Establishment And Application Of Time Series Forecasting Framework Based On Neural Network Model

Time series forecasting plays a pivotal role in social life and scientific research,which can provide important theoretical basis and strong data support for decision analysis and policy making.It is the current research focus and difficulty in the field of big data analysis and management science.How to effectively dig out the development rules and changing trends of things,and make reasonable estimates and inferences about the future state of things,and then further understand and master the evolution laws and mechanisms of the researched issues is a both significant and challenging task.As a research hotspot in academia,how to establish an effectual forecasting technique to achieve effective time series forecasting is the core innovation of time series forecasting research.Various studies have shown that there is no one best forecasting method that can be applied to all situations,while the individual forecasting model ignores the significance of data preprocessing,optimization and others to improve the forecasting model's performance,which often leads to poor forecasting results and cannot provide effective forecasting information.As a result,starting from making up for the shortcomings of individual forecasting models,research on hybrid forecasting methods with the excellent characteristics of "learn from others strong points to offset one's own weakness" can be carried out by taking full advantage of the unique merits of different methods,which can provide a new innovative perspective for time series forecasting research.Therefore,the main research work of this paper focuses on the innovation of time series forecasting technology,which is expected to construct a forecasting methodology with excellent forecasting performance and a wide range of applications.Under the background of the highlighted limitation of individual forecasting models and the emerging development trend of hybrid forecasting method in time series forecasting research,the modeling concepts of module and framework are introduced into time series forecasting,this study develops a series of effective time series forecasting framework based on neural network model to make up for the shortcomings in the existing research,which can further improve the forecasting model's performance.In order to verify the effectiveness and superiority of the developed serial hybrid forecasting frameworks,the developed hybrid forecasting frameworks are applied to energy market research with important research value and practical significance.Empirical research,comparative analysis and in-depth discussions are conducted,and the results reasonably verified that the developed serial hybrid forecasting framework can effectively achieve the time series forecasting in practical applications,which has a broad application prospect in future.The time series forecasting research carried out in this study has both theoretical and practical significance.From the theoretical perspective,this study can solve some existing problems in the current research,which can enrich related research in the field of time series forecasting and provide reference for researchers in related fields.From the practical perspective,this study has important practical significance.On the one hand,the research findings of this study are successfully applied to the forecasting of energy market,which can provide important reference for decision makers and help decision makers make correct decisions;on the other hand,the research findings of this study can also be applied to the time series forecasting research in other fields,which is of great significance to solve more practical issues in the real world.This paper is divided into six parts:Chapter one introduces forecasting and its significance,time series forecasting,research ideas and main research contents,as well as the main innovations and deficiencies of this study.In the second chapter,the main forecasting techniques in the study of time series forecasting are summarized from the perspective of main forecasting methods,and then based on the energy market,time series forecasting researches under the background of different indicator data are analyzed and commented from the perspective of time series forecasting research's application.In addition,the forecasting performance evaluation index,significance evaluation index and effectiveness evaluation index are introduced to evaluate the time series forecasting model.Chapter three develops a hybrid forecasting framework based on decomposition and reconstruction strategy.Empirical research shows that the proposed hybrid forecasting framework has excellent forecasting performance in forecasting electricity demand indicator,electricity price indicator and wind speed indicator of energy market.The fourth chapter researches the individual data preprocessing algorithm and proposes a new data preprocessing strategy called hybrid data preprocessing strategy,and finally,a hybrid forecasting framework based on decomposition and forecasting strategy is successfully developed.Empirical research shows that the hybrid forecasting framework has ideal forecasting performance in forecasting electricity price indicator of energy market.Chapter five proposes a new integration strategy and applies it to the research on time series deterministic forecasting and uncertainty analysis,and finally,a hybrid forecasting framework based on decomposition-forecasting-integration strategy is successfully developed.Empirical research shows that the hybrid forecasting framework has excellent forecasting performance in wind speed indicator of energy market.Chapter six summarizes the above research work and looks forward to future research directions.The main viewpoints and conclusions of this study include the following five aspects:First,artificial neural network models have become the fastest-growing method in the field of time series forecasting.However,the individual artificial neural network model still has some inherent shortcomings and cannot perform time series modeling and forecasting well,which limits its development and application to a large extent.In order to make up for the shortcomings of individual forecasting models and give full play to the advantages of different methods,on the basis of neural network model,this study develops a series of hybrid forecasting frameworks based on different strategies.Meanwhile,the developed hybrid forecasting framework is modularized in this study.Specifically,the data preprocessing module is constructed to preprocess the original data to reduce the non-linear and non-stationarity characteristics of the original data,which is expected to improve the performance of the hybrid forecasting framework from the data perspective.In addition,the optimization module is designed to optimize the basic forecasting model and provide support for the establishment of the optimal forecasting model,which is expected to improve the performance of the hybrid forecasting framework from the model perspective.Moreover,the basic forecasting model with better performance is selected and combined with the optimization module to develop a forecasting module,which is designed to further improve the hybrid forecasting framework's performance.Second,this study discusses the research and application of data preprocessing module in time series forecasting.To this end,a series of data preprocessing modules are constructed and applied to time series forecasting research,which can effectively reduce the non-linear and non-stationary characteristics of the original data.In chapter three,the improved complete ensemble empirical mode decomposition with adaptive noise algorithm is introduced to construct a data preprocessing module to remove noise information from the original data.In chapter four,a new data preprocessing strategy called hybrid data preprocessing strategy is proposed to construct a data preprocessing module to overcome the shortcomings of individual data preprocessing algorithm.In chapter five,the variational mode decomposition algorithm is introduced to design a data preprocessing module to provide support for the next step of modeling and forecasting.By comparing with other algorithms,the superiority of the constructed data preprocessing modules is verified,which contributes to the advancement of related research on data preprocessing technology.Next,this study discusses the research and application of the optimization module in time series forecasting.To this end,this paper introduces the multi-objective dragonfly algorithm,multi-objective grey wolf optimizer,and multi-objective salp swarm algorithm to construct optimization modules,and applies them to the time series forecasting research,which not only overcomes the shortcomings of the traditional single objective optimization algorithm,but also provide support for solving the multi-objective optimization problem in the hybrid forecasting framework.By comparing with a variety of optimization algorithms,the superiority of the constructed optimization modules is verified,which contributes to the related research of multi-objective optimization algorithms.Then,this study analyzes the selection and improvement of basic forecasting models in the serial hybrid forecasting frameworks.To this end,two types of neural network models with better performance are introduced as basic forecasting models in this study.One is the Elman neural network with memory function,which can more directly and vividly reflect the dynamic characteristics of the system.The other is the extreme learning machine model,which has the advantages of easy parameter selection,fast learning speed,good generalization ability and not easily falling into the local optimum.Subsequently,the detailed objective function is set according to the optimization goals of deterministic forecasting and uncertainty analysis,and then the optimal forecasting module is constructed in combination with the optimization module to further improve the hybrid forecasting framework's performance.Finally,this study takes the energy market as the background,and discusses the application of the developed serial hybrid forecasting frameworks in different indictor scenarios.Considering the potential influence of factors such as time,location and season on the model's forecasting performance,diverse experimental data is selected in the empirical research,which can ensure the fairness of experimental comparison and the reliability of experimental conclusions.The results show that the developed serial time series forecasting framework can obtain ideal forecasting results in practical applications,which not only make up for the shortcomings in the existing research,but also exhibits significant superiority over other comparative models and successfully realize the effective time series forecasting.The main innovations of this study are as follows:first,different from previous studies that only develop one hybrid forecasting model,this paper initially proposes the module and framework modeling concept,and then develops a series of hybrid forecasting frameworks for research on time series forecasting,which can obtain desirable results in practical applications.Second,this study focuses on strategy research,rather than proposing a specific model to solve particular problem.The research on serial forecasting frameworks is both gradual and independent,and one or more frameworks can be obtained to solve practical issues according to different application scenarios.Third,this study takes the multi-objective intelligent optimization algorithm as the main research line and introduces different multi-objective optimization algorithm to construct optimization modules,which can not only overcome the shortcomings of the traditional single objective optimization algorithm,but also provide support for solving the multi-objective optimization problem in the hybrid forecasting framework.Fourth,according to the specific hybrid forecasting framework,the optimal data preprocessing module is proposed by comparing and analyzing various data preprocessing algorithms,which ensure that the main features of time series data can be effectively identified and extracted.Fifth,in addition to selecting multiple evaluation metrics to measure the model's performance in time series forecasting,the model's significance,forecasting validity,the superiority of important components and modeling strategy,and other issues are also discussed in depth.The shortcomings of this study are as follows:First,this study only focuses on the research on single variable time series forecasting issues,while the research on time series forecasting framework beased on multiple influence factor can be carried out in the future research.Second,in the context of the energy market,the developed serial time series forecasting frameworks are applied to the short-term forecasting of different indicator data,while other application backgrounds and the research on longer forecasting horizon can be considered and carried out in the future research.Third,although the developed serial time series forecasting frameworks have achieved ideal performance,the specific algorithms and multiple parameters need to be artificially selected.As a result,how to realize the automatic selection of algorithms and the automatic setting of important parameters in the time series forecasting framework is the further research direction of this paper.