Short-term Electricity Load Forecasting with Mixture of Experts

Enhancing accuracy in electricity load forecasting is essential for efficient real-time power scheduling and energy market operations across residential and commercial sectors. Existing machine learning and deep learning approaches often struggle to adaptively capture the inherent complexity and non-linear temporal patterns, such as diurnal cycles, seasonal variations in electricity load data. To address this challenge, we propose MoE-MEL, a novel Mixture of Experts (MoE) based fusion framework that integrates multiple electricity load forecasting models. Instead of relying on a single model, MoE-MEL could leverage multiple linear-centric experts with a router that combines expert outputs with adaptive weights. Thus, each expert learns to specialize in distinct temporal patterns, while the router facilitates the dynamic orchestration of collaboration among these experts. Extensive experiments on hourly PJM electricity load dataset demonstrate that the proposed MoE-MEL could consistently outperform all baseline models in terms of forecasting accuracy.