Sensitivity Analysis in Predicting Photovoltaic Systems Performance Utilizing Random Forest

This study systematically investigates the sensitivity of the Random Forest (RF) model to training dataset size in predicting photovoltaic (PV) system performance, a critical factor for smart grid stability and reliability. Using comprehensive operational data, including dust accumulation and ambient temperature, from a 7.98 kW grid-connected PV system in Jordan, the impact of varying training fractions on model accuracy, Coefficient of Determination (R²), Adjusted R², and Mean Square Error (MSE) is assessed rigorously. Through Bayesian Optimization for hyperparameter tuning and multiple replicates, the findings consistently demonstrate that larger training datasets yield superior predictive performance and enhanced model stability. An optimal training fraction of 70% was identified, achieving high R² and low MSE in addition to minimum computational time. While the prediction accuracy exhibited remarkable robustness even with reduced data, significant degradation in other metrics and increased variability were observed with training fractions below 30%. This research underscores the paramount importance of judicious training data selection for developing robust and reliable RF-based PV forecasting models, thereby facilitating more effective renewable energy integration.