Collaborative Learning Integration for Enhanced Photovoltaic System Power Forecasting: A Novel Ensemble Approach

The integration of photovoltaic (PV) systems into power distribution networks presents challenges due to the unpredictable nature of power output, which affects grid reliability and energy management. Accurate forecasting of a PV power output is therefore critical. This study proposes a novel collaborative learning intergration (CLI) stacking ensemble energy forecasting model approach, designed to forecast real-time hourly power output of a 39.02 kWp photovoltaic system using the historical real time power generation output of the system and the hourly time correlating meteorological data. Unlike conventional methods that simply average model predictions or develop a meta learner, the proposed topology allows individual machine learning models to interact and learn from each other through mutual reinforcement, thereby enhancing forecasting accuracy. The algorithm is evaluated against benchmark independent and stacked models to analyze its performance as an improved forecasting model. The experiment results show that the proposed collaborative learning integration stacked ensemble model outperforms the five benchmark models, achieving an R-square (R2), mean absolute error (MAE) and root mean square error (RMSE) accuracy of 0.89, 1.44 and 2.47 respectively. The residual analysis highlighted that the collaborative learning intergration model was more effective at minimising long-term forecast errors compared to other models. The proposed model can be highly useful for predicting the performance of small and large-scale solar PV power plants in a multi-step ahead forecasting.