Research On Prediction Models And Applications Of Urban Spatial-temporal Sequences Based On Deep Learning

Aiming at the prominent problems such as air pollution and traffic congestion brought by the process of urbanization,this dissertation conducts research on urban spatial-temporal sequence prediction,one of the key techniques in unban computing.Due to the characteristics of non-linear,high-dimensional and spatial-temporal correlation of urban spatio-temporal sequences data,how to construct a prediction model of urban spatio-temporal sequences effectively,which is a common key technical problem of urban computing research.As a new machine learning method for processing big data,deep learning technology realizes the approximation of complex functions by simulating the multilayer non-linear structure of the visual cortex learning mechanism of the human brain,which has more powerful automatic feature extraction and expression capabilities.Based on this,this dissertation addresses air pollution prediction and traffic flow prediction requirements,which are the typical application scenarios of urban computing,and integrates the multi-level feature extraction and expression ability of deep learning with the nonlinear,multimodal and spatiotemporal correlation analysis methods of urban spatiotemporal sequence data.Then a variety of new deep learning models are developed for urban spatio-temporal sequence prediction based on strategies such as multimodal spatio-temporal correlation,spatio-temporal attention,variational self-coding and hybrid modeling.The main research results obtained are summarized as follows:(1)Aiming at the core problems of urban air pollution forecasting and early warning,a new deep learning model is proposed for predicting air pollution(mainly PM2.5),which is based on a hybrid deep neural network architecture for the first time to learn air quality related spatio-temporal sequence data,and capture the spatio-temporal correlation features and longterm dependence features implied in urban air quality data.The basic module of model design includes multiple one-dimensional convolutional neural networks(1D-CNNs)and bidirectional long-short-term memory networks(Bi-LSTM).The former is used to extract local trend features of sequence data and spatial correlation features between multiple sequences.The latter is used to learn the long-term dependence and spatio-temporal correlation features in the data.The entire model is based on multiple one-dimensional CNN and Bi-LSTM integrated hybrid deep learning frameworks to support the hidden representation and deep features learning of air quality-related urban spatiotemporal sequence data.Finally,experimental evaluation and analysis are performed based on two real air quality data sets.The experimental results show that the proposed model can effectively predict PM2.5 air pollution.(2)Based on the research of urban air pollution prediction under the hybrid deep learning architecture,a new semi-supervised learning learning strategy is adopted,and a novel variational autoencoder model of extended memory mechanism is developed,which aim to solve the problem of urban air pollution prediction.This is an end-to-end deep learning structure that combines traditional variational coding latent variables and memory vectors for deep representation learning of urban air quality-related spatiotemporal sequences.The encoder is based on a bi-directional GRU deep network,and the memory mechanism assists the encoder network to memorize and learn deep features,so as to obtain long-term dependent features and deep nonlinear correlation features of urban air quality data.The experimental results based on two real air quality data sets showed that the proposed model has better PM2.5air pollution prediction performance than the benchmark methods and the hybrid deep learning model.(3)In view of the bottlenecks faced by traditional traffic flow prediction methods,a multimodal deep learning model is presented for short-term traffic flow prediction.This model framework can adaptively learn multimodal traffic flow related sequence data through multimodal deep learning strategies,which can extract and learn the non-linear correlation features and long-term dependent features of multimodality data.The basic module of the proposed model is composed of two components: Convolutional Neural Network(CNN)and Gated Recurrent Unit(GRU)with attention mechanism.Each basic module corresponds to one modality traffic sequence data.The CNN component is used to capture local trend features,and the GRU attention module is used to capture long-term dependent features in sequence data.Finally,through the multi-modal deep learning integration framework,which can be dynamically integrated to perform fusion learning on shared representation features of different modality traffic data.By comparing and analyzing the experimental results of real traffic flow datasets,it is verified that the model can handle complex nonlinear urban traffic flow prediction problems.(4)Based on the study of multi-modal deep learning framework,a new sequence-tosequence spatial-temporal attention learning model is proposed to improve the accuracy and effectiveness of urban traffic flow prediction.This method adopts an end-to-end deep learning strategy based on sequence-to-sequence learning structure and assists the spatio-temporal attention mechanism,which can effectively learn the characteristics of spatio-temporal dependence and non-linear correlation in multivariate spatio-temporal sequence data of related urban traffic data.The analysis of multi-dimensional experimental results based on two real traffic flow datasets shows that the proposed model has the best traffic flow prediction performance compared with the traditional baseline method and the state-of-art method.