A Data-Driven Approach for Traffic Crash Prediction: A Case Study in Ningbo, China

In the past few years, fully connected Long Short-Term Memory (FC-LSTM) network has been widely used to predict traffic crashes in urban areas. This article attempts to improve the traditional prediction model by adopting Convolutional Long Short-Term Memory (ConvLSTM) network. ConvLSTM can effectively capture the spatial and temporal characteristics of traffic crashes within road network. It overcomes the shortcoming of the FC-LSTM model that ignores the spatial characteristics of traffic crashes. Therefore, the ConvLSTM model shows excellent performance when predicting traffic crashes. To verify the effectiveness of the ConvLSTM, this study uses historical crash data in the City of Ningbo to train the model and compares the result with that from FC-LSTM. The results show that ConvLSTM has better accuracy and lower loss values. Moreover, the model has higher calculation efficiency. Therefore, the ConvLSTM model is more suitable for predicting traffic crashes.