, an average of 178.8 qualified days out of 182 test days is obtained, with a standard deviation of 1.1 days. In ablation experiments, the increase in qualified days is mainly attributed to trend decomposition, while average errors and their variability across runs are reduced by the gated multi-scale block. On a public industrial load dataset, the proposed method achieves the lowest mean absolute error among the compared methods. The effectiveness of the proposed method for day-ahead industrial electricity consumption forecasting is demonstrated by these results.