Based on the Spatio-temporal Characteristics and CNN LSTM of the Short-term Prediction and Evaluation of Urban Road Traffic

基于时空特性和1D CNN-BiLSTM的城市道路交通短时预测与评价

Keyu Chen 2037673
Supervisor: Dr. Arodh Lal Karn

This paper analyzes the characteristics and limitations of the existing short-term traffic parameter prediction model. Based on the spatiotemporal information, the 1D CNN-BiLSTM model exploits the periodicity and similarity of traffic parameters in the temporal dimension, as well as the similarity transfer in the spatial dimension. In this model, the 1D CNN model inputs data with spatial characteristics (merging two traffic parameter data of adjacent roads), and can effectively extract the spatial characteristics of the data through a series of steps. The BiLSTM model can consider bidirectional timing information simultaneously, which means capture more temporal features than a single LSTM. Therefore, 1D CNN-BiLSTM short-term traffic parameter predictive model has the spatio-temporal information capacity of road traffic in the actual situation. Meanwhile, traffic and average speed, which are inherently related, are used as two parameters to construct the data set. Following model training and testing, it is demonstrated that the model has effective learning ability. Finally, the congestion prediction accuracy of the model is high, as demonstrated by the calculation of the congestion index with the prediction results. This indicates that the model has excellent congestion prediction ability.

The main conclusions include the following three aspects:(1) A CNN-BiLSTM short-term traffic parameter prediction model is constructed.(2) The effectiveness of the short-term traffic parameter prediction model is verified.(3) CNN-BiLSTM model has certain applicability for traffic operation state prediction.