A novel correlation feature self-assigned Kolmogorov-Arnold Networks for multi-energy load forecasting in integrated energy systems

Abstract

The prediction of multi-energy load in an integrated energy system (IES) is crucial for facilitating the integration of renewable energy and energy scheduling. However, the multi-energy load and its related variables exhibit strong coupling, correlation quality, and uncertainty. More specifically, the short-term correlation degree and stability of the load variables are inconsistent, significantly impacting the accuracy of the final prediction model. Therefore, this paper proposes a novel correlation features self-assigned Kolmogorov-Arnold Network (KAN) for multi-energy load prediction. Initially, a multi-decoder Informer model is utilized to encode the multi-energy load variables. The encoded features are fused using random sample self-combination and a correlation feature self-assignment module. Subsequently, the decoder is employed for energy co-decoding. The final decoded features are employed to construct a predictive model using interpretable KAN. The proposed algorithm is validated on an open-source dataset. Simulation results demonstrate that compared with Transformer and Informer algorithms, the average RMSE of multi-energy load prediction achieved by our proposed algorithm is reduced by 27.880% and 40.176%, respectively; Additionally, the robustness of the proposed model has been confirmed, and the relative error of prediction for multi-energy load data with and without noise is strictly limited to the range [−0.02, 0.02].