Research And Application Of Spatiotemporal Series Prediction Based On Multi-branch Networ

Spatio-temporal sequence forecasting refers to the prediction of future trends of a series of related variables using temporal and spatial information.Traditional methods perform poorly in dealing with such highly nonlinear problems as spatio-temporal sequence prediction,which cannot accurately predict future trends,while deep learning models have good prediction performance by learning spatio-temporal dependencies through a flexible network structure.However,existing deep learning models still have some problems,such as difficulty in capturing spatio-temporal dependencies,which lead to degraded prediction performance.In this thesis,we address the shortcomings of existing models and conduct a study on spatiotemporal sequence prediction with precipitation prediction and traffic flow prediction as the research objects:(1)Most previous spatiotemporal sequence prediction studies have not adequately captured the long-term spatiotemporal dynamic offsets and data heterogeneity in the data,resulting in inadequate modeling and poor prediction performance.In order to perform more accurate spatio-temporal sequence prediction,this thesis proposes a new deep learning model for precipitation prediction,called the two-branch self-attentive convolutional LSTM,which includes a short-term branch network and a long-term branch network,and the two branch networks predict the inputs at different time intervals to capture the heterogeneity of the rainfall data,in addition,a new fusion module is proposed to fuse the the outputs of the two branch networks.This thesis also constructs a new memory unit called 3DALSTM by embedding 3D convolution and self-attentive mechanisms to capture spatio-temporal dependencies and extract spatio-temporal features,which exhibits excellent prediction performance on real-world rainfall datasets.(2)As a kind of spatio-temporal sequence prediction,the traffic flow prediction dataset contains a large amount of spatio-temporal data,and the traffic data has strong periodicity and spatio-temporal dependence.Based on the two-branch network in Chapter 3 to explore the decoupling prediction of periodic spatio-temporal sequence data using multi-branch networks,a deep learning model for traffic flow prediction,called multi-branch graph attentional temporal convolutional network MGATCN,is proposed,which consists of three temporal branch networks for the temporal relationships of proximity time,daily cycle and weekly cycle time periods,respectively.modeling.Each network stacks multiple spatio-temporal blocks and is constructed using the graph attention module and gated temporal convolution module,and the feature stitching module and multi-branch fusion module.The spatio-temporal information is captured dynamically using the graph attention module and the gated temporal convolution module,and in addition,an attention mechanism is used to mitigate the oversmoothing phenomenon and deepen the network.Three large real-world traffic datasets are used as experimental samples for experimental evaluation with other methods,and the experimental results fully demonstrate that the performance of the methods in this chapter outperforms other deep learning models.