BiKAN-LoadNet: An Electric Load Forecasting Model Based on Bidirectional LSTM and Kolmogorov–Arnold Networks

Electric load forecasting plays an important role in smart grid scheduling and operation optimization. However, existing methods still have limitations in multi-target joint modeling, integration of external weather features, and capturing nonlinear relationships. This paper proposes BiKAN-LoadNet, a multi-target load forecasting model that combines Bidirectional Long Short-Term Memory (BiLSTM) networks with a Kolmogorov–Arnold Network (KAN). The model uses BiLSTM to capture long- and short-term dependencies in load sequences, applies a residual connection to retain recent fine-grained variations, and integrates a weather feature encoding module to incorporate the impact of meteorological factors. The output layer adopts KAN to enhance nonlinear fitting capability and achieve joint prediction of daily average, maximum, and minimum loads. Experiments on real-world electric load datasets show that BiKAN-LoadNet achieves significant improvements in RMSE, MAE, and MAPE compared to existing methods. The results demonstrate that BiKAN-LoadNet effectively combines temporal patterns with external weather information, achieving high accuracy and strong stability in multi-target load forecasting tasks, and providing an efficient solution for smart grid load management. The code for reproducing main experiments are available at https://github.com/lianghesiit-code/BiKAN-LoadNet.