Efficient Lightning Outage Estimation of Distribution Lines via Machine Learning

This paper presents a novel approach for estimating the lightning performance of overhead distribution lines by offering an alternative to conventional statistical techniques. Precisely, the presented technique combines a machine learning algorithm with Monte Carlo method. The approach is structured into three main phases. First, a pre-processing phase generates the training set, obtained collecting simulations of indirect lightning events. Next, the training set is used to train a machine learning algorithm. Finally, a test set - comprising both direct and indirect events - is created and the performance of the algorithm is evaluated. Flashover is always assumed for direct events, while for indirect events, the trained machine learning algorithm estimates the induced voltage. The outcome is used to determine if a flashover occurs. The performance of the proposed approach is assessed for two test cases, respectively involving a single phase and a three phase system. The benchmark, in both cases, is the identification of the flashover computed with the traditional Monte Carlo method. The results on the test cases demonstrate that the proposed approach can provide similar lightning outages estimation as the traditional method but with greater computational efficiency. According to the results, for single-phase system the percentage deviation between the two approaches is lower than 5% and for three-phase system the percentage deviation is lower than 17%. On average, the proposed approach consumes 0.6% of the simulation time required by the Monte Carlo method.