Answers are grounded in this paper and cite their sources.
Accurate short-term load forecasting is challenged by multi-scale load fluctuations and weather effects. This study proposes CVGKAN, a hierarchical decomposition and gated recurrent unit–Kolmogorov–Arnold network (GRU–KAN) framework. Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) separates historical load into intrinsic mode functions and a residual. Sample entropy and K-means reconstruct them into three components, and variational mode decomposition (VMD) refines only the high-complexity component. Each component is combined with historical weather and calendar variables and encoded by an independent GRU. A KAN output layer replaces the conventional linear head to forecast each component. These forecasts are summed to obtain the total load prediction. In 30 runs per horizon on the Panama electricity load dataset, CVGKAN achieved the lowest mean root mean square error (RMSE) for one-hour-ahead and direct 12- and 24-step forecasts. Its mean RMSE was 12.2%, 7.6%, and 7.6% lower than that of the baseline with the lowest mean RMSE at each respective horizon. Ablation results support the contributions of hierarchical decomposition, selective VMD, and KAN. One-hour-ahead experiments on four additional datasets also show the lowest mean errors among the compared models.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: