Spatio-Temporal Sequence Prediction Model Based On Multi-Layer Attention Mechanism

As a key technology of spatio-temporal data analysis,time-space data prediction has been widely used in fields such as smart city construction and smart manufacturing upgrades.In the process of building smart cities,accurate air quality forecasts can help regulatory agencies issue air pollution warnings scientifically,quickly complete accident prevention and resource allocation,and improve people’s quality of life.In the field of intelligent manufacturing,accurate supply chain output values forecasting helps companies make correct production decisions,strengthen the organic collaboration with supply chain companies,and increase the overall production capacity of the industrial field.The performance of spatio-temporal series prediction model is restricted by many factors,including:(1)Lack of neighborhood information.For example,there are only 35 monitoring stations in Beijing,and the available neighborhood information is insufficient.At present,spatial interpolation method is mostly used to solve the problem of data sparseness,which will introduce a lot of noise and affect the prediction accuracy.(2)Rough spatial view.At present,convolution neural network is mostly used as a spatial view capture method,without considering the dynamic wind field and other factors,and the spatial relationship mining is incomplete.(3)Single times data relationship.At present,most of the researches take the cyclic neural network as a single time series relationship captures method,only considering the longterm development pattern of time-space series,or only considering the direct effect of shortterm data,without considering the multiple effects of periodicity and short-term mutation at the same time,which makes insufficient use of the existing time series data.Since the spatial dependence on the data is dynamically restricted by many factors,and the time dependence has problems such as loops and mutations,effective spatio-temporal data modeling is extremely challenging.Existing spatio-temporal data prediction models mostly to use static spatial views and a single time data onto prediction,and do not fully consider the complex effects of dynamic spatial correlation and multiple time relationships on the data.In order to solve this problem,the spatio-temporal sequence prediction model based on multilayer attention mechanism(STSPM)is proposed for spatio-temporal data prediction.The main research contents are as follows:Firstly,this paper puts forward a spatial dependency analysis method of spatio-temporal data,extracts the nonlinear correlation between features,analyzes the direct and indirect factors that affect the spatial relationship of data,and integrates cross-domain features hourly to build a dynamic spatial view.Secondly,LSTM encoder is used to obtain the long-term and medium-term development patterns and short-term abrupt changes of spatio-temporal data respectively,and attention mechanism is used to dynamically calculate the influence weights of multiple time series relationships.Finally,LSTM decoder is used to fully integrate historical features to predict spatio-temporal data.Finally,based on the real data set,the prediction performance of STSPM model proposed to this paper is compared with that of six existing baseline models(RNN,LSTM,seq2 seq,STRes Net,DA-RNN and Geo MAN).The experimental results show that the performance of this model in the prediction of time-space series is better than the existing six baseline methods,and it has higher prediction accuracy.