Anomaly Detection for Grid-Connected Photovoltaic Array via Graph Attention Mechanism

Abnormal conditions of field photovoltaic (PV) array such as open-circuit, short-circuit, and partial shading are embedded in DC side voltage/current curves. Besides, meteorological factors as solar radiation, wind speed, and ambient temperature can also influence fault behaviors. To realize abnormal identify of PV array under environmental interference, this paper presents a graph attention network (GAT) based fault detection algorithm for PV array. Fault simulation experiments are conducted on grid-connected PV array, and the voltage/current curves of the PV DC side under different states (normal, open-circuit, short-circuit, and partial shading) were collected to build dataset. One-dimensional convolution and two parallel graph attention layers are adopted to extract temporal and dimensional features of the voltage/current series. A gated recurrent unit (GRU) is employed to capture the long-term dependencies of the time-series data. Fully connected (FC) layers and variational auto-encoder (VAE) are combined optimized for detecting and locating the PV abnormal events. Model performance are compared with Robust Anomaly Detection (OmniAnomaly), Transformer Networks for Anomaly Detection (TranAD), and Long Short-Term Memory (LSTM), result show that the proposed grid-connected PV array fault detection model achieves an accuracy of 96.8% on the test dataset, providing an effective method for fault diagnosis of grid-connected PV systems under different meteorological conditions.