Prediction Of Mixed Spatio-temporal Traffic Flow Based On Dual Path Network

With the continuous growth of the population and means of transportation in domestic cities,the existing resources in the city are increasingly difficult to meet people’s travel needs,and traffic accidents and traffic jams frequently occur.Therefore,it is very important to launch an intelligent transportation system.As the core of the intelligent traffic system,traffic flow prediction can predict the current or future flow through the past flow data laws,and can provide scientific planning for the city’s traffic system.Because traffic flow is greatly affected by time factors,most traditional traffic flow predictions are based on time series methods,and cannot extract spatial information features.However,traffic flow is also closely related to spatial characteristics and is affected more by spatial factors.If the characteristics of time and space can be obtained at the same time when predicting traffic flow,then the accuracy of traffic flow prediction will be greatly improved.Therefore,this paper proposes a mixed spatio-temporal flow prediction model based on dual-path neural network,ST-DPN-ext to predict traffic flow.By analyzing the impact of different historical time on flow,dual-path neural network and LSTM are used to predict different flows.The data is processed with time and space features,and then the weight is dynamically integrated for each time through the attention mechanism,and finally combined with weather factors to make predictions.Experimental results show that compared with other traffic prediction models,our proposed ST-DPN-ext has better prediction effect.The research results of this article are as follows:1.Traditional traffic flow prediction methods usually only deal with the temporal or spatial characteristics of traffic flow.In this paper,the ST-DPN model is proposed.For the time characteristics,the DPN network is designed to extract the spatial characteristics.This method can extract the spatial features more effectively than the ordinary convolutional neural network;for the time characteristics,LSTM is used to time Feature extraction.By combining the two to deal with the spatio-temporal characteristics of the flow rate,it is possible to predict more effectively.2.Most traffic forecasting methods are based on a single time domain.This paper analyzes the time characteristics of flow through actual data,and obtains three characteristics of flow proximity,daily periodicity and weekly trend.Experiments prove that the prediction effects of different time-domain features are different,that is,the degree of influence is different.Therefore,we propose the ST-DPN-ext model,using the attention mechanism to assign weights to different time-domain prediction maps after the ST-DPN model training,and Combine it dynamically.Compared with the single time-domain features of ST-DPN training,the prediction effect of the ST-DPN-ext model is better.3.Analyze the influence of weather factors on traffic flow through actual data,and further combine weather factors to make predictions.The experimental results show that the ST-DPN-ext integrates weather factors,and the predicted value fits better with the true value.