Photovoltaic power forecasting model employing epoch-dependent adaptive loss weighting and data assimilation

Abstract

Accurate prediction of photovoltaic (PV) power output is crucial for optimizing energy management systems and enhancing grid stability. This study presents the Physics Constrained PV Power Prediction Network (PC-P${}^3$reNet), a dual-layer deep learning framework optimized for scenarios where local environmental data remain consistent while PV system characteristics vary. The framework integrates a physics-based model to calculate theoretical PV power outputs, which are then compared with actual measurements using the Huber Loss function. A unique feature of PC-P${}^3$reNet is its adaptive loss weighting, which dynamically adjusts the balance between theoretical and measured data across different training epochs. This feature allows the model to initially leverage theoretical insights for learning and later refine its predictions based on measured data, effectively capturing both trends and variability. The model’s performance was evaluated using data from four PV stations in Australia. The model demonstrated superior performance in multi-step forecasting compared to other methods. It achieved a minimum mean absolute error (MAE) of 0.1837 at the No. 18 power station. The mean square error (MSE) improvement was 4.68% higher on average for the proposed model than the baseline method.