Photovoltaic (PV) system performance is linked to climatic conditions in which the system operates. This leads to the Köppen-Geiger-Photovoltaic (KGPV) climate classification. KGPV is created by overlaying four irradiation levels with the commonly used Köppen-Geiger climate zones. Potential drawbacks of this approach are that the climate features are not considered in a combined manner in the sorting process and that the KGPV zones inherent a dependence on precipitation. We propose a machine-learning approach to address this deficiencies and improve PV climate classification. First, supervised learning is used to evaluate the correlation between climate features and a PV system's specific energy yield. We find that the inclusion of the darkest and brightest irradiation months as well as UV irradiation improves accuracy, while wind speed, relative humidity, precipitation and annual mean daily temperature difference have little impact on accuracy. Subsequently, k-means clustering combined with comprehensive qualitative analysis, identifies a PV classification based on seven climate features and 21 clusters. A mountainous climate characterized by moderate to low temperature and high irradiation is uncovered compared to KGPV. Moreover, this new PV climate classification reduces the sum of squared errors by 58 % compared to KGPV clearly signifying a more accurate PV climate classification approach.