Adaptive Attention-Based Spatio-Temporal Forecasting Framework for Urban Traffic Speed Prediction
TL;DR
The results show the effectiveness of spatial graph learning, temporal convolution and adaptive attention in forecasting traffic speeds across the benchmark urban transportation datasets and provide a promising way to apply the proposed method in real scenarios.
Abstract
Spatio-temporal traffic forecasting, with reliable temporal and spatial information, is a crucial component of any urban transportation network for intelligent transportation systems and the management of mobility. This study proposes an Adaptive Spatio-Temporal Forecasting (ASTF) framework based on Graph Neural Networks (GNNs), Temporal Convolutional Networks (TCNs) and an adaptive attention mechanism. The GNN models spatial relationships between interconnected traffic sensors and TCN models temporal patterns and changing traffic conditions. Adaptive attention additionally enhances prediction by putting more weight on important sensor positions. The framework is tested on the well-known METR-LA and PEMS-BAY benchmark datasets that includes measurements of traffic speed from urban road networks. The results show the effectiveness of spatial graph learning, temporal convolution and adaptive attention in forecasting traffic speeds across the benchmark urban transportation datasets and provide a promising way to apply the proposed method in real scenarios.