CPERNN: An RNN-based Multi-scale Variable Interaction Model for electrical Time Series Forecasting

Long-Term Time Series Forecasting plays a critical role in key domains such as electrical load prediction. Particularly, electrical time series data exhibit characteristics like multi-variability, high frequency, and strong periodicity, posing challenges for modeling long-term dependencies. Due to its recursive computation and error accumulation, traditional Recurrent Neural Network (RNN) architecture struggle to effectively capture long-term temporal dependencies. Moreover, the existing channel-independent strategy overlooks dynamic interactions between variables. To address this issue, this study proposes CPERNN, an RNN-based multi-scale variable interaction forecasting model. Its core innovation lies in the design of a variable relationship compensation mechanism, effectively mitigating the limitations of RNN in capturing long-term dependencies and the restricted variable information interaction under the channel-independent strategy. The model employs a segment-wise strategy to model long-term dependencies and utilizes a multi-scale variable interaction feature compensation module to enhance variable modeling capabilities, as well as adopts parallel multi-step prediction to improve overall accuracy. Experiments on the dataset Electricity demonstrate CPERNN's outstanding performance in 720-time-step forecasting tasks, achieving an MSE of 0.216 and reducing MAE to 0.311, significantly outperforming existing mainstream methods. This highlights its broad application prospects and academic research value in complex electrical time series forecasting tasks.