Research And Application On Prediction Methods For Multidimensional Sparse Spatio-temporal Data

A large number of spatial and temporal prediction problems exist in scenarios such as air quality monitoring,infectious disease transmission,weather forecasting,international import/export logistics,and traffic flow analysis.It is necessary to construct efficient and accurate spatio-temporal prediction models to provide theoretical and technical support for practical applications.In recent years,a series of research results have been obtained for large scale and high resolution spatio-temporal data prediction,but the research on prediction methods for sparse spatio-temporal data is still in the initial stage.A series of bottlenecks are still to be solved,such as feature representation of sparse spatial data,the overfitting tendency of prediction models,and the interference of data uncertainty.To address these bottlenecks,this paper takes sparse spatio-temporal data in air quality monitoring and COVID-19 surveillance as the starting point,and develops a prediction method for multidimensional sparse spatio-temporal data from the perspective of cross-fertilization of artificial intelligence theory and prior knowledge.The main research content are as follows:(1)To overcome problems coming with the spatial sparsity of air quality monitoring data and the reading drifts,cross interference as well as the time-varying and spatially heterogeneous monitoring accuracy of low-cost sensors,a low-cost sensor calibration model based on spatio-temporal attention mechanism is proposed.Firstly,we model this spatiotemporal calibration model as Encoder-Decoder structure.Under this framework,shortterm encoder aggregates real-time monitoring data from micro stations by spatial attention mechanism and learn the short-term fluctuation representation of pollutants through Gated Recurrent Unit.Long-term encoder utilize historical time series from the static station to learn the long-term trend and periodic pattern of pollutant concentration,and provides information from the prediction perspective.Finally,the decoder consider both long-term features and short-term features to generate the calibration results.We calibrate the sensor readings of carbon monoxide and ozone on two real-world datasets,the proposed model outperforms the 8 baseline models and has better robustness.(2)To address the challenges posed by the temporally sparse,lagging and underreporting COVID-19 surveillance data and the dynamic transmission pattern of disease,a prediction model based on meta learning for the number of new infected individuals is proposed.Firstly,spatial dependency is captured by the Graph Attention Networks that aggregate neighborhood embeddings.Gated Recurrent Units learn the temporal pattern of fused features and give the prediction results from short-term view.Meanwhile,the parameters of the infectious disease dynamics model SEIR are dynamically generated by the hidden states from Gated Recurrent Unit,by which the prediction results in long-term perspective is given.Additionally,the initial parameters of the model are optimized by the meta-learning algorithm MAML before the training phase.Finally,we evaluate the proposed model on US state-level dataset,and the results show that our model outperforms six baseline models and converges quickly to the optimal parameters even with small scale datasets.