Reseach And Application Of Seasonal Time Series Forecasting Model Based On Data Feature Driven Decomposition

Time series forecasting is one of the most widely used data-driven forecasting methods.Early classical time series forecasting models are mainly based on the theory of stochastic process and mathematical statistics.With the development and cross application among the computer science,mathematics,management science,statistics and biological engineering disciplines,computational intelligence methods such as machine learning methods obtains the rapid development,and with the parameter estimation and structural flexibility as well as the superiority of the forecasting performance,these methods have become a hot area of research in the field of time series forecasting.However,either the classical time series forecasting model or the advanced machine learning model,any kind of method has its advantages and disadvantages and is insufficient to cope with all types of time series forecasting problems.Especially under the background in the era of Big Data,the interaction of the Internet,information technology and human life results in the rapid accumulation and expansion of data resource,in the face of such a large amount of information,almost all of the model structure,parameter estimation and operation calculation,stability and adaptability will be challenged.Based on the above research background,on the basis of combing and summarizing the traditional statistical forecasting model,machine learning model,single model,combined model and hybrid forecasting model,this paper focuses on the construction of complex seasonal time series forecasting model based on data feature driven decomposition.Specifically,the forecasting model established in this paper has a systematic theoretical framework,which mainly covers the following four aspects:First,the construction of the forecasting model is based on the idea of data characteristics driven.Most of the current forecasting model pay more attention to the impro vement of the method itself,but less to the data feature as a starting point to cut into the model construction.The volatility of seasonal time series data as a distinct data type is driven by the mix of multiple potential features,and if one could dig out the different components and perform the effective analysis,the established model has more specific,rational and robust performance.Secondly,aiming at the structural features of the seasonal time series data,the model construction adopts the solution seeking mindset based on decomposition,the purpose of which reduces the uncertainty risk caused by the high complexity of the original data,and makes it easier to mine and analyze the internal pattern of the time series.Thirdly,the advanced computable intelligence method is taken as the core element of the established forecasting model.In this regard,machine learning model outperforms the classical time series model in forecasting accuracy and applicability.Fourthly,feature selection,as an important direction,is ignored by most of the mixed forecasting models.The existing mixed forecasting models pay more attention to the method itself,but few studies carry out in-depth feature analysis.Based on the above four aspects,this paper fully considered the unique fluctuation pattern of seasonal time series data and carried out constructing complex forecasting model for seasonal time series.The research content of this paper consists of six parts:Chapter one introduces the research background,ideas of the research topic,main research content,research significance,research innovation and deficiencies.Chapter two systematically discusses the data feature driven decomposition and feature identification method,which is the basis of the following research.In Chapter three,a hybrid forecasting model is established based on the designed denoising strategy based on data feature driven decomposition and machine learning,considering that the seasonal time series data are disturbed by the irregular fluctuation information.Chapter four lengthens the perspective of data feature driven decomposition but is different from the idea of denoising.Based on the solution seeking mindset by decomposition-ensemble,the theoretical framework of ensemble strategy is constructed and a series of decomposition integration model is proposed.Chapter five expands and deepens the research in Chapter three and Chapter four,but different from the modeling ideas of denoising and decomposition ensemble,the forecasting model is built based on the perspective of feature selection by the data feature driven decomposition,and the model construction is embedded into the feature selection process.Chapter six summarizes the research of this paper and looks forward to the future research direction.For the verification of the typical data sets,the research work and innovations of this paper are mainly summarized as the following three aspects:Firstly,the existing single model mainly has two defects:On the one hand,the potential irregular fluctuation component in time series makes the forecasting model unable to more effectively capture the data generation mechanism;the other is that most of the existing single models suffer the modelling setting and unstable structure.Against the above,this paper designed a seasonal time series hybrid forecasting model by using singular spectrum analysis,support vector regression,and the cuckoo search algorithm,the improvement of the model is based on the following two aspects:? aiming at the irregular feature of the seasonal time series data,this paper designed the seasonal time-series data noise reduction process,which can perform the separability decomposition of the potential composition of time series data,and then separate the irregular component;?by introducing artificial intelligence algorithm,the defect of traditional machine learning method falling into the local optimization can be avoided.Secondly,based on the decomposition idea for solving forecasting problems,this paper built up a series of hybrid models.One side,for two or more seasonal modes,if we could achieve the separability decomposition of components contacting the specific meaning,the uncertainty caused by the mix of multiple components would be reduced;according to this thought,this paper proposed a decomposition method for decomposing seasonal time series data,this method can adequately extracted the structural component and the decomposed component has stronger interpretability and separability.On the other hand,based on the decomposition-modeling ideas,to establish the targeted solutions for sub-problems and integrate them,this paper established the single model by the analysis of characteristics of the single component,and in order to be able to effectively integrate single forecasting results,build the linear regression integration strategy,nonlinear regression integration strategy,integration strategy based on intelligent search algorithm,generating several decomposition-ensemble forecasting models.Finally,in view of most of the hybrid models,two problems are remarkable:?)when considering only the time series data itself,the selection of input features of the forecasting model can only rely on the delay feature of the original series,which is difficult to dig deep into the potential information in the data;?)Due to the integration of different methods,many hybrid models suffer the disadvantage of high computational complexity,and it is difficult to integrate the hybrid model with feature selection because feature selection is a complex search process.Aiming at the deficiency in research,this article attempts to combine the establishment of the hybrid forecasting model with feature selection process based on data driven decomposition:? based on data feature driven decomposition was proposed for constructing feature space for single variable time series,which fully holds the effective lag information for the forecasting modeling,and then more effectively mining the generation mechanism of time series data;? in order to search the optimal feature subset from the built feature space,this paper proposed a new hybrid feature selection algorithms by combing the traditional Filter with Wrapper method.On the one hand,the intelligent search algorithm was used for improving Filter method to perform the dynamic search feature subset globally.On the other hand,the learning algorithm was embedded to assess the feature subset,which makes the selection process simple,quick and optimization;? to establish a series of hybrid forecasting models for seasonal time series,three multi-scale frequency-domain decomposition methods including singular spectrum analysis,integrated empirical mode decomposition and empirical wavelet transform were adopted,Support vector regression was selected as the forecasting model for evaluating feature subset,and the global solution advantage of Cuckoo Search algorithm was utilized.This paper verifies the theoretical value and application value of the proposed series models through the study of typical case data sets,which can not only enrich and supplement the existing research system of time series forecasting methods,but also provide important references for solving the related problems.In detail,starting from data feature driven decomposition and based on multi-scale decomposition method,a systematic prediction model construction framework is established in this paper.Secondly,for the seasonal time serie forecasting,this paper the construction idea of hybrid prediction model based on decomposition-denoising,the construction idea of series of hybrid prediction models based on decomposition-ensemble,and the construction idea of hybrid prediction model based on decomposition and optimal feature selection.Finally,the serial hybrid prediction model proposed in this paper based on data feature driven decomposition systematically integrates different methods,which can not only improve the forecasting accuracy,robustness and generalization ability of models,but also mine the potential fluctuation characteristics of seasonal time series data.This paper mainly has the following two shortcomings:?the construction of the model mainly utilized the advanced multi-scale frequency-domain decomposition technology,machine learning model and artificial intelligence search algorithm.In the future research,more decomposition technologies and machine learning methods could be considered for the extended research;?this paper only used a kind of seasonal time series data.Therefore,the different sampling frequency for seasonal time series data can be taken into account in future research,and exogenous factors may be incorporated into the modeling framework to enhance the interpretability and application value.