Construction And Application Of The Optimal Combination Time Series Prediction Model Based On Deep Learning

With the development of time series prediction technology,data modeling,prediction.and analysis have become more efficient,and obtaining higher accuracy and stability prediction results has also become one of the research hotspots.In recent years,combination prediction models have become an important research direction in time series prediction.Combination prediction refers to the combination of multiple different prediction models to comprehensively utilize the information provided by various prediction methods and obtain the final prediction results in an appropriate manner.The existing theory and practice show that combining several single models can focus on the advantages of each model to make it more adaptive and effective.This compensation principle can provide more possibilities for models to further improve prediction performance.However,due to the complex and ever-changing characteristics,coupled with some disturbances such as noise,random disturbances,or missing values.the data often exhibits chaos,nonlinearity,and multi seasonality.The various methods that make up composite models have their own applicability and advantages and disadvantages.How to develop appropriate forms of composite prediction models based on specific situations remains a challenging task.Therefore,this study attempts to build an optimal combination time series prediction model based on weight allocation strategy,information integration strategy and deep neural network fusion strategy.Firstly,the constructed model revolves around combinatorial modeling theory,comprehensively applying methods such as feature selection algorithms,deep learning models.classical machine learning models,and multidimensional sample tools,and verifying its effectiveness in terms of model prediction accuracy,stability,and result significance.Secondly,computational intelligence methods are introduced to solve the screening of individual models and the optimization of model weight coefficients,in order to combine the predictive advantages of various models to achieve optimal prediction results,compensate for the prediction defects of individual models,and improve prediction performance.The constructed model expands the application scenarios of the combined time series prediction model,solves the problem of optimizing the construction of the combined model under different data conditions,breaks through the limitations of the utilization of raw data information,improves the shortcomings of the current combined model that only considers a single constraint and ignores diversified constraints,and makes up for the limitations of traditional combined time series prediction models.This study is divided into six chapters:The first chapter is an introduction,which elaborates on the research background and topic selection basis,research significance,research ideas and content,as well as the innovation and shortcomings of the research.Chapter two is the research foundation,including literature and theoretical foundations.In the literature foundation section,the development context of combined time series prediction models was summarized;In the theoretical foundation section,the relevant concepts,theories,and performance evaluation criteria of the combined time series prediction model are introduced.Chapters three to five concentrate the main part of the research,which focus on different data types and construct optimization combination models based on different combination forms and modeling strategies,and provide detailed explanations.Meanwhile,the constructed models are applied to data from different fields for empirical research.Specifically,Chapter three constructs a linear optimization combination prediction model based on intelligent optimization weight allocation strategy for complex time series with chaotic instability,in order to utilize competition and cooperation among multiple models to achieve optimal results;The empirical results of stock index time series prediction indicate that the constructed linear optimization combination prediction model based on weight allocation has good performance.In chapter four,a linear combination forecasting model based on information integration strategy is constructed for the time series with obvious characteristics of extreme points to accurately fit the extreme point data;The empirical results using the time series of electricity prices as an example show that the constructed combination model has good predictive performance.Chapter five constructs a nonlinear combination prediction model based on deep neural network fusion for time series containing multi-scale external information,in order to achieve deep mining and utilization of information;The application analysis results of carbon price time series show that the optimized combination model constructed has good predictive performance.Chapter six is a summary and outlook of the entire study.The research work mainly includes the following aspects:Firstly,in response to the difficulty of fitting complex time series with chaotic instability,a member model weight determination mechanism based on swarm intelligence optimization algorithm was designed,supplemented by model screening and multivariate assistance.A linear optimization combination prediction model based on intelligent optimization weight allocation strategy was proposed,effectively improving the prediction ability.The model introduces a multi-input multi-output neural network structure,which simultaneously outputs point and interval prediction results,further improving prediction efficiency.Specifically,multiple classic machine learning and deep learning models are selected as candidate models for undifferentiated prediction of stock index time series,and a model screening mechanism is used to select the optimal prediction model using diverse constraints;Using multi-objective optimization algorithms to calculate the weights of each member model,the final results of stock index prediction and interval prediction are obtained.In addition,feature selection technology is used to process the raw data and select appropriate variables to assist in prediction.The stock index is a complex time series with chaotic instability.The constructed optimal combination model is applied to the prediction of the stock index time series to verify its performance.The empirical results show that the designed linear combination prediction model based on intelligent optimization weight allocation strategy performs well,verifying the effectiveness of the model.Secondly,the existing combination prediction weight parameter solutions rely too much on prediction models,resulting in inaccurate prediction of extreme points.To solve this problem,this study designed a linear optimization combined time series prediction model based on information integration strategy,and combined data preprocessing and swarm intelligence parameter optimization methods to predict and analyze the time series.The relational operation based on the information integration operator can break through the limitation of the conventional combination model on the inaccuracy of extreme point prediction,and is more sensitive to the changes of the maximum and minimum values.When it is applied to the combination prediction problem,it is expected to obtain the results closer to the actual observations at the extreme point.Due to the extreme price and high noise characteristics of the electricity price time series,the constructed combination model is applied to the prediction of the electricity price time series to verify its performance.In order to appropriately reduce the volatility of the time series,a data preprocessing method is used to denoise the electricity price time series,and then a series with strong periodicity and regularity is used as auxiliary variables to construct appropriate input forms for training the sub model.Multi objective optimization algorithms are used to optimize the parameters of the sub model,ensuring that the model predicts the time series with the optimal structure and performance.The established combination prediction model effectively integrates the predictive advantages of individual models,compensates for the shortcomings of individual models that cannot accurately predict extreme points,and ultimately obtains more accurate point prediction and interval prediction results,proving the superiority of the model.Thirdly,for the problems of multi-scale external information screening and modeling difficulties,this study creatively constructs a combination method based on deep neural networks.At the same time,combining the advantages of two-stage feature selection technology,deep learning models,and swarm intelligence parameter optimization methods,a nonlinear optimization combination prediction model is proposed and applied to time series prediction research.Specifically,due to the many factors that affect data changes in reality,it is necessary to use effective feature selection methods to screen numerous external information variables.In this study,a two-stage feature selection method was selected to screen features of multi-scale external information,and then data input structures were constructed based on the filtered features of different scales;Use the constructed dataset to train the sub model and use optimization algorithms to adjust the parameters of the sub model to obtain the optimal model structure;The sub model prediction results are combined and trained through a deep learning network to obtain the final time series prediction results.Due to the obvious nonlinear characteristics of carbon prices and their linkage with factors such as energy markets,financial markets,climate,macroeconomics,and public attention,the constructed optimal combination model is applied to carbon price time series prediction to verify its performance.The empirical results indicate that the nonlinear optimization combination prediction model based on deep neural network fusion has superior prediction performance and efficiency.The main possible innovations of this study are as follows:(1)Establish optimized combination time series prediction models suitable for chaotic instability,multi extreme point characteristics,and multi-scale external information data,and expand the application scope of combination time series modeling.(2)In terms of model construction,focusing on combinatorial prediction theory,with deep learning as the core,combined with model screening mechanism,swarm intelligence optimization algorithm,and operator function integration form,the shortcomings of existing combinatorial models are compensated for.The prediction ability of the model is improved from the perspectives of improving prediction accuracy,increasing prediction efficiency,improving prediction stability and effectiveness,enriching the theoretical research of combinatorial time series modeling.(3)In terms of data preparation,different data preprocessing methods and feature selection methods are used to make data input more adaptive,enriching the data preparation foundation for combined time series modeling.(4)The optimal combination prediction model in this study combines multiple prediction models with diverse constraints,creatively constructing a point interval optimization mode,and solving the problem of selecting individual models and optimizing weight coefficients by introducing computational intelligence methods,providing more theoretical references for solving optimization problems.