Outlier Detection Algorithm of Photovoltaic Power via Multivariate Dependence Modeling Based on Vine Copulas

Abstract

Outlier detection is of great significance to ensure reliable and intelligent operation and maintenance of photovoltaic power (PV) plants. The PV operation data of supervisory control and data acquisition (SCADA) suffer from a high proportional of abnormality due to equipment failures and/or human activities. As a remedy, we present a novel outlier detection method, which leverages vine copulas for modeling the multivariate dependence of PV power output and different external environmental conditions. A generalized confidence interval (CI), obtained by quantile computation, is then utilized to give upper and lower thresholds for identifying outliers, which can be further optimized with the best selection of conditions. For algorithmic improvement, we combine the method with manual exclusion of three typical types of outliers in a priori. The effectiveness of the proposed method is verified by extensive case experiments using a mixture of real-world and artificial PV plant running data.