A Hybrid Algorithm for Changepoint Aware Long-Term Seasonality Detection of Mobile Network Base Stations

Abstract

Automated capacity planning for mobile networks requires long-term forecasting of traffic demand by using historical patterns. To decide the correct time of investment and correct capacity expansion size or to improve the accuracy of forecasting algorithms with exogenous features, both seasonal decomposition, and seasonal period identification improves decision accuracy. We design a hybrid algorithm to calculate these features on live network data with improved accuracy which uses piecewise Seasonality Trend Decomposition with Loess (STL) decomposition and Prophet library’s regression with Laplace prior under the hood. Combining both methods with the awareness of their weak and strong parts and leveraging overall output with changepoint and similarity analysis help us to improve our accuracy around 18.6% comparing the average of single usage of these methods. We also provide and present some special cases that increase problem complexity and decrease decomposition accuracy.