A multi-energy load forecasting method based on the Mixture-of-Experts model and dynamic multilevel attention mechanism

Multi-energy load forecasting is essential for energy management and scheduling optimization in Integrated Energy Systems. To address the challenges posed by the complex coupling, non-linearity, and significant load fluctuations among multi-energy loads, this paper proposes the Empirical Mode Decomposition-Mixture-of-Experts-Mamba-Attention forecasting model, which considers local features. In the feature engineering phase, this study employs correlation analysis to select key influencing factors, thereby reducing the interference of irrelevant factors. It also uses Empirical Mode Decomposition to decompose the highly complex load data, extracting components at different frequencies to effectively reduce the influence of non-linearity and noise. Furthermore, the decomposed load signals are reconstructed to extract clearer signals, providing a solid foundation for subsequent forecasting. In the forecasting model, this study introduces the Mixture-of-Experts-Mamba-Attention framework. This model employs a Mixture-of-Experts approach to simultaneously forecast cooling, heating, and electrical loads, with the Gating network dynamically assigning different weights to enhance forecasting accuracy. Additionally, the Mamba-Attention mechanism strengthens the model's ability to capture local features through a multi-level structure. The experimental results demonstrate that the performance metrics for cooling, heating, and electrical loads significantly outperform traditional methods, with MAPE values of 2.35 %, 2.02 %, and 2.67 %, respectively. These results validate the effectiveness and superiority of the proposed model.