Neural network-based power forecasting framework for solar PV plants under high soiling and degradation conditions

Soiling and degradation are two major challenges that reduce the performance and forecasting accuracy of solar photovoltaic (PV) systems. While degradation effects have been studied, the combined influence of soiling and degradation on power forecasting remains largely unexplored. Soiling is particularly difficult to model because it has no direct correlation with irradiance or temperature and varies with cleaning schedules. As a result, power output can differ under identical environmental conditions depending on the soiling ratio (SR), leading to increasing forecasting errors. This study presents a transformer-based forecasting framework that explicitly corrects for both soiling and degradation. The proposed method first removes soiling effects from historical data, enabling the model to learn clean generation patterns. Forecasted power is then adjusted in real time using the SR. To handle degradation, the model employs periodic retraining (PR), maintaining accuracy as system capacity decreases. Tested on a 504 kWp rooftop PV plant, the framework achieved a mean relative error of 3.57 %, RMSE of 30 kW, and MAPE of 9.8 % over one year. Long-term analysis shows RMSE reduction of 32 % and MAPE reduction of 69 % over 20 years, confirming its robustness for reliable PV forecasting.