Interpretable short-term electric load forecasting based on Bayesian Optimization for AM-CNN-KAN

Abstract

Accurate power load forecasting is crucial for modern power system planning, yet existing methods often struggle to balance high accuracy with model interpretability. To address this research gap, we propose AM-CNN-KAN, a novel hybrid model integrating Kolmogorov-Arnold Networks (KAN), Convolutional Neural Networks (CNN), Multi-Head Attention Mechanism, and Bayesian Optimization. While previous KAN-based studies focus mostly on basic predictive performance, our approach uniquely positions AM-CNN-KAN to explicitly address both multi-scale feature extraction and black-box interpretability issues in power systems. Compared to the baseline KAN model, AM-CNN-KAN achieved significant improvements in MAE, MSE, RMSE, and MAPE by 0.072%, 7.52%, 2.44%, and 1.23%, respectively. Furthermore, we integrated the SHAP method to decompose feature contributions, thereby ensuring that the forecasting mechanism is highly interpretable. Overall, our method not only achieves state-of-the-art accuracy but also provides practical and transparent guidance for grid operations.