Research On Base Station Traffic Prediction Method Based On Machine Learning

The current construction scale of wireless communication base stations is constantly expanding,and wireless communication technologies are being updated.Domestic communication equipment manufacturers and operators are committed to the construction of wireless communication networks at home and abroad.4G wireless communication networks face problems such as complex network structures,diversified services,and rapid growth in user requirements.For areas where the demand for wireless networks is growing rapidly,capacity needs to be expanded as soon as possible to ensure service quality.The traffic usage and the load of each base station can most directly and effectively reflect the user requirements for the wireless network.Equipment manufacturers usually decide whether to expand capacity by predicting the amount of traffic in the base station in the future.For some newly added base stations,the historical running time is short,so how to use a short traffic time series to predict the next long period of traffic becomes a problem that communication operators need to solve urgently.This paper studies the method of base station cell traffic prediction based on machine learning.The research on the traffic prediction method in this paper is based on a batch of base station data provided by ZTE Corporation.The research goal is to predict the base station's traffic in the next month while using historical base station data for up to six months.This is a long-term prediction problem compared to existing traffic prediction methods and time series prediction methods.Three different models are used for traffic prediction:Seasonal Autoregressive Integrated Moving Average model(SARIMA),BP-Adaboost model,and Wide and Deep model.The SARIMA model is a time series prediction model based on statistical hypothesis testing.This paper proposes a method for determining the order of the model based on grid search to determine the order of the model,which has been verified on the long-term prediction problem of network traffic.The BP-Adaboost model uses BP neural network as a weak predictor and combines Adaboost to form a strong predictor to achieve the best of all.Compared with the traditional ARIMA model and BP neural network prediction method,the prediction accuracy is improved.The Wide and Deep model combines artificially selected features with time series features automatically learned through deep neural networks,and the prediction accuracy is better than traditional neural network prediction methods.Based on the above methods,designed and implemented a base station traffic prediction platform.This platform is provided to the operators of base stations and includes functions such as data visualization analysis,offline model training,online traffic prediction,and data management.