Urban road speed prediction based on multisource feature bidirectional long short-term memory

Traffic speed prediction plays a vital role in the intelligent monitoring and scheduling of urban traffic. The urban traffic status is affected by a variety of complex factors, considering multiple features to improve the performance of speed prediction model is an urgent problem to be solved. This paper establishes a multisource feature bidirectional long short-term memory (MF-BiLSTM) framework, which considers the influence of weather, air quality and temporal attribute on the traffic state. The bidirectional network structure captures the characteristics in both forward and backward periods, thus effectively capturing the influence of meteorological factors, environmental changes and time-related characteristics on the operation state of urban traffic. On the above model, experiments are conducted on a time-series dataset of the road speed in Chengdu, which is computed based on the DidiChuxing open-source algorithm. Five baselines of the recurrent neural network (RNN), autoregressive integrated moving average (ARIMA), convolutional neural network (CNN), stacked long short-term memory (LSTM) model and single LSTM model are executed for validation. The results indicate that MF-BiLSTM outperforms the above baselines with MAPE improving by 2.94%, 1.02%, 2.45%, 1.10%, 1.67%, respectively, and exhibiting a higher stability under different road grades. In addition, at different weather and air quality levels, MF-BiLSTM exhibits a more accurate and smoother prediction performance. An intelligent traffic prediction system was developed and open source based on MF-BiLSTM, which facilitates the visualization of prediction results.