Research On Multi Perspective Spatio-temporal Situation Prediction Based On Deep Learning

With the rapid development of artificial intelligence,big data,the Internet of Things and other technologies,more and more information is beginning to be digitized into a form that can be processed by various advanced algorithms.Time and spatial information play an identifier role in these data,and is an important basic component of the complex data information.It is becoming more and more important.It can support the prediction and inference of the spatial location of space-time objects,hot spots of events and other elements,thus mining valuable knowledge and change trends,urban planning and management,traffic congestion prediction.Crime early warning,prevention and control and other scenarios have been widely used.Spatial-temporal situation data is characterized by both contextual time series-related attributes and spatial dynamic changes,and changes in two dimensions are interrelated and indivisible,so spatial-temporal situation prediction needs to explore both spatial and temporal change patterns simultaneously.The general practice is to divide geographic space into multiple grid areas according to latitude and longitude,and use the number of hot spots of events as the statistical index of the grid area,so that the spatial signal is processed as input to similar pixel-level images,learn the latent representation of spatial scale with CNN,and learn the correlation of time scale with RNN.However,many real-world data cannot be effectively represented in Euclidean space using grid formats with pixel-level rules.For example,in traffic flow prediction,different traffic nodes are not adjacent in Euclidean space,but the connectivity between traffic nodes is critical to traffic flow prediction.Therefore,in order to improve the accuracy and validity of space-time situation prediction,it is necessary to dig into the potential semantic information among the fused data in addition to Euclidean spatial correlation.To solve the above problems,a new integrated neural network(DMEN)for multi-view space-time situation prediction is presented,which integrates the predictive output based on geographic information with the predictive output based on semantic information.The DMEN model combines the structure of Pred RNN,GC-Pred RNN and the integration layer to capture space-time situations from different perspectives,and it can also learn some potential high-level situations in the real world without additional data.In order to verify the validity of our proposed model,this paper evaluates and validates the prediction effect of the model on three real datasets of urban hot spots in different areas.The experimental results show that DMEN has better prediction effect than the previous spatial and temporal situation prediction models,and has generalization and validity in spatial situation prediction tasks for different types of hot spots.The results of visualization experiments in this paper also show that DMEN achieves high prediction accuracy and good prediction results in both numerical values and distribution details.The main work and innovations of this thesis are:(1)A method for mining and extracting semantic association information is proposed.The self-semantic graph generated by adjacency matrix is used to mine the latent semantic association information between regions,and the dataset is transformed into an acceptable input semantic form for the model.This makes the prediction model have good generalization and validity in the task of space-time situation prediction for different types and regions.(2)A new in-depth learning method,graphical convolution prediction recursive neural network GC-Predrnn,is proposed.This model can better orient semantic information for space-time prediction.It solves the problem that traditional prediction models only consider the influence of geographic neighborhoods,but ignore the deep-level relationship between regions,and improves the accuracy and performance of space-time situation prediction.(3)A new multi-view integrated model framework for space-time situation prediction is proposed,which combines the structure of Pred RNN,GC-Predrnn and the integration layer,and integrates the models by using the improved Stacking method.Solves the problem that the traditional prediction model predicts from a single angle and results in distorted prediction results.