Short-Term Traffic Flow Forecasting A wide and deep approach with periodic feature selection

Accurate traffic flow forecasting can improve the effectiveness of traffic management schemes, leading to reduced traffic congestions, emissions, costs, and improved health and productivity of commuters. However, the complex non-linear, periodic, and dynamic spatial-temporal nature of traffic flow data makes traffic forecasting challenging. This work proposes a deep-learning model, based on wide and deep architecture, to mine both periodic and spatial-temporal features. The wide section consists of a fully connected layer optimized to mine the periodic feature, while the deep portion consists of a conv-LSTM module, which extracts the spatial-temporal features. The network is made more efficient by analyzing traffic data periodicity and selecting the most relevant features, namely the weekly pattern. Simulation results demonstrate that our proposed model outperforms existing approaches whilst requiring one less input, thereby reducing the amount of pre-processing required. The evaluation of model accuracy considers the GEH statistic, used by traffic modellers, as well as traditional criteria adopted in traffic forecasting.

Recommended citation: Martin Esugo, Qian Lu, Olivier Haas. "Short-Term Traffic Flow Forecasting A wide and deep approach with periodic feature selection." TechRxiv. May 17, 2022 http://komehz.github.io/files/2022-05-17-paper-002.pdf