Sarcouncil Journal of Engineering and Computer Sciences

Abstract: Rooftop photovoltaic (PV) systems are essential for urban renewable energy adoption, but their efficiency is compromised by meteorological variability, requiring accurate performance evaluation and predictive modeling for effective grid integration and maintenance planning. This empirical investigation analyzes 60 grid-connected rooftop PV installations at the Hong Kong University of Science and Technology (HKUST) campus, using a high-resolution dataset from January 2021 to December 2023. The dataset features inverter-level PV parameters (DC/AC voltages, currents, power output) and meteorological data (global horizontal irradiance, ambient and module temperatures, relative humidity, wind speed/direction, atmospheric pressure, precipitation) at 1-minute intervals, encompassing over 94 million records with >92% completeness. Three machine learning algorithms—Support Vector Machine (SVM), Random Forest (RF), and Multilayer Perceptron (MLP)—were trained to forecast normalized PV power output using selected weather features. Models were optimized through grid search and 5-fold cross-validation. Performance metrics included Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²). The RF model achieved the highest accuracy (RMSE: 4.8 W, MAE: 3.6 W, MAPE: 5.2%, R²: 0.96), outperforming SVM (RMSE: 6.2 W, MAE: 4.5 W, MAPE: 6.8%, R²: 0.93) and MLP (RMSE: 5.1 W, MAE: 3.8 W, MAPE: 5.9%, R²: 0.95). SHAP-based sensitivity analysis identified global horizontal irradiance (GHI) as the primary influencer (45% importance), followed by temperature (25%) and humidity (15%). Annual degradation averaged 1.0%, largely due to humidity-induced potential-induced degradation (PID). These results validate ML's role in subtropical PV forecasting, facilitating yield optimization and cost reductions of 10–15% through predictive maintenance.