Time-series Data Modeling And Analysis Method For Internet Of Thing

With the development of the Internet of Things(IoT),the connotation of urban construction has changed from digital city to smart city.Smart city refers to the application of information and communication technologies to perceive,analyze and integrate various data of the core system of urban operation,so as to make intelligent responses to various demands of services,including peo-ple's livelihood,environmental protection,public security and industrial and commercial activities.However,due to the non-stationary and non-linearity of IoT time-series data,the existing time-series data analysis technologies can not effectively support diverse requirements in IoT.Thus,it is essential to in-vestigate effective time-series modeling and analysis methods.The develop-ment of Deep Learning,provides a promising and efficient solution for IoT time-series data modeling and analysis.However,due to the missing data and short sampling period,a small amount of time-series data can be collected in some scenarios.How to exploit different size of datasets to build time-series data modeling and analytical algorithms oriented towards their typical char-acteristics,and how to implement the algorithms in large-scale networks,to provide efficient and accurate services for the IoT applications are key ques-tions in this dissertation.Firstly,we design a short-term time-series forecasting algorithm based on ensemble learning for small datasets.Then,we focus on investigating the DL techniques required for large time-series datasets:for the single point scenarios consist of one or a few time series,two novel long-term forecasting algorithms based on Deep Neural Networks(DNNs)are proposed;for the network-level complex scenarios consist of numerous time series,we study effective time-series forecasting and completion algorithms,to support the requirements of applications in networks.The main contributions of this dissertation are summarized as follows(1)Time-series data modeling and short-term forecasting algorithm for small datasetsFor small time-series datasets,design a novel ensemble learning-based short term time-series forecasting method.In this dissertation,a novel ad-dictive formulation is first proposed to model the time-series.Specifically,the addictive formulation decomposes the time-series data into two parts,i.e.,the extrinsic-variational component described by the external factors,and the intrinsic-stationary component representing the internal structure of the se-ries itself.Thereafter,a boosting regression learner consists of several simple learners,is further developed to model the extrinsic-variational component and intrinsic-stationary component.After localizing the global attributes,regres-sion algorithms such as multilayer perceptrons are used to learn the extrinsic-variational component by mining the relationships between series patterns and the corresponding local attributes.Then,the intrinsic-stationary component is naturally learned by the classic time-series methods such as AutoRegres-sive Moving Average(ARMA)models.Experimental results show that the proposed method improves the accuracy of time-series forecasting on small datasets,as compared to the existing approaches in the literature.(2)Time-series data modeling and long short-term forecasting algorithm for big datasetsFor big time-series datasets,the existing time-series forecasting technolo-gies mainly mine the relationships between time-series data and external fac-tors,which are usually difficult to forecast accurately.To overcome this dif-ficulty,the dissertation designs two DL methods,Time-Dependency Convolu-tional NN(TD-CNN)and Cycle-based Long Short Term Memory(C-LSTM)network.They only utilize the historical measurements and are independent on the external factors,termed as factor-independent algorithms.By modeling the time series data as pixels and rearranging them into a two-dimensional pic-ture,TD-CNN transforms the temporal correlation of load series into the spatial correlation and keeps the long-term memory.Moreover,in order to extract the temporal correlation between the long-term sequences with lower complexity,the proposed C-LSTM method generates a new short series from the original long series without information loss.The LSTM is then applied to model the dynamical relationship of the series with shorter time steps.Experimental re-sults show that the proposed methods outperform the existing method,with greatly reduced computation complexity(3)Networked method for time-series forecasting technologies based on DeepClusterFor the network-level scenarios,the dissertation studies the feasibility of applying the existing time series forecasting technologies for large-scale net-works.It includes two modules,i.e.,DeepCluster and DeepPrediction.By deep representation learning,the DeepCluster can effectively cluster the time-series and obtain the typical variational patterns of time series.More specifically,to fully exploit the traffic periodicity,the raw series is first divided into a number of sub-series for triplet generation.Then,a binning strategy is designed to vi-sualized the time-series.The CNNs with triplet loss are utilized to extract the features of shape.Thereafter,a model sharing mechanism is further proposed to build predictions through Patten Models(PMs)in DeepPrediction.PM is built for a type of series patterns,instead of Individual Models(IM)exclu-sive for one time-series or Whole Models(WM)universal for all time-series.Our networked method can not only significantly reduce the number of predic-tion models,but also improve their generalization by virtue of being trained on more diverse examples.Furthermore,the proposed framework over a selected urban road network is evaluated.Experiment results show that the DeepClus-ter can effectively cluster the road segments,and PM can achieve comparable prediction accuracy against IMs with less number of prediction models.(4)Time-series data completion algorithm oriented towards networksIn IoT,the premise of decision based on big data technologies is the com-pleteness of data.However,incomplete traffic time-series data are nearly in-evitable due to the constraints of sensors and data transmission.The high com-putational complexity of existing tensor-based completion technologies pre-vents their network-level implementation.Therefore,to tackle these problems,this dissertation designs a batch-oriented time-series data completion algorithm for large-scale networks.In order to preserve different temporal mode of IoT time-series data,we first model the time-series data by tensors so that time-series completion becomes a tensor completion problem.Thereafter,a traffic data completion algorithm is further developed based on the Generative Ad-versarial Net(GAN).More specifically,the formulated tensors are interpreted as samples from an unknown traffic distribution and a GAN is then proposed to learn this distribution.Based on the learned GAN,an objective problem of using the generative tensor approaching the raw tensor corresponding the in-complete tensor is established.By considering effects from adjacent times and nodes,the goal is to minimize the distance between the generative tensor and the incomplete tensor,as well as make the the generative tensor look real.Our experiments on real datasets show that the proposed method can effectively batch incomplete tensors with fast recovery process.