Forecasting fine-grained city-scale cellular traffic with sparse crowdsourced measurements

Abstract

With the rapid development of wireless technology and the edge computing applications, an increasing number of 4G/5G infrastructure are densely deployed to meet the booming cellular traffic demands. Monitoring and forecasting urban cellular traffic is fundamental for urban planning, network resources allocation, traffic engineering, etc. In this paper, we address the crowdsourcing-based urban cellular traffic prediction problem, i.e., to predict the city-scale fine-grained cellular traffic patterns based on partial user-generated measurements. We propose a novel deep generative adversarial network (GAN) model called CrowdGAN to solve the problem. Specifically, CrowdGAN employs a convolutional Long Short-Term Memory (LSTM) network to extract spatio-temporal features from sparse traffic maps, and adopts a novel design of co-training a generator and a discriminator under the supervision of an accuracy assurance network to generate a high-resolution cellular traffic map for prediction. We implement the proposed CrowdGAN in TensorFlow and evaluate its performance using two real-world cellular traffic datasets. Extensive experiments show that CrowdGAN significantly outperforms the baselines on a variety of performance metrics, and achieves at least 47% reduction in root-mean-squared error compared to the state-of-the-art.