PAKMamba: Enhancing electricity load forecasting with periodic aggregation and Koopman analysis

Abstract

In recent years, with the increasing proportion of renewable energy in power systems, the difficulty of system dispatching has also increased. Accurate power load forecasting is an important prerequisite for achieving flexible dispatching. Power load exhibits significant daily and weekly periodicity and non-stationary characteristics. Current deep learning models cannot fully capture the periodicity and non-stationary characteristics of power load, leading to insufficient prediction accuracy and scalability. To address this issue, this paper proposes a new prediction model, PAKMamba, which consists of a dual-layer Mamba, a Periodic Aggregation module, and a Koopman Temporal Detector. The dual-layer Mamba handles the forward and backward dependencies of the sequence in parallel. The Periodic Aggregation module is used to extract the periodic properties of the sequence to capture its local features. The Koopman Temporal Detector, combined with Koopman dynamics theory, more effectively handles the non-stationarity between sequences. Validation on five datasets demonstrates that PAKMamba achieves more accurate predictions compared to other benchmark models, with an average MSE of 0.1324 and an average MAE of 0.252 for four-step predictions on the AEP(American Electric Power) dataset.