Comparing Time Series Classification And Forecasting To Automatically Detect Distributed Generation

This paper introduces tools for the automatic detection of "hidden" behind-the-meter solar generation in case where there is no monitoring or connection agreement contract with the system operator. The objective is to reach the highest precision while discriminating the nodes with and without solar generation. The proposed methods are based on exogeneous information (smart meter and temperature data) and artificial intelligence techniques consisting of neural networks as well as analytical classification algorithms. A wide range of models differing in size, architecture and number of parameters has been investigated, and the best performing ones are presented in the article. The first method involves time series classification (TSC), and the second involves time series forecasting (TSF). Open-access consumption data were used for the training of the neural networks. The implemented solutions were tested across all the nodes of the simulated electrical grid and the sensitivity of the tools was analyzed with regard to the level of PV penetration. One of the proposed tools is able to detect up to 100% of new PV installation, depending on the exogenous conditions.