PV monitoring and fault detection: Evaluation of machine learning for prediction of PV soiling in Northern Cape, South-Africa

Abstract

Renewable energy sources, and thus PV are experiencing exponential growth due to most current energy production still relies on fossil fuels, and energy demands are steadily increasing. If the performance of PV could be increased, the result will be more production per installation. One significant performance loss for PV is soiling on the modules. Research has been done to statistically indicate optimal cleaning intervals. Some attempts using conventional methods to predict soiling have been conducted as well, suggesting environmental features like wind and humidity are relevant factors for predicting soiling. With the increase in popularity and availability of machine learning – is it possible to use machine learning to predict soiling? If it is possible, this could lead to quick and precise implementation of algorithms to predict instantaneous losses due to soiling. Which would further lead to an exact optimal cleaning schedule, reducing both costs and losses. With a test site in close proximity to a solar plant in Kalkbult, South-Africa, and the machine learning approach called artificial neural networks; this thesis tried to identify if this relationship exists, and if so, to what extent. The results were encouraging, but not conclusive. There was indications the two features average humidity and maximum wind speed could relate to a daily change in performance with R2 scores around 0.1–0.28. However, more accurate data and designated experiments are needed to reduce uncertainties for a more conclusive remark.