Study Of Time Series Prediction Model Based On Deep Spatio-Temporal Characteristics

By exploring the interdependence of various time series data at historical events and identifying their pertinent properties,people can predict future time data,which is widely utilized in many areas such as finance,energy,medicine,and meteorology.Due to the recent rapid advancements in sensor,computer,and communication technology,there is an unprecedented opportunity for time series prediction,but it is also a significant difficulty.A more complete view of the depth of its internal characteristics is given by the abundance of time series data,such as long-term complex time dependence,dynamical changing spatial characteristics and unknown fused spatialtemporal correlation.How to extract such features efficiently and accurately has become a key scientific problem in time series prediction.Therefore,it is of great importance to explore time series prediction models based on deep spatial-temporal characteristics.Four time series predicted models are proposed to address the aforementioned fundamental scientific problems.These models are validated on publicly available datasets,such as the meteorological dataset SML2010,the stock dataset Nasdaq100,and a time series dataset under realistic application scenarios of underground coal gasification.The main work and contributions are as follows.1)To address the problem of long-term complex temporal dependence,we propose a multivariate time series predicted model based on Hybrid Temporal Convolutional LSTM Network.The dilated causal convolution and residual networks are introduced to extract the temporal features of multivariate time series,expanding the perceptual range of the network and better learning the relationships between the time interiors of multivariate time series.The traditional RNN models such as LSTM are combined with temporal convolutional modules to solve the problem of gradient disappearance or gradient explosion of multivariate time series data.Experiments show that the model has excellent performance in multivariate time series prediction compared to other prediction methods based on temporal characteristics.2)To address the problems of long-term complex temporal dependence and dynamical changing spatial characteristics,a dual-stage attention-based Conv-LSTM network is proposed.The model introduces a convolutional layer to extract the spatial features.A new method for pre-processing time series data is presented in order to improve the convolution operation.On the basis of extracting the spatial characteristics of multivariate time series,LSTM networks are combined with attention mechanisms to extract their temporal correlation more precisely.Experiments show that the model has excellent performance in multivariate time series forecasting compared to other forecasting methods of the same type.3)The two-stage Transformer model of graph information embedding based on temporal and spatial features is also proposed to address the problems of long-term complex temporal dependence and dynamical changing spatial characteristics.To get more accurately characterize the temporal and spatial features,the model constructs adaptive spatial and temporal graph structures.Extraction of spatial and temporal correlation of multivariate time series by combining the spatial Transformer module,HTC-LSTM module and temporal Transformer module respectively.Experiments show that the model has excellent performance in multivariate time series forecasting compared to other forecasting methods of the same type.4)To address the problem of unknown fused spatial-temporal correlation,a multivariate time series predicted model based on adaptive graph neural network with fused spatio-temporal features is proposed.We construct a new fused spatio-temporal graph that fuses the entire sliding window individual spatial graphs into a large fused spatio-temporal graph based on the interrelationship of spatial graphs in each time dimension,which preserves the hidden temporal,spatial,and temporal connections.We also construct an adaptive adjacency matrix by node embedding method without any guidance of a priori knowledge,and dynamically optimise the parameters in a datadriven manner during the training process to obtain the optimal spatio-temporal connections.Experiments show that the model has excellent performance in multivariate time series prediction in comparison with other models.The findings can further guide the design of time-series-based decision-making and control,so that the corresponding system proceeds in a favourable direction.This thesis consists 44 images,36 tables and 153 references.