Fuzzy clustering-based mixture of GRU experts for short-term photovoltaic power forecasting

Abstract Accurate photovoltaic (PV) power forecasting is essential for grid integration and energy management, yet remains challenging due to the inherent variability of solar irradiance and the concept drift phenomenon where data distributions evolve over time. This paper proposes FCM-MoE, a fuzzy clustering-based mixture of experts framework for ultra-short-term PV power forecasting. Unlike conventional day-level weather classification approaches, FCM-MoE employs a 28-dimensional state vector computed from rolling windows at 15-minute resolution, combined with PCA dimensionality reduction and Fuzzy C-Means (FCM) clustering to achieve fine-grained weather regime identification that captures intra-day transitions. To address data scarcity in minority weather regimes, we introduce a warm-start training strategy that initializes regime-specific expert models from a pre-trained global model, enabling effective knowledge transfer while preserving specialization capability. GRU networks serve as the expert backbone, offering competitive performance with reduced computational overhead compared to LSTM. Comprehensive experiments on real-world PV data from Gansu Province, China, demonstrate that FCM-MoE achieves normalized root mean square errors (nRMSE) of 8.00% and 10.44% for 1-hour and 2-hour ahead forecasting respectively, representing 22.7% and 25.6% improvements over the global GRU model for these operationally critical horizons.