Enhancing power load forecasting accuracy under high renewable energy penetration with a MSN-KAN framework: A novel approach to mitigate non-stationarity and enhance interpretability

• We propose an interpretable long-term time series forecasting method based on the MSN-KAN framework. • A novel MSFDN approach is proposed to model complex trend variations and seasonal patterns in non-stationary data. • The SNeLU activation function enhances the model’s robustness and generalization capability. • We innovatively propose a quantifiable evaluation metric, NWE, for assessing forecasting models. • The MSN-KAN model outperforms state-of-the-art methods in prediction accuracy while offering interpretability. With the increasing penetration of renewable energy, electric power load forecasting faces growing challenges due to strong non-stationarity, complex temporal patterns, and the limited interpretability of deep learning models. Existing forecasting methods often fail to effectively capture multi-scale dynamic features while providing transparent and reliable model explanations. To address these issues, this paper proposes an improved Kolmogorov-Arnold Networks (KAN) architecture—Multi-scale Normalized-Feature Dynamic (MSN-KAN)—incorporating an interpretable Multi-scale Feature Dynamic Normalization (MSFDN) mechanism to model complex trends and seasonal variations in non-stationary load data. By integrating the advantages of ReLU and GELU, the Smooth Soft NeLU Activation Function (SNeLU) is adopted to enhance model robustness and generalization capability. An innovative Nonlinear Weighted Entropy (NWE) metric is developed, integrating multi-dimensional evaluation indicators including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R 2 ) to establish a quantitatively consistent model evaluation framework. Experimental results demonstrate that the MSN-KAN model exhibits superior performance while improving model interpretability. This approach provides an efficient and interpretable solution for load forecasting in modern power systems with high renewable energy penetration.