Long-term electricity load forecasting model based on multi-period and multi-scale analysis

Existing long-term power load forecasting models face significant challenges due to the complexity of non-stationary electricity data, inadequate information representation, and suboptimal prediction accuracy. To address these issues, this study introduces PowerNet, a novel forecasting framework based on a multi-period and multi-scale approach. PowerNet first employs Fast Fourier Transform (FFT) to perform multi-period analysis, transforming one-dimensional time series data into two-dimensional (2D) tensors to capture latent periodic patterns. A multi-scale information representation module is then implemented using computer vision-inspired techniques, including Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and non-local attention mechanisms to extract robust features from these tensors. Finally, a fully connected layer performs feature reconstruction, while residual connections facilitate multi-step load forecasting. PowerNet was evaluated against state-of-the-art models such as Informer, LogTrans, and ETSformer across two public electricity datasets and one weather dataset. At prediction horizons of 96, 192, 336, and 720 time steps, experimental results demonstrate that PowerNet achieves competitive or superior accuracy on the evaluated datasets. Notably, compared with Autoformer, PowerNet reduces the Mean Absolute Error (MAE) by 18.77% to 37.41% and the Mean Square Error (MSE) by 18.23% to 47.74%, significantly enhancing the reliability of long-term electricity load forecasting.