Research On Traffic Flow Prediction Method Based On Parallel Neural Networks

In recent years,with the improvement of people’s living standards,the demand for urban travel is increasing sharply,and the rapid increase of travel vehicles has not only caused parking difficulties and traffic congestion in major cities,but also increased the frequency of traffic accidents.As an important part of intelligent transportation system,accurate and real-time traffic flow prediction is of great significance to alleviate traffic problems.However,most of the current methods focus on short-term time-series traffic flow prediction,and few methods for long-time series prediction are studied,and long-time series traffic flow prediction still faces serious challenges;secondly,due to the problems of gradient disappearance and memory constraints,most existing methods are difficult to use the full contextual information of long-time series,and the model robustness is low,resulting in low performance of long-time series prediction;finally,the traffic flow data often have complex spatio-temporal correlations,which makes the task of traffic flow prediction much more difficult,and most of the existing traffic flow prediction methods predict them at the short-term time series level,with different degrees of defects in mining the spatio-temporal characteristics of traffic data.To address the above problems,this thesis conducts research on both time-series and spatio-temporal forecasting in the following two aspects:(1)For the traffic flow prediction problem of long time series,this thesis proposes a parallel time series prediction model called LDformer.First,Informer is combined with LSTM to obtain the depth representation capability of the time series.Second,we propose the parallel encoder module to improve the robustness of the model,and combine the convolutional layer with the attention mechanism to avoid the value redundancy in the attention mechanism.Finally,this thesis proposes a probabilistic sparse attention mechanism combined with Uni Drop to reduce the computational overhead and mitigate the risk of losing some key connections in the sequence.Experimental results on two real traffic flow datasets from the California highway network show that LDformer outperforms the state-of-the-art baseline for most of the results in different long time sequence prediction tasks.We also performed experimental validation using datasets from other domains to demonstrate model validity,and conducted ablation experiments for each module.(2)In this thesis,we propose a parallel spatio-temporal traffic flow prediction model MGNN considering the temporal and spatial correlation of traffic flow data.we use the idea of meta-learning with three graph neural networks,namely graph convolutional network,graph attention network,and Chebyshev graph convolutional network,to input traffic flow data.First,we combine the advantages of the three graph neural networks by assigning weights to them through meta-learning;second,we run the three graph networks in parallel to improve the robustness of the overall model;finally,this thesis uses a gating mechanism to fuse the output results of the three components and further extract the information to obtain the final prediction results,thus improving the accuracy of the final prediction.Experimental validation on two real California highway network traffic flow datasets and comparison of the prediction results with those of deep learning models commonly used for prediction show that the proposed prediction model in this paper has good performance and can provide an effective basis for traffic management and control.