A Robust Ensemble Framework with Context-Gated Dynamic Fusion and Temporal Residual Correction for Photovoltaic Power Forecasting

Accurate forecasting of Photovoltaic(PV) power generation is critical for the stability and economic operation of modern smart grids. However, the inherent intermittency and volatility of PV power, driven by highly variable weather conditions, present significant challenges for prediction models. This paper proposes a novel, robust, and explainable ensemble learning framework to address these challenges. Our core contributions are fivefold: (1) A Context-Gated Mixture of Experts (MoE) mechanism that employs a lightweight gating network to dynamically assign weights to diverse base learners (LSTM, GRU, Transformer, and tree-based models) based on real-time input features, local prediction variance, and historical residuals, enabling superior adaptation to non-stationary environments. (2) A Temporal Residual Corrector (TRC) module that learns the complex error patterns between the MoE output and the true values, acting as a powerful post-processor for fine-grained prediction correction. (3) A knowledge distillation architecture to compress the powerful “MoE+TRC” teacher model into a lightweight student GRU network, drastically reducing computational cost for real-time deployment. (4) A robust uncertainty quantification method that fuses prediction intervals from MC-Dropout, LightGBM Quantile Regression, and Conformal Prediction to produce reliable confidence estimates. (5) Test-Time Augmentation (TTA) to enhance model robustness against noisy or missing input data. Comprehensive experiments on a real-world, multi-location PV dataset demonstrate that our framework significantly outperforms all baseline models and conventional ensemble techniques across multiple metrics, providing highly accurate, reliable, and uncertainty-aware forecasts for practical energy scheduling applications.