Research On Traffic Flow Prediction Algorithm Based On Deep Learning

In traffic forecasting scenarios,a real-time and accurate prediction method is crucial for controlling and managing urban traffic flow.However,on the one hand,traffic flow data has complex nonlinear spatial correlation,which means different road nodes have different connectivity relationships.On the other hand,it has a dynamic temporal correlation.That is,the spatial correlation of road nodes will change with time.At the same time,the traffic flow data has a certain periodicity as a whole,but there are anomalies and uncertainties locally.Therefore,how to efficiently mine hidden correlations in traffic flow data has become an important task in traffic prediction.In addition,the scale of traffic flow data is huge,and the existing methods require a lot of parameters and time for training and prediction,which cannot meet the real-time requirements.Therefore,real-time traffic prediction is also an important research direction.In order to mine the complex spatio-temporal correlation of traffic flow data,reduce the training time of the model and achieve real-time performance,this thesis proposes three new traffic flow prediction methods.Furthermore,it proves the superiority of the proposed method through comparative experiments.The main work is as follows:(1)Aiming at the problem that the existing traffic prediction methods lack flexibility in the process of extracting spatial feature information and ignore the correlation between local and global spatio-temporal features,we propose a new method termed traffic prediction method integrating graph wavelet and attention mechanism.This method uses wavelet transform and an adaptive matrix to extract the local and global spatial features of traffic flow respectively.It combines the improved recurrent neural network to extract local temporal characteristic information.At the same time,this model uses the attention mechanism to capture the temporal and spatial dynamic variability and applies a spatio-temporal feature fusion mechanism to fuse the local and global temporal and spatial features.Experimental results show that this mothod can efficiently extract the spatial and temporal features of real traffic datasets.Moreover,it can outperform the existing methods.(2)To settle the problem that the distance-based adjacency matrix cannot fully express the adjacency relationship of the roads in the traffic network,and the recurrent neural network(RNN)based traffic prediction method cannot be parallelized and difficult to train,this thesis proposes a method termed the traffic flow prediction method for traffic pattern similarity,and this method uses normalized cross correlation algorithm to calculate the traffic patterns of each road.Then selecting the road with the highest traffic pattern correlation as the neighbor road.Meanwhile it adaptes the residual diffusion graph convolutional network to extract the spatial characteristics of traffic flow.In addition,the temporal convolutional network(TCN)and gated tanh units(GTU)are employed to extract the temporal correlation of traffic flow.Moreover,it can outperform the existing methods.(3)In view of the fact that most of the current traffic flow prediction methods focus on the accuracy of the models,while ignoring the training time,inference time and data privacy of the models.This thesis proposes a cloud-edge collaborative short-term traffic flow prediction method.The Fed Avg algorithm framework is used to deal with the parameter update problem of edge and cloud models.At the same time,a graph division method is used to divide the adjacency matrix that combines traffic similarity and distance similarity into several subgraphs.The subgraphs are assigned to the corresponding edge server for local training.In the end,the method uploads the training parameters to the cloud and updates the corresponding global model parameters.The experimental results show that this method can effectively reduce the training time and inference time of the model under the premise of ensuring a specific prediction performance and has good robustness.