A lightweight model with spatial–temporal correlation for cellular traffic prediction in Internet of Things

Accurate cellular traffic prediction becomes more and more critical for efficient network resource management in the Internet of Things (IoT). However, high-accuracy prediction results are usually accompanied by high computational capacity requirements. Although many lightweight neural network models have been proposed, some lightweight mechanisms will easily destroy the features of the raw data. Not all lightweight mechanisms are suitable for network traffic prediction. Therefore, this study proposes and optimizes an input data conversion method to extract the features of spatio-temporal dependencies based on convolutional neural network (CNN) architecture. In addition, we also propose a lightweight neural network model to reduce the computational cost for cellular traffic prediction problem and use mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) to evaluate the prediction accuracy. The experimental results show that the proposed model is better than CNN, ConvLstm, and Densenet as well as can greatly reduce the parameters of the neural network while maintaining prediction accuracy.