Design And Implementation Of Deep Model For Spatial-Temporal Data

In recent years,with the wide application of spatio-temporal data,various analysis of spatio-temporal data has also emerged,and has gradually become a research hotspot with theoretical and applied value.There are many types of spatio-temporal data,including position trajectory data(including traffic trajectory data,logistics trajectory data,GPS data,etc.),remote sensing big data(including satellite remote sensing data,industrial control system sensor data,etc.),and media data associated with spatial locations(refers to digital text,images,sounds,videos,and other media data that have spatial location characteristics and change over time,such as communication data,city surveillance video data,social network data,etc.).Spatio-temporal data generally consists of multiple time series,each of which has a temporal dependence,and these sequences have a spatially related relationship.Spatio-temporal data prediction is aim to predict the future changes of spatio-temporal data by integrating known multiple data information,which has gradually become an important research direction in spatio-temporal data analysis.In the spatiotemporal data prediction task,how to effectively learn the spatio-temporal correlation in the data is an important challenge.However,most methods focus on the temporal or spatial features of the spatio-temporal data.How to analyze the association of spatiotemporal features and achieve accurate prediction is still difficult.Therefore,this paper utilizes the graph convolutional network GCN to effectively deal with these irregular spatial structures,and uses the long-short term memory neural network LSTM to learn the time dependence in spatio-temporal data.In addition,the process of traditional space-time prediction model is often composed of multiple independent modules,but the cost of cooperation between each module is generally expensive.Therefore,this paper proposes an end-to-end spatio-temporal prediction model GCLSTM by combining the characteristics of GCN and LSTM,which can effectively train the entire model.Secondly,since the input sequence of spatio-temporal data is generally long,it is difficult for existing models to learn a reasonable vector representation,so the performance of the model will be very poor.Aiming at this problem,this paper uses two attention mechanisms to pay more attention on the input characteristics related to the prediction task in temporal and spatio,and proposes a deep model GLAT based on the spatio-temporal attention mechanism.Finally,the two spatio-temporal models are combined with the Seq2 Seq model and applied to the traffic field.Two traffic-time prediction methods based on spatio-temporal correlation are studied and proposed,which have certain application significance.The specific research content includes the following parts:(1)In view of the irregular network structure of dynamic networks,this paper proposes to use the graph convolutional network GCN to effectively deal with these irregular spatial structures,and to use the long-short and short-term neural network LSTM to learn the time dependence in dynamic networks,and combine the two advantages of this proposed a new end-to-end depth model,GCLSTM,which enables link prediction for the overall network.Each LSTM unit is followed by a GCN unit.Each time,different structural relationships are used as additional information to further update the time characteristics and transmit information layer by layer.The prediction model contains more feature information and the performance is better.(2)When the input sequence is very long,the model is difficult to learn a reasonable vector representation,resulting in poor performance of the model.Aiming at the above problems,this paper proposes a spatio-temporal attention-based deep model GLAT which is capable of processing long input data sequences,and utilizes spatial attention mechanism to learn the spatial correlation between hidden state and cell state at each moment,and adopts the temporal attention model to learn the temporal information of the link states of the network node,which focuses on the part of the learned spatio-temporal feature that is most relevant to the task,thereby improving the dynamic link prediction performance.(3)Application verification of traffic speed prediction.In the real world,the transportation network is composed of a large number of criss-crossing roads,and the traffic state of each specific road section may be affected by its adjacent road sections.However,the traditional traffic prediction model Seq2 Seq model aims to learn the temporal dependence within the road intersection,rather than the spatial dependence,thus ignoring the influence of adjacent road segments.This paper uses the GCLSTM model and the GLAT model combined with the Seq2 Seq model,and integrates the structural information of the traffic network and the temporal information of the traffic speed of each road to help speed prediction of the entire traffic network.