Clustering Sensitivity Analysis for Gaussian Process Regression Based Solar Output Forecast

Power system operations are becoming more challenging with the increasing penetration of renewable-based re- sources such as photovoltaic (PV) generation. In this regard, obtaining accurate solar power output forecasts allows a deepening penetration of renewable-based resources in a secure and reliable way. In this paper, we propose a probabilistic framework to predict short-term PV output taking into account the uncertainty of weather data as well as the variability of PV output over time. To this end, we use datasets comprising of meteorological weather data such as temperature, irradiance, zenith, and azimuth and solar power output. We cluster these data in categories and train a Matérn 5/2 Gaussian Process Regression model for each cluster. More specifically, we cluster the data into one to eight different partitions by making use of the k-means algorithm. In order to identify the optimal number of clusters we use the Elbow and Gap methods. We compare the results obtained for the different number of clusters with the (i) 5-fold cross-validation; and (ii) holding out 30 representative days as test data. The results showed that the optimal number of clusters is four, since in comparison to higher number of clusters the increase in the forecast error was marginal.