Research On Traffic Flow Prediction Algorithm Based On Knowledge Distillation

With the acceleration of China’s urbanization and the rapid increase of residents’ car ownership,the road congestion has become more and more serious.Traffic flow prediction plays an important role in alleviating urban traffic congestion.Most of the existing traffic flow forecasting models are based on deep learning,which has complex structure and high training cost,and can not be well deployed to small equipment or users.In order to reduce the training cost of deep learning model,knowledge distillation method has become a research hotspot.This thesis analyzes the temporal and spatial characteristics of traffic flow data.In order to effectively capture the temporal and spatial dependence of traffic flow,a traffic flow prediction model based on improved graph wave network is proposed.At the same time,the method of knowledge distillation in other fields is introduced into the task of traffic flow prediction,and the traffic flow prediction model based on improved graph wave network is taken as a teacher model,taking the dynamic attention network model with linear components as another teacher,a dual-teacher-student model is proposed to solve the problem of large training cost of deep learning model.The research contents of this thesis are as follows:(1)According to the time dependence of spatiotemporal data,a triple time convolution network is proposed.By merging the three time convolution networks,the capture of time dependence in the data is enhanced.(2)The linear component is integrated into the model.The autoregressive model is added as the output of linear component,triple time convolution network and graph convolution network to achieve the combination of linear and nonlinear,which improves the precision of the model.(3)For the first time,the knowledge distillation method is introduced into the field of traffic flow prediction.By using the classical knowledge distillation method,the single-teacher-student model and the double-teacher-student model are constructed,and the teacher leads the training of the student model by simulation,which reduces the time cost under the condition of less precision deviation.(4)The method is verified on two public traffic data sets.The experimental results show that the traffic prediction model based on the improved graph wave network is better than most of the contrast models,at the same time,the first attempt of the knowledge distillation method also achieved good results.