A method to predict solar photovoltaic soiling using artificial neural networks and multiple linear regression models

• Published in: Energy systems (Berlin. Periodical) Vol. 11; no. 4; pp. 981 - 1002
• Main Authors: Chiteka, Kudzanayi, Arora, Rajesh, Sridhara, S. N.
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.11.2020; Springer Nature B.V
• Subjects: Algorithms; Alternative energy sources; Artificial neural networks; Cleaning; Economics and Management; Energy; Energy dissipation; Energy loss; Energy Policy; Energy Systems; Modelling; Neural networks; Operations Research/Decision Theory; Optimization; Original Paper; Parameters; Particulate matter; Performance prediction; Photovoltaic cells; Photovoltaics; Precipitation; Prediction models; Regression analysis; Regression models; Relative humidity; Soils; Solar cells; Solar energy; Wind direction; Wind speed; Feature selection; Soiling; Multiple linear regression; Photovoltaic performance; Artificial neural networks
• ISSN: 1868-3967; 1868-3975
• DOI: 10.1007/s12667-019-00348-w

The adverse effects on performance and reliability of soiling on solar photovoltaics are the major areas of concern in today’s era. Environmental and meteorological solar photovoltaic soiling parameters were investigated for three 100W p PV collectors installed at Harare Institute of Technology, Harare, Zimbabwe. The Boruta algorithm implemented in the random forest technique was used to select the most influential parameters in a given set of parameters used in soiling predictive modelling. Five most important variables which are PM 10 , relative humidity, precipitation, wind speed and wind direction were identified and used in modelling. Two soiling predictive models were developed using Artificial Neural Networks together with Multiple Linear Regression. The five selected most influential soiling variables were used in the two predictive models and the performance of the models was adequate with R adj 2 of 97.91% and 79.69%, respectively, for Artificial Neural Networks and Multiple Linear Regression. Moreover, the Residual Mean Square Error Values for the two models were 1.16% and 4.9% with Mean Absolute Percentage Errors of 6.3% and 10.6%, respectively, for Artificial Neural Networks and Multiple Linear Regression. The measured data indicated a mean daily loss in efficiency η ¯ l of 0.083% and a standard deviation of σ l of 0.00973%. The investigation revealed that soiling prediction is of paramount importance as it give the basis for the determination of mitigation activities. If the energy loss due to soiling is known in advance, cleaning procedures will be planned ahead of time. The energy supply by such a solar photovoltaic system will be known in advance leading to the determination of an alternative energy source to cater for the deficit created by the anticipated energy loss due to soiling and the subsequent cleaning procedure.