A BiLSTM-Based Method for Detecting Time Series Data Anomalies in Power IoT Sensing Terminals

Abstract

With the development of power IoT, all kinds of sensing devices in the sensing layer have increased, and a large amount of time-series data is collected every moment. However, the data will inevitably be abnormal due to the external environment or equipment, etc. To ensure that the anomalous data collected by the sensing terminal in the power IoT can be detected, a BiLSTM-based anomaly detection model for time series data of the sensing terminal in the power IoT is proposed. Firstly, the Bi-LSTM can capture bi-directional timing information to build a prediction model. Secondly, multiple thresholds are set up, and the predicted value and the data collected by the sensing terminal are calculated as residuals and then compared with multiple thresholds, and the majority result is taken to determine whether the data is abnormal or not, avoiding the misjudgment of a single threshold.