A Novel Anomaly Forecasting in Time-Series Data: Feedback Connection between Forecasting and Detecting Algorithms with Applications to Power Systems

Abstract

This article is concerned with developing a novel structure of machine learning-based anomaly forecasting, by which both forecasting the future states and detecting the anomalies in these states can be achieved at the same time. The main idea of this article is to introduce a feedback connection to combine several algorithms with respect to the forecasting and the detecting in a single algorithm. More precisely, Xgboost and long short-term memory are used for forecastor and one-class support vector machine and robust random cut forest are used for detector. Combining those 2 × 2 schemes leads to the overall four algorithms, and future anomalies can be detected before they occur. The effectiveness of the proposed algorithms is verified through some comparative simulations of an IEEE 3-bus system with various faults. More interestingly, the detecting accuracies obtained through the two schemes of taking robust random cut forest are shown to be improved by 10% than those of employing the one-class support vector machine. For the forecasting part, Xgboost is regarded as involving the fastest prediction speed for online implementations, and thus the combination of Xgboost and robust random cut forest can be the most suitable choice for anomaly forecasting for power system fault events.