Novel Forecasting for Photovoltaic Installation Output Using Transfer Learning Genetic Programming

Abstract

Accurately predicting photovoltaic (PV) power generation is crucial for optimizing energy allocation and integrating solar energy into smart grids. PV generation forecasting faces challenges due to varying forecasting horizons, seasonal changes, and location-specific environmental factors. Transfer learning can improve accuracy and reduce computation by using knowledge from related tasks. Genetic Programming (GP) offers transparent, interpretable symbolic regression models compared to black-box methods. In this paper, a transfer-learning-based GP is applied to the power generation forecasting model of PV installation across different seasons. In the proposed method, mutual information is employed to identify the useful knowledge from constructed models as the source domain, and captured shared generation patterns are then embedded into the initial population of the GP model for the target domain, guiding the search process. The proposed method is evaluated on a case study leveraging the homologous power generation patterns across different seasons for the same PV installation to assist in constructing forecasting models. Experiments on real-world seasonal data show the method outperforms state-of-the-art algorithms, improving both accuracy and model explainability. These results highlight the potential of transfer learning in GP for PV installation generation forecasting, offering both performance improvements and model transparency, which are crucial for real-world deployment and interpretability.