Research On Urban Population Flow Prediction Method Based On Residual Network

Urban computing is a city-based process that integrates disciplines such as urban planning,intelligent transportation,sociology,and economics.The process of acquiring,integrating,and analyzing urban big data to improve urban management.The prediction of urban crowd mobility has an important significance on the areas of traffic management,risk assessment,public safety and other areas,and it is also another research hotspot in urban computing.The deep learning method acquires the high-level abstraction of the data unsupervisedly through the hierarchical nonlinear transformation,so that the feature selection is no longer dependent on the task itself and can reduce the time consumption of the adjustment process,but when we have a neural network with significant depth,it will bring about some problems,such as optimization difficulty and gradient disappeared.The residual network alleviates the problems coming with extending the depth of the network layers.How to use the residual network to carry out effective urban crowd flow prediction is a beneficial exploration.In this paper,based on the analysis of the existing residual network structure,we propose an urban crowd flow prediction method named ST-DPResNet,which is based on dual spatial-temporal residual network.In order to model the city grid region and the spatial features which contain temporal proximity,periodicity,and trend,the proposed method takes DenseNet as the backbone and integrates the ResNet path so as to fully utilize the spatial-temporal features of the data to improve the prediction performance of urban crowd flow.According to Experimental results,ST-DPResNet has higher prediction accuracy and better model convergence.On the other hand,in order to simultaneously extract the temporal features and spatial dimension features,it is necessary to extend the 2D convolutional network to 3D,which can further improve the effect of urban crowd flow prediction.Taking the high computational cost of 3D convolution into account,the 3D convolution need to be modified and a pseudo 3D spatial-temporal residual network named P3DST-ResNet is proposed for urban crowd flow prediction.The method decomposes a 3×3×3convolution kernel into two convolution kernels of 1×3×3 and 3×1×1.The former is used to obtain spatial dimension information,and the latter is used to obtain time dimension information.Through decomposing 3D convolution,the method caneffectively reduce the amount of compution.Experimental results show that P3DST-ResNet has a higher prediction accuracy.