Effect of Climate on Photovoltaic Yield Prediction Using Machine Learning Models

Abstract

and clouds create fluctuations in the power generation, which result in voltage unbalance, voltage rise and voltage flickers in networks with high PV presence. [3] Forecasting technologies can help grid operators with the scheduling and dispatching of this renewable energy source more effectively. Due to the stochastic nature of PV power generation, machine learning (ML) approaches have gained popularity in forecasting tasks. [4] The main characteristic of ML algorithms is that the coefficients of the model are obtained automatically by using training data. [5] There is a huge amount of algorithms inside this family, whose complexity ranges from linear regression techniques to deep learning neural networks.Despite abundant literature, most studies consider <5 PV systems that are generally located in the same areas. [6] Most researchers study systems in Europe, the US, Australia, or China, which are population-dense areas, but receive relatively small solar intensity (except for Australia). Few studies have considered PV systems subjected to diverse meteorological conditions in order to study the effect that climate has on the performance of ML models.Pasion et al. forecasted the PV power of 12 northern-hemisphere locations subjected to seven different climate regions. [7] The selected algorithm best predicted the data in a hot-dry climate region, and a mixed-humid climate region had the secondbest model performance, with coefficient of determination (R 2 ) scores of 0.968 and 0.962, respectively. In contrast, the model performance for the tropical rainforest site was the poorest with an R 2 score of 0.908. Do et al. forecasted two PV systems, one in Guadeloupe (North America, tropical climate) and the other in Lille (Europe, mild temperate climate). [8] They discovered that the system in Lille required a longer training duration but achieved a lower normalized root mean squared error of 10.69 % compared to the one in Guadeloupe (error of 11.97 %). In ref. [9], two systems located in tropical (Singapore) and mild temperate (Australia) climates were also considered for a probabilistic forecasting model. As opposed to the previous study, they reported worse metrics for the system located in the mild temperate climate, due to the highly variable Australian weather and thunderstorms in the summer season. Zhang et al. considered three distant PV systems located in the USA, Denmark, and Italy. [10] Because of the unique weather and climate characteristics of each location, the modeling parameters and best features of the optimum ML model differed between sites. Finally,