Research And Application Of Time-series Data Prediction Based On Deep Learning

With the rapid development of big data and artificial intelligence technology,as well as the popularization of various mobile sensing devices and mobile technologies,people have obtained massive spatio-temporal data in the context of geography,that is,nearest neighbor multi-point spatio-temporal data.Analyzing and modeling these spatio-temporal data can provide important decision-making support for social and economic activities such as urban planning,disaster prevention,and resource management,thus generating significant economic and social benefits.In recent years,there have been various prediction models for spatio-temporal data both domestically and abroad,but there are still certain limitations in terms of spatial correlation,operational efficiency,prediction accuracy and so on.Currently,the challenges in predicting nearest neighbor multi-point time series data are as follows: 1.there is a large difference in the sample size and distribution of data between different nodes,and the features exhibit significant heterogeneity,making it impossible to accurately model multiple points using the same model;2.this type of spatio-temporal data has complex spatial dimensions and multiple features,and existing models cannot effectively capture its high-dimensional nonlinear spatio-temporal coupling characteristics,thus requiring the fusion of multiple spatio-temporal features;3.spatio-temporal data prediction tasks require extremely high timeliness and accuracy,but the large scale and amount of data make it difficult for existing models to be trained quickly,thus making it impossible to obtain prediction results in a timely manner.In response to these challenges,this thesis designs a combined deep learning model based on transfer learning and graph structure analysis theory for multi-point spatio-temporal data prediction,and combines the advantages of the two theories from different perspectives to construct a more efficient subgraph transfer spatio-temporal prediction model.The main work and contributions of this thesis are summarized as follows:(1)Aiming at the problems existing in the process of multi-point time series data preprocessing,several data reprocessing methods are proposed.Which is,the data component decomposition algorithm,source domain data selection algorithm and small sample data enhancement algorithm.(2)A two-stage combined deep learning prediction model based on transfer learning is constructed to predict the nearest neighbor multi-point data.In this method,the basic model obtained by pre-training is transferred to several neighboring nodes for fine-tuning training,which can not only retain the spatio-temporal characteristics of the nodes in this area,but also deal with the heterogeneity of each node effectively.Therefore,the prediction accuracy of multi-point in the whole region is better.This method can solve the problems of large difference of sample size and different data distribution among different nodes in space.Taking the landslide displacement prediction task in Baishuihe area of the three Gorges Reservoir and the traffic flow prediction task of Pe MSD8 highway in the United States as examples,compared with other widely used time series prediction methods,this model provides a better solution for the prediction of neighbor multi-point time series data.(3)A parallel–series prediction model based on graph structure analysis is proposed.This model combines the parallel Transformer network and the attention-based graph convolutional neural network to encode the spatial multi-node temporal cycle properties and dynamic spatial association.The parallel Transformer network in this model can capture the temporal association of multiple nodes efficiently and accurately,while the attention-based graph convolutional neural network can analyze the spatial coupling properties between nodes.The model can assign attention weights from different angles to the aggregated information of neighboring nodes in the graph to obtain more accurate spatio-temporal prediction results.The method can solve the problems such as coupling of multiple spatio-temporal characteristics in multi-point prediction.Taking the California highway dataset Pe MSD4 and Pe MSD8 traffic flow prediction tasks as an example,the model provides a certain reference value for the prediction of graph-structured spatio-temporal data compared with other widely used traffic flow prediction methods.(4)A spatio-temporal prediction model based on subgraph-transfer learning is proposed.The model combines subgraph partitioning,graph structure analysis theory,migration learning and other methods for more efficient prediction of nearest neighbor spatio-temporal data.Firstly,the full graph data is divided into multiple balanced subgraphs,and the basic model based on graph structure analysis is pre-trained on the source domain subgraph,and then the basic model is transferred to other subgraphs for fine-tuning training.At the same time,the attention mechanism of "backbone nodes" between sub-graphs is added to analyze the adjacency of nodes between sub-graphs.Taking the Pe MSD8 traffic flow forecasting task as an example,this method is better than other models mentioned above in terms of training time and model accuracy.