Quantifying Soiling Losses Without Sensors
Abstract
Soiling is defined as the accumulation of dust, pollen, and other particulates on photovoltaic (PV) modules, and it significantly impairs energy generation by reducing light transmittance. Accurately quantifying these losses is challenging due to the dynamic interplay of environmental factors and the high cost of dedicated, permanent monitoring stations. This study proposes a novel genetic algorithm (GA) optimized approach for estimating photovoltaic soiling losses, designed to reduce reliance on physical sensors. The model predicts soiling ratios using only two widely available input variables: hourly rainfall and PM₁₀ particulate matter concentration. The use of a GA to optimize both the parameters and the logic of rainfall-dependent state transitions ensures high predictive accuracy without requiring complex, hard-to-measure inputs. Training was conducted on a 3-year and 5-month dataset from a 30 MW-scale PV plant, with an extra year of data being used for validation. The proposed model was benchmarked against two established models from the PVLib package: the Hsu model, used with its original parameters, and the Kimber model, optimized using the same GA for a fair comparison. The algorithm demonstrated a superior coefficient of determination (R² = 0.72), representing a performance improvement of 12% over the optimized Kimber model (R²=0.59) and 32% over the Hsu model (R²=0.50). By achieving high accuracy with minimal data requirements, this solution directly addresses industry priorities for low-cost, high-reliability soiling estimation.
Keywords
Soiling, Genetic Algorithm, O&M Cost Reduction, Sensorless Monitoring, Data-driven modeling