Evaluating Performance and Predicting Soiling Loss at a 1.1 MW Solar PV Plant Using Enhanced Physics-Based and Machine Learning Models

Abstract

This dissertation presents a comprehensive study of a 1.1 MW grid-connected solar photovoltaic (PV) power plant at the University of Louisiana at Lafayette, focusing on evaluating different PV technologies and optimizing soiling loss predictions in the challenging Louisiana climate. The solar PV plant, one of the largest in the state, features 4,142 modules utilizing mono-crystalline silicon, poly-crystalline silicon, and copper indium gallium selenide (CIGS) technologies. Over one year, from September 2019 to August 2020, performance metrics such as energy yield, performance ratio, capacity factor, and levelized cost of energy were assessed and compared against simulated results from the System Advisor Model (SAM) and PVsyst. CIGS technology demonstrated superior performance, suggesting a strategic focus for future installations in similar climates.

In the second part of the dissertation, the DustIQ device is utilized to monitor soiling at the outdoor testing facility. Through adaptations and optimizations of well-known physics-based models (Kimber and HSU), enhanced versions were developed to more accurately reflect soiling recovery post-rainfall, achieving significant improvements in prediction accuracy with root mean squared error reductions and mean absolute percentage errors of less than 1%.

Finally, the effectiveness of machine learning techniques for predicting soiling losses was evaluated, comparing these methods to traditional models. Advanced algorithms such as Random Forest, Support Vector Regression, Multilayer Perceptron, Long Short-Term Memory, and Gated Recurrent Unit were employed, focusing on three feature sets that incorporated lagged environmental data. These machine learning approaches, particularly Random Forest and SVR, excelled in utilizing 20 and 30-day lagged feature sets of enhanced features, highlighting the potential of integrating complex temporal data to enhance predictive models.