Anomaly Detection and Trend Prediction in Intelligent Operations Based on Prophet and S-ESD

Abstract

Anomaly detection and trend prediction based on KPI business indicators play a key role in intelligent operation and maintenance in various fields. The types of indicators analyzed vary greatly for different scenarios of operation and maintenance, but all have time-series characteristics. In this paper, three core indicators are selected based on time series to build anomaly detection and trend prediction models. After preprocessing the index data, a statistical-based S-ESD time series anomaly detection method is firstly implemented. Anomaly detection is performed on three core indexes, and the detected anomalies are extracted and corrected. Use the revised KPI indicator data to further build a Prophet trend prediction model based on decomposable (trend + season + holiday), select MAPE as the evaluation index, perform model training on the indicator data, predict the next three days by step size, and get the fitting Curve and trend for the next three days. After parameter tuning, the MAPE of the Prophet prediction model is less than 0.1, the running time is shorter, and the trend prediction accuracy is higher, which can be practically applied to the field of intelligent operation and maintenance.