A Scalable Deep Convolutional LSTM Neural Network for Large-Scale Urban Traffic Flow Prediction using Recurrence Plots

Abstract

Short-term traffic prediction is critical for urban traffic congestion control and management. The past two decades have seen a rapid increase in short-term traffic prediction models. However, the majority of traffic prediction models focus on junction or link traffic parameter prediction, rather than network-wide prediction. For effective urban traffic congestion management and future planning, network-wide traffic parameter prediction becomes critical. This paper, therefore, proposes a scalable deep learning framework that learns traffic flow parameters as images and predicts multi-step traffic flow. The input traffic network time series is converted to a series of recurrence plots. A deep 2-dimensional Convolutional Long Short-Term Memory (ConvLSTM) architecture is applied to perform representation and sequential learning. We evaluated the performance of our proposed model using real-world road traffic network data obtained from sensor-collected data in California, USA. The performance of our predictive approach is benchmarked against state-of-the-art deep learning traffic prediction models. The experimental results highlight the potential of the model in handling large-scale urban traffic data and substantiate the value of the approach when applied to large-scale urban traffic flow prediction.