Residual-Convolutional Network Based Deep Learning Approach for Anomaly Detection in Photovoltaic Substation Systems

With the development of technologies for new distribution power grid, the integration of distributed sources has increased significantly, posing transformation demands on operation of photovoltaic (PV) substations. Currently, PV substations commonly face insufficient monitoring facilities, which bring difficult to meet the requirements for refined control. Therefore, this paper proposes a method for identifying data anomalies in PV substations, caused by factors such as measurement errors, communication failures, and issues inherent to PV equipment. Initially, a convolutional neural network framework is employed, proposing a residual fusion network structure that leverages deep learning of historical data features to construct an anomaly detection model. Subsequently, the model processes through convolutional, pooling, and batch normalization layers to calculate the normalized index for target conditions, building a classifier for high-precision identification of substation anomaly data. Finally, the model's ability to accurately identify data anomalies and distinguish between measurement, communication, or PV equipment issues is validated using simulated substation measurement data. It is demonstrated by experimental results, that the aforementioned improved residual-convolutional architecture offers superior detection accuracy and better noise resistance compared to traditional deep learning methods.