Time Series-Based Electric Load Forecasting with Mixture of Expert System

In this paper, a novel approach for electric energy prediction utilizing a hybrid expert system integrated with a time series model is introduced. The proposed method leverages the concept of a hybrid expert system to construct multiple time series expert models, each equipped with its corresponding gating module, thereby enabling the selection of diverse expert models for enhanced electric energy prediction accuracy. The methodology comprises several key steps. First, the construction of a time series expert model involving the integration of three distinct time series methods. Second, the training of the time series expert model, encompassing hyperparameter optimization and parameter optimization procedures for each constructed model. Third, the model blending process is based on the hybrid expert system, wherein the gating module is jointly trained alongside the time series expert models, allowing for the dynamic mixing of expert models based on varying input characteristics. Finally, the electric energy prediction is executed using the mixed expert combination within the hybrid expert system. Experimental results on real-world datasets demonstrated the effectiveness compared to competitive baselines.