Machine Learning Driven PV-Climate Classification

Abstract

Technological advances, cost reduction, depletion of fossil fuels, environmental concerns, and growing energy demand are expanding photovoltaic solar energy (PV) in more latitudes and locations. A simple and effective procedure to assess the PV potential of a particular region is to analyse its climatic conditions. In general, climatic studies use the K¨oppen-Geiger (KG) climate classification as a reference. However, KG is solely based on temperature and precipitation, resulting in an unsatisfactory scheme for analyses in the PV field, since the most important variable, solar irradiation, is not considered. Thus, in 2019 Ascencio-V´asquez et al. developed a new worldwide classification based on temperature, precipitation, and solar irradiation: the K¨oppen-Geiger-Photovoltaic (KGPV) climate classification. Even though KGPV is a good improvement, it just consists of a simplified version of the KG groups subdivided into four levels of irradiation: low, medium, high, and very high. Hence, the climate parameters are not considered in a combined manner in the sorting process.

In this project, a new worldwide climate classification directly applicable to PV has been developed. Machine Learning proved to be a convenient tool to achieve this objective. First, supervised learning served to identify and assess the climate variables more correlated to the specific energy yield. More specifically, a Linear Regression model was implemented. Subsequently, these variables were used to create the classification by applying k-means, a clustering algorithm. The classification was optimised following a comprehensive qualitative analysis, resulting in a scheme based on seven climate variables and 20 clusters. By contrast, KGPV considers five variables. Even though it contemplates 24 groups at first, half of them are neglected based on a land-surface ratio and population density criterion, resulting in a classification based on 12 clusters. Hence, the methodology proposed in this work enables identifying new relevant regions. Moreover, “Machine Learning driven PV-climate classification” presents a satisfactory correlation with the specific energy yield, except for very low values, where the correlation is minor.

Lastly, the relationship between climate and degradation rate was explored. The complexity and non-linear behaviour of degradation demand an alternative approach. Random Forests was proposed, but it showed poor performance. It is necessary to be able to predict non-linearities and, at the same time, keep a logical mathematical relation between the supervised and clustering algorithms. In this regard, Multivariate Adaptive Regression Spline (MARS) might be a promising option.