Short-term speed prediction of urban roads based on multi-source feature fusion

Short-term traffic forecasting is an important means to prevent urban congestion. In this paper, the urban traffic operation time series data sets are used for modeling, combining with various characteristic variables of weather conditions, air quality index (AQI) and travel time index (TTI). A bidirectional long and short-term memory network framework (MF-BiLSTM) based on multi-source feature fusion was established to predict the average road speed. The comparison experiment selects three benchmark models: ARIMA, LSTM and CNN, and a bidirectional long short-term memory network (BiLSTM) model using only road speed time series data sets, and compares the prediction accuracy of these models under three prediction durations. The results show that MF-BiLSTM has the best prediction performance in terms of prediction error measurement, the prediction accuracy is 2.93%, 1.95%, 2.87%, 0.89% higher than that of ARIMA, LSTM, CNN, and BiLSTM, respectively. In addition, comparing the prediction accuracy of each model under different weather conditions and air pollution levels, MF-BiLSTM has the smallest deviation.