A Novel Short-Term Load Forecasting Model Based on Bidirectional Long Short-Term Memory and Kolmogorov-Arnold Network

With the development of modern power systems and the widespread adoption of smart grids, accurate short-term load forecasting has become increasingly important. Although Bidirectional Long Short-Term Memory (BILSTM) networks excel at capturing bidirectional dependencies in time series data, they exhibit limitations when dealing with highly nonlinear data and often lack interpretability. This paper proposes a novel prediction model based on BILSTM and Kolmogorov-Arnold Network (KAN). The model captures bidirectional contextual information through BILSTM and enhances nonlinear representation capabilities using KAN with learnable activation functions, thereby improving both interpretability and prediction accuracy. Experimental results demonstrate that the BILSTM-KAN model reduces the Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) by 16.8%, 16.39%, and 16.6%, respectively. Compared to existing models, the BILSTM-KAN model exhibits significant advantages in short-term load forecasting tasks.