Utilizing Deep Belief Networks for Power System State Estimation and Anomaly Detection

The state estimation and anomaly detection of power systems highly rely on accurate and comprehensive real-time data. However, many power systems face data quality issues such as sensor failures, data loss, or noise pollution. Therefore, this study proposes a power system fault detection and recovery time prediction model based on deep belief network (DBN). Firstly, various fault feature extraction methods have been proposed for power system operation data, including electrical quantity features (such as voltage and current), time-domain features (such as power fluctuations, frequency fluctuations, etc.), and normalization methods have been used to preprocess the raw data to eliminate outliers, missing values, and redundant data. Then, deep belief networks are used for pre training of fault features, and unsupervised learning is used to train each layer of restricted Boltzmann machines (RBM) to extract potential features of power system data. This process optimizes the weights by comparing the divergence (CD) algorithm, and further trains the backpropagation algorithm during the fine-tuning stage to improve the accuracy of fault detection. In terms of recovery time prediction, by establishing the relationship between recovery time and fault type, combined with the feature extraction ability of DBN, the recovery time of different fault types can be predicted. The experimental results show that the model exhibits good performance in fault detection and recovery time prediction, which can effectively improve the stability and fault response speed of the power system.