Base station traffic prediction using XGBoost-LSTM with feature enhancement

With the development of information technology, base station traffic prediction is becoming more and more vital in allocating resource, and finally improving terminal users' quality of experience. Temporal and periodic characteristics are important for handling the issue of efficient and accurate traffic prediction. Considering these characteristics, this study proposes base station traffic prediction using extreme gradient boosting-long-short-term memory (XGBoost-LSTM) with feature enhancement. First, the collected dataset is preprocessed, especially realising missing values filling. Then, to mine the tidal property, feature engineering is performed, which contains feature creation and feature selection. More importantly, the variance contribution of the indicators is calculated based on the factor analysis. The variance contribution of the indicators is used to determine the weights of each selected features. Finally, the XGBoost-LSTM model is adopted to predict the traffic of base stations. By observing the predicted values, the authors find that the simple combination of XGBoost and LSTM can achieve great improvement. Experimental results show that the proposed scheme can get much better performance when compared with competing algorithms.