In this paper, a data-driven approach is explored to evaluate the impact of weather conditions on the reliability of urban distribution system. The severity of power outages is divided into two levels according to the number of days with outages in one week. The actual outage records from the local utility are used for the analysis in this study. First, the difference of weather conditions under the two outage levels are intuitively described with the Kernel Density Estimation (KDE). Then, an extreme gradient boosting algorithm is applied to build a classification model for evaluating the outage levels of the local distribution system under given weather conditions. The importance of weather features on the outage level is discussed with the built model. Finally, the performance of the proposed data-driven model is assessed with the Receiver Operating Characteristic (ROC) curve.
Keywords kernel density estimation, distribution system, outage, ROC