Research On Multivariate Time Series Prediction Method Based On Spatio-temporal Correlation

Multivariate time series prediction is a hot research topic of time series data mining,which has been widely applied in many fields,e.g.,meteorology,economy,environment,energy,and traffic.With the increasing scale of multivariate time series data,it is challenging to effectively capture the correlations between different time steps in the temporal dimension and the correlations between series in the spatial dimension and provide valuable information for different types of prediction tasks.Traditional prediction methods cannot thoroughly learn the spatio-temporal correlations in various multivariate time series prediction tasks,resulting in difficult prediction results to meet practical needs.Due to the excellent feature extraction ability of deep learning models that can learn the complex and dynamic interdependence among time series.To this end,for different multivariate time series prediction tasks,including single-step prediction,multi-step prediction,single-target prediction,and multi-target prediction,the dissertation introduces the advanced theories and technologies in deep learning,and proposes a series of multivariate time series prediction method based on spatio-temporal correlation to capture semantic spatio-temporal correlation,multi-dimensional spatio-temporal correlation,multi-scale spatio-temporal correlation,and multi-type spatio-temporal correlation.The main research contents and innovations of the dissertation are as follows:(1)For across multi-step prediction tasks of multivariate time series,the dissertation proposes a correlational graph attention-based Long Short-term Memory network to capture semantic spatio-temporal correlation.The model is a nested network that nests the correlational attention mechanism in the graph attention mechanism to extract semantic spatial correlations and transform sequence features into higher-level features to obtain sufficient expression capability.A graph attention mechanism is introduced to calculate the weight coefficients between the current and previous time steps to capture the temporal correlations among different time steps.Experimental results show that the proposed model can significantly improve accuracy in single-step and across multi-step prediction tasks of multivariate time series.(2)Existing multi-step prediction methods cannot extract relevant features that affect future values from multiple dimensions,resulting in a lack of learning dependencies.Based on the studies on multivariate time series single-step prediction,the dissertation proposes a multidimensional spatio-temporal attention-based Recurrent Neural Network to address this issue.The model captures local and global spatio-temporal correlations from the perspectives of onedimensional and two-dimensional spatio-temporal.Experimental results show that considering multi-dimensional spatio-temporal correlation can significantly improve the performance of multivariate time series multi-step prediction tasks.(3)For single-target prediction tasks of multivariate time series with geospatial characteristics,the dissertation proposes a joint network of non-linear graph attention and temporal attraction force to capture multi-scale spatio-temporal correlations.First,the original time series data is divided into multi-scale subsequences from the temporal and spatial dimensions.Then,the non-linear graph attention mechanism and the temporal attraction force mechanism are combined to capture the spatio-temporal features of these subsequences.Finally,multi-scale feature fusion is performed.Experimental results show that the proposed model effectively predicts the future values of a specific location.(4)For multi-target prediction tasks of multivariate time series with geospatial characteristics,the dissertation proposes a joint model of autoregressive and deep neural networks to capture the multi-type spatio-temporal correlations.A spatial module is designed to fuse semantic and geospatial correlations in the spatial dimension.In the temporal dimension,a temporal module is designed to extract temporal correlations of individual time series.Moreover,the traditional autoregressive linear model is integrated with parallel with a deep neural network to improve the robustness of the prediction model further.Experimental results show that the proposed model can effectively predict the future values of multiple locations simultaneously.