Cellular Traffic Prediction Models Using Convolutional Long Short-Term Memory

Abstract

Precise cellular traffic modeling and prediction is essential to future big data-based cellular network management for providing autonomic control and user-satisfied stable mobile services. However, the traditional methods have difficulty learning the complex hidden patterns of the users’ traffic data from cross-domains because of their shallow learning characteristics. Deep learning (DL)-based methods could somewhat identify these hidden patterns by learning the underlying spatial and temporal features and their dependencies. Yet, they too have constraints in handling the noisy and sparse data, reducing the prediction accuracy with increased computation time and associated storage costs. Therefore, this paper presents an intelligent cellular traffic prediction model (ICTPM) using two improved deep learning algorithms to tackle the negative impacts of noisy and sparse traffic datasets. Firstly, the Enhanced Stacked Denoising Auto-Encoder (ESDAE) is introduced to eliminate the noise in the traffic data by an adaptive Morlet wavelet transform. Secondly, Multi-dimensional Spatiotemporal Sparse-representation Convolutional Long Short-Term Memory (MDSTS-CLSTM) is used to learn the hidden patterns by extracting the spatial-temporal dependencies and predict the cellular usage in the presence of data sparsity problem. This MDSTS-CLSTM is developed by combining the Long Short-Term Memory (LSTM) with the Convolutional Neural Networks (CNN) and improvising the multi-dimensional feature learning, spatial-temporal analysis, and sparse representation properties of the hybrid DL algorithm. Evaluated over real-world cellular traffic cross-domain datasets from Telecom Italia and Open-CellID, the proposed ICTPM outperforms the state-of-the-art methods with 5-10% better performance enhancements.