Detecting Behind-the-Meter PV Installation Using Convolutional Neural Networks

Abstract

Increased penetration of behind-the-meter (BTM) PV installations can cause numerous challenges in planning and operation of distribution systems. Utilities must accurately record the installed PVs in their territory and keep their PV database updated. However, many utilities do not have enough visibility on the actual installed PVs due the growing number of unauthorized PV installations as well as the complexity of data tracking and updating the databases even for authorized PVs. In this paper, a data-driven classification method is proposed for detecting BTM PV installation using convolutional neural networks and synthetic net load profiles generated from AMI data. The network is trained and tested on 50 folds of the dataset and the testing classification accuracy per each fold is calculated. Results show that the median of per-fold testing accuracies is 98.9%. In terms of average error, only 0.7% of the customers with PV are not detected. This is significantly less than the 6% error in the next best method. The impact of training data parameters, such as the size of dataset and label errors on the accuracy and computational time of the method is also studied and characterized. Using only the available AMI data, the proposed method can help utilities accurately monitor BTM PV systems and keep their databases updated and thus avoid the costs of operation and planning errors.