Photovoltaic power estimation and forecast models integrating physics and machine learning: A review on hybrid techniques

Abstract

Photovoltaic (PV) models are essential for energy planning and grid integration applications. The models used for PV power conversion typically adopt physical, data-driven, or hybrid approaches. Different hybrid techniques, combining elements of physical and data-driven methods, have been effectively developed for PV power estimation and forecasting, leveraging the strengths of both native methods. The data-driven approach allows models to account for unmodeled uncertainties and nonlinearities that purely physical models cannot capture. Meanwhile, the guidance of scientific theory makes the models less dependent on data, thereby improving generalization, interpretability, and accuracy. This review paper provides a comprehensive overview of hybrid methodologies for PV power estimation and forecasting. The main available hybridization methods are summarized and discussed under a novel methodological classification into three primary categories: physics-informed machine learning models, optimized physical models, and physics-guided models. Furthermore, these hybrid models are compared in terms of methodology, applications, and elucidating the merits and demerits of different techniques. By offering insights into these hybrid models, this review lays a foundation for researchers in this field and contributes to the advancement of PV power estimation and forecasting methodologies. • Hybrid models outperform traditional physics and data-driven methods by merging them. • Physics rules prevent illogical prediction and overfit, and improve generalization. • Data-driven methods offer flexibility, addressing unmodeled/seasonal phenomenon. • Hybrid approaches span collaborative to physics- or data-driven-centric methods. • Current models exploit more ML than physics techniques, and fair benchmark is crucial.

Topics & Concepts