A Spatial-Temporal Deep Learning Framework for Network-based Traffic Flow Prediction

Authors: Yang Zhang*, University College London, Yibin Ren, SpaceTimeLab, University College London, Tao Cheng, SpaceTimeLab, University College London
Topics: Quantitative Methods, Spatial Analysis & Modeling, Temporal GIS
Keywords: spatial-temporal dependency, road network, graph convolution, residual long short-term memory
Session Type: Paper
Day: 4/4/2019
Start / End Time: 3:05 PM / 4:45 PM
Room: Capitol Room, Omni, East
Presentation File: No File Uploaded

Short-term traffic flow prediction is important in many practical applications, such as real-time route planning and congestion alleviation. The network-based traffic flow forecasting is challenging due to the spatial-temporal dependency considering the topology of a citywide road network. Recently, various deep learning (DL) methods have been proposed for grid-based short-term traffic flow prediction, which cannot model the spatial-temporal data with network structure. This paper develops a multi-layer residual graph convolution long short-term memory network (RGC-LSTM) for road-network-based traffic forecasting. This model integrates a novel graph convolution operator (spatial domain) and a proposed residual LSTM structure (temporal domain) to automatically extract spatial-temporal features from network-structured data accounting for the real topology of the road network. The proposed model has fewer parameters, lower computational complexity and faster convergence rate. The framework is illustrated by a case study using a large-scale traffic flow data for 10-min, 20-min, 30-min and 60-min traffic forecasting, which shows the effectiveness of our approach over various state-of-the-art baselines.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login