Analyzing And Forecasting Time-series Data Via Neural Ordinary Differential Equations

Time series data are widely available in real life,especially in smart cities,intelligent industries and the general trend of big data information era,a large amount of time series data can be easily obtained.However,how to effectively utilize these time series data in real-world scenarios(e.g.,urban traffic monitoring and forecasting,data analysis and intelligent scheduling in industry)has been the focus of attention.In this thesis,we study how to effectively analyze and predict two different kinds of time series data,urban traffic flow and industrial hydropower flow,by using deep learning methods.Existing time series forecasting methods are mainly based on deep neural networks,such as residual networks,convolutional neural networks,and recurrent neural networks.Although these existing approaches have many advantages,they all suffer from high model complexity and violent use of memory parameters,and fail to show strong robustness in the absence of data.In addition,existing methods do not provide an effective trade-off between prediction accuracy and efficiency,offering a high degree of flexibility to decision makers.In this thesis,we propose two methods,TODE and FODE,to address the problem of urban traffic flow prediction and urban traffic flow super resolution,and present Deep Hydro method to predict the hydropower flow time series,whose main elements are respectively as follows: 1)TODE first introduces neural ordinary differential equations to the problem of urban traffic flow analysis and prediction,and utilizes a discretization and then optimization method to improve and balance the prediction accuracy and computational efficiency.This continuous dynamic system can better cope with the situation of partial data missing,and also performs more complex learning on factor fusion.2)FODE,based on TODE,tackles the problems of numerical computational instability and inaccurate gradient calculation in NODE,and also improves the model performance while the memory overhead remains unchanged.The improved enhanced flow normalized flow method additionally takes into account the distribution probability of external factors and strengthens the constraint on urban traffic flow.3)Deep Hydro proposes a novel conditional latent recurrent neural network by combining ODE with RNN,and then introduces a continuous normalized flow to continuously vary the distribution of latent variables for capturing the uncertainty information in multivariate time series data.Finally,the proposed method is validated on two different datasets,urban traffic flow and hydropower flow,and the experiments demonstrate that our proposed time-series models based on neural ordinary differential equations outperform other methods and consumes the least amount of memory and parameters,which demonstrates the efficiency of the proposed model.In addition,the Deep Hydro model proposed in this thesis has been deployed on the application side of National Grid Group and has achieved some initial results.