A multi-scale feature extraction and fusion framework based on wavelet Kolmogorov–Arnold networks and parallel Bi-directional gated recurrent units for electric load forecasting

Short-term electric load forecasting remains challenged by the dual requirements of accuracy and robustness due to the combined effects of strong seasonality, multi-scale spikes, and stochastic disturbances. To address this, we propose a novel multi-scale forecasting framework, NP-WavKAN-Fusion, which integrates Neural Prophet for data decomposition and a Wavelet-based Kolmogorov–Arnold Network (WavKAN) with learnable wavelet kernels for multi-scale encoding. This fusion model utilizes a Bi-directional Gated Recurrent Unit (BiGRU) to capture long-term temporal dependencies and an adaptive feature fusion gate (AFF) to dynamically re-weight static and dynamic features for final load predictions. Extensive experiments on two public datasets from Australia and Morocco show that NP-WavKAN-Fusion consistently outperforms traditional models, reducing the mean absolute error by at least 30 %. For multi-step forecasting tasks, NP-WavKAN-Fusion maintains error inflation within 15 %, demonstrating superior performance compared to state-of-the-art long-sequence models such as Informer and PatchTST. The Diebold–Mariano test confirms that NP-WavKAN-Fusion yields statistically significant improvements, with 19 out of 20 comparisons showing lower errors. Ablation studies show that removing either the Neural Prophet component or the AFF significantly increases the forecasting error, validating the necessity of our layered denoising and fusion strategies. The proposed NP-WavKAN-Fusion framework demonstrates strong potential for real-world applications in electric load forecasting, offering robust performance under various temporal and non-stationary conditions.