Abstract:Currently, deep learning methods have become prevalent in the field of electrical load forecasting. These approaches have shown a great potential to map complex nonlinear feature interactions. However, many existing electrical load forecasting models based solely on deep learning suffer from various limitations including the inability to perceive and integrate multi-scale features, the absence of continuous information of electrical load series and capturing fine-grained and hidden temporal pattern of electrical load series. In order to address these issues, in this paper we propose an improved model based on neural ordinary differential equations (NODEs), which possesses the ability of adaptive fusion, multi-scale feature perception, and representation. This model strengthens the effective decomposition of multi-scale features with the NODE-series block and enhances the multi-scale feature extraction and fusion ability with the NODE-split block. The experimental results show that the proposed model outperforms five baseline models and three block ablation experiments prove the necessity of the blocks.

Electrical Load Forecasting Based on the Fusion of Multi-Scale Features Extracted by Using Neural Ordinary Differential Equation