Siamese Convolutional Neural Network-Based Anomaly Detection for Distributed PV Inverter

Monitoring distributed photovoltaic (PV) inverter operation statuses, especially anomaly/fault detection, is of great significance for the reliability and durability of PV systems. Considering the rare real-world fault data collected from distributed PV inverters, it is still an outstanding problem how to perform accurate anomaly detection for distributed PV inverters. To this end, this paper proposes a Semi-Supervised Learning Enhanced Siamese Convolutional Neural Network (SSL-SCNN) Framework for solar inverter anomaly detection. Relying on the contrastive structure, the proposed SCNN can effectively solve the small sample problem and alleviate the influence of the class imbalance. The designed SSL mechanism lifts the generalizability of the SCNN model by offering a form of transductive learning to introduce new information from testing samples, which adapts the model to the diverse operational characteristics of industrial solar inverters. The proposed method builds on and is tested on real-world solar inverter operation data and fault information. The numerical results show that compared with commonly-used anomaly detection methods, the SSL-SCNN model exhibits better anomaly detection performance while providing satisfactory generalizing capacity on new samples.