Research Of Traffic Flow Prediction Algorithm Based On Deep Learning

With the continuous advancement of industry,the rapid development of the automobile industry,the number of vehicles such as private cars,taxis and buses has increased dramatically.Many convenient and convenient means of transport have brought convenience to people's daily travel and also to traffic.The road has brought great pressure.Traffic congestion has gradually begun to affect people's travel and daily activities.Traffic congestion has become an urgent problem to be solved.Short-term traffic flow forecasting technology can provide users with traffic flow trends in the short-term future.Users can reasonably select the travel route and travel time according to the traffic flow information in the short-term future,avoid congested road sections,and increase the degree of road access.This paper analyzes traffic flow data and proposes a short-term traffic flow data prediction algorithm based on deep learning based on traffic flow data characteristics.The main work and contributions of this article are as follows:(1)The system introduces the research background and significance of traffic flow forecasting algorithms,summarizes the domestic and foreign research status and research hot spots of traffic flow forecasting algorithms,analyzes traffic flow data,and summarizes the main features of traffic flow data as traffic data.Space-time characteristics and periodic characteristics.The traffic flow prediction algorithm is designed for traffic flow data characteristics.(2)For existing traffic flow prediction algorithms,complex data preprocessing,insufficient analysis of data features,and low prediction accuracy exist.This paper proposes a short-term traffic flow data prediction algorithm based on deep learning.This paper combines convolutional and LSTM networks to generate Conv-LSTM module,and uses Conv-LSTM module to extract temporal and spatial characteristics of traffic flow data,and the temporal characteristics.Full integration with airborne features,compared with existing algorithms,can better integrate the temporal and spatial characteristics of traffic flow data,avoid feature isolation,and use Bi-LSTM to extract the periodic characteristics of traffic flow data.The feature fusion is extracted to obtain the prediction result.Finally,the neural network is optimized to improve the accuracy and improve the generalization performance.The neural network prediction algorithm proposed in this paper has the end-to-end characteristic that can directly process the original data,eliminating the data preprocessing part and reducing the manpower consumption;fully analyzing the characteristics of the traffic flow data,compared with the existing algorithm has a higher accuracy.(3)For the problem of large amount of calculation and long training time for deep learning,this paper sets up a multi-GPU parallel computing accelerating neural network environment to accelerate the computation of neural networks.The neural network will cause large amount of calculations during the training due to forward propagation and backward gradient calculation.In order to improve the ability of deep learning server to shorten the neural network training time,this paper builds a multi-GPU parallel computing environment and performs neural network calculations.Parallel computing accelerates,shortens computing time,and increases system availability.