Research On Crowd Flow Prediction Method Based On Graph Neural Networks

With the increase of the population and the number of vehicles,more and more attention has been paid to the traffic planning and management.Reasonable traffic planning and public traffic layout can reduce the urban congestion and traffic safety hazards.In order to optimize the efficiency of urban traffic management and improve the travel experience of citizens,grid-based urban traffic flow forecasting methods are widely used in Intelligent Transportation System(ITS).However,the existing work faces many problems,due to the randomness,the highly nonlinear spatial-temporal dependencies and periodicity of traffic flow,accurate prediction becomes very challenging.Previous studies usually divided cities into regions of the same size and coded history data as heat map.For cities with complex terrain,heat maps contain a lot of invalid data,the data in some regions is zero at any time,such as rivers,lakes,mountains,etc.Invalid data will affect the model’s acquisition of spatial correlation.In order to reduce this effect,the traffic flow data is encoded as a graph,and a multi-graph neural network model is proposed to solve the one-step traffic flow prediction problem.This thesis first constructs two K-nearest neighbor graphs through the Euclidean distance and the Pearson correlation coefficient between regions,and then obtains the spatial dependencies of the two graphs through the spatial module composed of a graph attention network and a Cheb Net.Then,a LSTM and a self-attention mechanism are deployed to obtain the time dependence of all regions.Finally,the prediction result is carried out by a fully connected layer.In addition,in order to improve the efficiency of traffic flow forecasting,a multi-step traffic flow forecasting model is proposed.In the one-step traffic flow prediction task,the next prediction result depends on the data of several time interval before,if the iterative method is used for multi-step prediction,the previous prediction error will be propagated to the subsequent prediction results.In order to solve this problem,this thesis focuses on modeling the time correlation of multi-step traffic flow prediction periodically,and constructing the input of the model from three time ranges: long-term,medium-term and short-term,and the attention mechanism is used to judge the influence of different periods of data on the traffic flow at each target time step.finally,multi-step traffic flow prediction is completed based on seq2 seq model.Finally,according to the GPS data of taxis in Chongqing city,this thesis constructs the Taxi CQ data set and then evaluates the one-step and multi-step traffic flow prediction models on the Taxi CQ and Bike NYC.The experimental results show that those two models proposed in this thesis have higher prediction accuracy than other compared models.