Physic and machine learning in photovoltaic systems modelingn: exploring hybrid approaches

Abstract

Hybrid modeling approaches that combine physical and data-driven techniques have shown potential for photovoltaic (PV) power estimation by leveraging the strengths of both individual techniques. Despite their increasing popularity, different hybridization approaches for PV applications and the impact of physical model selection on hybrid model performance remain underexplored. This thesis presents a comprehensive study on hybrid models for PV power prediction, introducing a novel classification into three categories: Physics-Informed Machine Learning (ML) models, Optimized Physical Models, and Physics-Guided Models. As a key contribution, an Optimized Physical model based on the energy balance of PV systems is proposed, with Bayesian inference applied to mitigate parameter uncertainties. The model is integrated into a Physical Model Chain (PMC) and combined with ML techniques using two hybridization approaches: Physics-Informed ML, which relies on ML models, and Physics-Guided models, which prioritize physical modeling. Combinations of four ML models (Multiple Linear Regression – MLR, K-Nearest Neighbors – KNN, Extreme Gradient Boosting – XGB, and Random Forests – RF) are tested with three PMC compositions: standard (PMC1), optimized model choice (PMC2), and the proposed energy balance-based (PMC3). The results show that ML model choice has the greatest impact on performance, since ensemble methods (RF and XGB) consistently present the lowest MAE values (0.030–0.032 kW/kWp), while MLR reaches 0.036–0.037 kW/kWp MAE. In comparison, PMC choice has a smaller effect, with differences across PMC1–PMC3 of 0.001 kW/kWp in MAE for most cases. Nevertheless, the proposed PMC3 surpasses classical models in both hybrid approaches, providing the most accurate foundation for hybrid frameworks. Overall, the Physics-Informed ML models combining PMC3 and XGB or RF yield the best accuracy (0.030 kW/kWp). This result demonstrates the potential of ML models to capture the underlying physics of PV conversion through feature engineering and accurately describe PV power conversion compared to primarily physical hybrid models. In conclusion, this work demonstrates the impact of physical model selection on hybrid model performance, compares hybridization strategies, and offers guidance on balancing accuracy, complexity, and generalizability in PV power modeling.