The Methodology for Predicting the Monthly Urban Traffic Accidents

Abstract

The number of traffic accidents is an important evaluation index for measuring urban road traffic safety. A methodology was proposed to achieve the accurate prediction of the city traffic accident number using the monthly road traffic accident data in combination with the cross-sectional data such as the weather data, and the city GDP data. First, the cross-sectional data was verified whether it correlate to the traffic accidents concerned. Second, the traffic accident data was smoothed to extinguish the peak values to help the LSTM model capture the dependency hiding in the time-series accident data for the prediction. Finally, considering the uncertain influences incorporated in the test data which cannot train the LSTM model, the three standard deviation criteria was employed to rectify the predicted values made by the LSTM model. The results show that the finally predicted values achieve a better fit to the ground truth, and the accuracy of the prediction is improved by 22%, compared to that predicted by the LSTM model without smoothing the row data and rectifying the predicted values.