Research And Implementation Of Deep Learning Based Grid Data Anomaly Detection Method

Grid data anomaly detection is an important tool to ensure safe operation of power systems and improve power quality.Traditional anomaly detection methods have some limitations,such as sensitive to data distribution and noise and difficult to handle high and complex data.In order to overcome these problems,this paper proposes an approach for anomaly detection in multivariate time series data from sensor networks under power grid systems.The proposed approach considers the strong time-series and different periodicity of the data,and uses a recurrent neural network self-encoder reconstruction error model,combined with multi-scale analysis and attention mechanism,to improve detection performance.To address the challenges of incomplete data labels and severe sample imbalance,the paper employs integrated learning techniques to build a cleaning network that reduces the impact of outliers on classifier accuracy.Specifically,this paper accomplishes the following work.First,for the characteristics of high dimensionality,multivariate and nonlinearity of grid data,this paper proposes a self-encoder-based anomaly detection algorithm for multivariate time-series data to discriminate anomalies using reconstruction errors.In this paper,the Seq2Seq model is constructed using a bidirectional long and short-term memory network to capture the temporal characteristics of the data,and an attention mechanism is used to prevent information forgetting,thus maintaining good reconstruction performance on long sequences.In addition,this paper also considers that different sensors have different period characteristics,so multi-window training is used instead of single window to accommodate anomalies at different time scales.In terms of threshold selection,this paper dynamically determines thresholds based on extreme value theory to adapt to data with different distribution types and achieve a balance between accuracy rate and recall rate,while reducing the false detection rate and leakage rate.Second,to address the problems of incomplete labels and data imbalance,this paper proposes a sample cleaning and anomaly detection strategy based on the degree of sample difficulty.Abnormal scarcity causes data imbalance problems.Therefore,in this paper,positive and negative samples are constructed manually by clustering,and feature extraction networks and classification networks based on spatially selected pass units are used to distinguish between positive and negative samples and treat anomalies as one of the hard-to-score samples.In addition,anomalies in the training data affect the final result,to address this problem this paper implements anomaly sample cleaning based on Gaussian mixture model and AdaBoost method to enhance the robustness of the overall method.The experimental results show that the accuracy rate of the model using the cleaning network is improved by 3.76%and the recall rate is improved by 3.75%compared with that before cleaning.Finally,in order to validate the proposed model approach in grid scenarios,an anomaly detection and visualization system is built based on a smart grid platform and validated using two grid datasets in different scenarios.The system visualizes the detection results and model indicators by means of graphs and charts,which facilitates further analysis and judgment by researchers and improves grid system security,and has practical application value.