Hybrid Data‐Driven Modeling and Prediction of Photovoltaic Soiling Losses: Balancing Accuracy and Simplicity

Photovoltaic (PV) system efficiency is significantly affected by soiling, leading to energy losses and increased operational costs. Within the SERENDI‐PV and CACTUS projects, this study refines soiling loss modeling through two complementary approaches: the stochastic quantifying soiling loss (SQSL) method and a machine learning (ML)‐based predictive model. SQSL quantifies soiling losses using electrical performance data, independent of meteorological inputs, while the ML model forecasts losses using environmental parameters such as temperature, humidity, wind speed, and particulate matter concentration. The SQSL method identifies soiling phases—cleaning periods, stable periods, and soiling periods—by analyzing PV performance trends. Monte Carlo simulations generate probable soiling profiles, assessing uncertainty and informing maintenance strategies. Additionally, SQSL enables classification of soiling accumulation patterns, distinguishing between gradual and rapid soiling events. The ML‐based approach integrates artificial neural networks and regression models to predict soiling losses, applying data preprocessing techniques to enhance accuracy. Regional climate variations and site‐specific soiling characteristics are incorporated to improve predictive performance. Preliminary results validate the robustness of both methodologies. SQSL‐generated profiles align closely with observed soiling trends, and ML models effectively capture seasonal variations. Field studies at a utility‐scale PV plant in Spain further demonstrate the location dependency of soiling patterns. Comparative analysis of SQSL predictions and real soiling measurements shows prediction errors between 1% and 2.6% over an 8‐day forecast period. These findings support the integration of predictive soiling models into PV monitoring platforms, optimizing maintenance strategies, and minimizing energy yield losses.