Anomaly Detection and Prediction of Base Station KPI
Qingyi Li, Jingyi Zhu, Yongjia Yu, Wen Hai Shi · 2022
This paper uses the periodic time series analysis method to study the short-term prediction of time series business data in specific operation and maintenance scenarios, identify and correct abnormal values of historical data series of the performance index of the operator's base station KPI, and realize the short-term multi-step prediction with high accuracy. This study innovates the detection method of outliers using the Nonhomogeneous Poisson Process (NHPP) and multilayer perceptron (MLP), and completes the multi-step time series prediction utilizing three periodic prediction methods, AUTO-ARIMA, Prophet and Trendy and Seasonal Linear Model (TSLM), based on the correction of outliers. The competitive side of the National University Big Data Challenge acknowledged the forecast result, and it was awarded the fifth place of the first prize in the competition. In intelligent operation and maintenance settings, this prediction model can serve as a guide for index prediction.