Multi-criteria Modeling of Photovoltaic Module Soiling Losses in a Grid-Connected Power Plant within the Sudano-Sahelian Zone (Burkina Faso) Using Artificial Neural Networks and Linear Regression

Abstract

Soiling of photovoltaic modules is a major factor affecting the performance of solar power plants, particularly in dry and dust-prone environments. This study developed and evaluated predictive models for estimating the soiling losses (SL) of photovoltaic modules at the Zagtouli grid-connected solar power plant in Burkina Faso. Field-measured environmental and meteorological variables were used to examine relationships between local operating conditions and module soiling. After variable screening, five predictors were retained for modelling: exposure duration, ambient temperature, wind speed, relative humidity and PM10 concentration. Two modelling approaches were compared: multiple linear regression (MLR) and an artificial neural network (ANN). The MLR model provided a useful baseline and explained approximately 94% of the variability in the soiling losses. However, its linear structure limited its ability to represent more complex interactions among climatic and particulate variables. The ANN model, developed using a 5-15-1 multilayer perceptron architecture, showed stronger predictive performance. The optimised model achieved coefficients of determination of 98.42% during training and 98.45% during validation, with low error values for the mean square error (MSE = 0.015), root mean square error (RMSE = 0.122) and mean absolute error (MAE = 0.086). These findings indicate that exposure duration was the dominant predictor, while the nonlinear modelling approach better represented the combined effects of meteorological and particulate factors. The proposed model provides a site-specific basis for understanding photovoltaic soiling behaviour in a Sudano-Sahelian environment and may support more informed maintenance planning for grid-connected photovoltaic plants operating under similar climatic conditions.