Research Of Parking Prediction Algorithm Based On Deep Graph Neural Network

As an important link in the field of intelligent transportation,smart parking is an effective solution to the problem of ”difficult parking” in cities.One of the key points of smart parking is to accurately grasp the information of parking resources.However,due to the high cost of sensing equipment and maintaining real-time parking availability information,not all parking spots in the city provide real-time parking information services,which leads to an information asymmetry between drivers and parking lots.Therefore,there is a need to mine parking data information and enhance the interoperability and sharing of parking resource information.In addition,many parking lots only collect and release parking information instead of systematically analyzing related information and data,and of forecasting the parking availability for the time to come.Given that in order to provide better parking services,it is necessary to study how to accurately predict the availability of parking spots.This paper divides the predictions for parking space availability into the following two subones,namely,the prediction for real-time parking space and the prediction for future parking space.For the former one,this paper proposes a parking space prediction algorithm based on inductive graph convolutional network,with the goal of predicting real-time parking space information of a larger range of parking lots from a limited number of real-time parking spaces.The generalized model built by inductive graph convolutional network relying on inductive learning has strong generalization ability,which makes it unnecessary to retrain the model when new parking nodes are added.At the same time,the graph convolution module can effectively capture the potential spatial correlation between parking lots.For the latter one,this paper proposes a parking space prediction algorithm based on spatiotemporal deformable graph convolutional network to realize multi-step prediction of free parking spaces in the future.The spatiotemporal deformable graph convolutional network is mainly composed of a spatial dependency capture module and a temporal dependency capture module.The spatial dependency module uses graph convolution to capture the local spatial features between parking lots,and compensates for the graph volume through an adaptive adjacency matrix.It can capture the potential connection between different parking lots? the time-dependent capture module applies one-dimensional temporal convolution and one-dimensional deformable convolution to work together in the time domain to capture temporal features,Among them,deformable convolution mainly makes up for the deficiency of one-dimensional temporal convolution that may lead to the loss of local information when the expansion rate increases.By combining the temporal dependency capture module and the spatial dependency capture module to construct a spatio-temporal convolution layer to effectively capture the dynamic spatio-temporal dependencies between parking lots,multi-step prediction of available parking lots in the future time period is achieved.In order to verify the effectiveness of the proposed method,several sets of comparative experiments are designed to evaluate the proposed methods on the parking datasets in six areas of Beijing.Experimental results show that the two methods proposed in this paper achieve better performance than the baseline methods in their respective problems.That is to say,the real-time parking space prediction algorithm based on the inductive graph convolutional network improves the prediction accuracy by about 5%,and the future parking space prediction algorithm based on the spatiotemporal deformable graph convolutional network improves the accuracy by about 12%.