Hybrid physical-machine learning estimation of photovoltaic soiling losses from meteorological reanalysis data in Africa and Pacific islands

Soiling losses are a major source of uncertainty in photovoltaic energy yield, particularly in regions exposed to high aerosol concentrations and intermittent precipitation. These losses are strongly modulated by meteorological conditions making their quantification a key challenge in energy meteorology. Estimating soiling losses is challenging due to complex interactions between deposition processes, cleaning events such as rain, wind-driven dust transport, and proximity to local aerosol sources.

Soiling losses can be derived from irradiance measurements using paired modules subjected to differing cleaning schedules. In this work, one year of measurements from monitoring networks in West Africa and Pacific islands are used. Meteorological drivers are extracted from ECMWF reanalysis products, including precipitation and particulate matter.

We evaluate two widely used semi-physical soiling models as benchmark, HSU and Kimber, and develop a hybrid physical-machine learning framework that integrates a physics-based empirical model with XGBoost trained on meteorological reanalysis data. Model performance is assessed using temporal cross-validation across all stations and a leave-one-out approach to evaluate spatial portability, followed by an application to a real-world photovoltaic case study in Mali.

The hybrid model significantly improves soiling losses estimation compared to semi-physical benchmarks across most sites. However, its performance decreases in environments characterised by persistently low soiling, highlighting the importance of physical constraints for extrapolation beyond the training domain.

These results highlight the potential and limitations of hybrid physical-machine learning approaches for meteorology-driven soiling assessment, supporting maintenance decisions and photovoltaic energy yield optimization.