Machine Learning models for the estimation of the production of large utility-scale photovoltaic plants

Abstract

Photovoltaic (PV) energy development has increased in the last years mainly based on large utility-scale plants. These plants are characterised by a huge number of panels connected to high-power inverters occupying a large land area. An accurate estimation of the power production of the PV plants is needed for failure detection, identifying production deviations, and the integration of the plants into the power grid. Various studies have used Machine Learning estimation techniques developed on very small PV plants. This paper deals with large utility-scale plants and uses all the available information to represent the non-uniform radiation over the whole studied solar field. Variables measured in up to four meteorological stations and distributed across the plant are used. Three PV plants with 1, 2 and 4 meteorological stations have been used to develop Machine Learning models. The hyperparameters were systematically optimised, demonstrating the improvements by comparing with a simple model based on Multiple Linear Regression. The best results were obtained with the Random Forest technique for the three PV plants, providing a RMS error value ranging from 1.9% to 5.4%. The final models were compared with those found in the literature for tiny PV plants showing in general much better performance.