Conjoining Congestion Speed-Cycle Patterns and Deep Learning Neural Network for Short-Term Traffic Speed Forecasting

Forecasting accurate traffic conditions is essential to regional traffic management. Since congestions are usually caused by regular activities, capturing speed-cycle patterns for congestions is useful for short-term traffic forecasting. In this study, we propose a novel approach, namely Neighbor Subset Deep Neutral Network (NSDNN), to forecast spatio-temporal data; the approach conjoins Deep Neutral Network (DNN) and the subset selection method, in order to extract useful inputs from nearby roads. Appropriate input subsets can be selected for DNN training via congestion cycle patterns, in order to reduce input data dimensions and to avoid artificial high correlations from free-flow traffic in off-peak hours. Furthermore, speed data with time lag is also embedded into the DNN model to generate the multi-timestep forecast model. Experimental results demonstrate that the proposed NSDNN achieves higher accuracy, compared to other conventional methods including the Autoregressive Integrated Moving Average (ARIMA), the correlation method and the k-Nearest Neighbor (kNN). In addition, NSDNN is also comparable to the Long Short Term Memory (LSTM) Neural Network, namely NSLSTM, when the same selected input subset is used. The forecasting system can be used by logistic companies to produce better route planning and fleet management in assigning vehicles.