Research On Deep Spatio-temporal Model Structure

The spatio-temporal model is an important model for describing the spatio-temporal topological relationship,which can be applied to the fields of traffic management,medical diagnosis,military confrontation and others.Owing to the development of machine learning methods such as deep learning structures,and the rapid increase in spatio-temporal data,the deep spatio-temporal model based on black box simulation of structured spatio-temporal data has become a research hotspot in the field of spatio-temporal models.There are two main types of deep spatio-temporal models,the spatio-temporal discriminative models and the spatio-temporal generative models.The spatio-temporal discriminative model is used to predict the conditional probability distribution of spatio-temporal data,while the spatio-temporal generative model is used to predict the joint probability distribution of spatio-temporal data.Due to the complexity of temporal and spatial attribute information in a large amount of spatio-temporal data,it is difficult to extract the time and space features for the deep spatio-temporal model structure,which poses serious challenges to the predictive ability and training speed.This article starts with improving prediction accuracy by studying the structure of deep spatio-temporal model.The main work of this article is as follows:(1)Aiming at the problem of large errors of the spatio-temporal discriminative model when predicting the short-term and medium-term interval data,the deep spatio-temporal convolutional long short-term memory network(DSTCL)is established to improve the prediction accuracy.When predicting the conditional probability distribution,compared with the traditional structure of spatio-temporal discriminative models,the processing of spatial attribute information is added.DSTCL is designed as a multivariate network structure including convolutional neural networks and long short-time memory networks.Experiments show that the DSTCL can extract the temporal and spatial attribute information at the same time.Compared with the mainstream structure of spatio-temporal discriminative models,it can predict the data of long-term,medium-term and shor-term interval,and reduce errors when predicting short-term and medium-term interval data effectively.(2)Aiming at the problem that traditional generative adversarial networks lack the ability to extract structured spatio-temporal data attribute information,the spatio-temporal generative adversarial network(STGAN)is established to improve the prediction accuracy.When STGAN constructs the discriminator,the structure of data attribute extraction is added,making the discriminator a kind of spatio-temporal discriminator.The spatio-temporal discriminator detects the data.It must not only complete the process of the traditional discriminator to distinguish the real data sequence and the generated data sequence,but also determine whether the data sequence meets the requirements for time and space coherence.Experiments show that the sequence generated by the STGAN has the spatio-temporal logic,and the prediction accuracy is improved.Distributed training of multiple STGANs can form a distributed STGAN(STGAN-D).The bootstrap aggregating-down stochastic gradient descent(Bagging-Down SGD)is used to update the network gradient.While ensuring the prediction accuracy,the training speed is improved.(3)In order to improve the diversity and stability of the generated sequence,and further improve the prediction accuracy,the multi-discriminator STGAN(STGAN-MD)is established.The spatio-temporal discriminator is designed into two parts: a temporal discriminator and spatial discriminators when predicting the joint probability distribution.Experiments show that compared with STGAN,STGAN-MD improves the stability and diversity of the generated sequences,and the prediction accuracy is improved.However,when the number of samples in generated sequences is large,the training time will increase significantly compared to STGAN.(4)In order to adapt to the requirements of the future battlefield environment with a higher degree of intelligence,by fully considering the time,space factors and individual intelligence and applying the deep spatio-temporal model structure of this article,an intelligent confrontation simulation platform centered on the spatio-temporal agent is established.The platform includes two sub-platforms,which are a confrontation simulation platform with multiple spatio-temporal agents and an intelligent confrontation platform based on target prediction.The experimental results verify the effectiveness of the deep spatio-temporal model structure in this article.And show that in the confrontation simulation platform with multiple spatio-temporal agents,the STGAN-MD can discover the hidden rules in military confrontation and can also meet the needs of individual intelligence.In the intelligent confrontation platform based on target prediction,the DSTCL can predict the information of the maneuvering target;the STGAN and the STGAN-MD can be used to generate spatio-temporal agents in the case of different numbers of predicted samples.