A review of deep learning-based approaches and use cases for traffic prediction

Traffic prediction is one of the most critical aspects of transportation planning and transportation system management and operations. Over the past decades, researchers and practitioners have explored numerous methodologies to improve the accuracy of traffic prediction models. These include analytical, simulation-based, and data-driven approaches. In recent years, data-driven traffic prediction methods have been gaining more attention due to their benefits in solving complex problems in a much simpler way. All these promising solutions are being supported by ubiquitous traffic sensing technologies that provide high-resolution spatiotemporal data in real time. Yet, traditional data-driven approaches in many cases fail to accurately predict traffic due to the presence of sharp non-linearities caused by transitions among free flow, traffic breakdowns and recovery, and congestion. The application of deep learning methods has created an opportunity to overcome these challenges by allowing the modeling of non-linear behaviors in traffic data. Deep learning methods decode the traffic data into a high-level representation, thus capturing subtle changes in traffic behavior. In addition, it consists of non-linear modules that allow very complex functions to be learned. These methods are more suitable to model sharp discontinuities in traffic data. This chapter presents a review of commonly used deep learning models to solve traffic prediction problems such as feed forward neural networks, long short-term memory neural networks, graph convolutional neural networks, and so on. The chapter also includes applications of such models for real-world traffic prediction problems.