Construction Of Time Series Combined Forecasting Model Based On Machine Learning And Its Application In Energy Field

As an important tool for human beings to understand and explore the law of development of things,time series research has made great progress since the 20th century and has been widely used in various fields of society.With the rapid development of society economy and the continuous updating of science and technology,the complexity of time series is increasing.How to mine valuable information from massive data to forecast its future development trend has become one of the hottest spots in the field of time series research.As an important branch of time series research,time series forecasting plays an important role in social and economic development,providing strong theoretical basis and data support for decision analysis and policy formulation in related fields.Following that,building accurate and efficient time series forecasting models is a core of time series forecasting research.However,in the face of time series widely existing in various industries of society,the complex data characteristics make it difficult to exist a perfect model to adapt to all application scenarios,and factors such as noise and random perturbations often make a single forecasting model perform poorly and fail to meet the practical needs of problem solving.Therefore,how to combine the advantages of multiple methods and techniques to build forecasting models with stronger performance becomes a new perspective of time series forecasting research.Combined forecasting model integrates the strengths of multiple methods to compensate for the shortcomings of individual models and improve forecasting performance.Traditional combination forecasting models forecast time series based on the forecasting performance of individual models,which improves the overall performance of the model to a certain extent.However,the construction of combined model considering only the forecasting performance of individual models does not fully exploit the advantages of the model,and the forecasting performance of the model needs to be further improved.With the continuous development of computer technology,data analysis and optimization techniques have provided new ideas for the establishment of combined models.In order to build more adaptable forecasting models based on different application scenarios,this paper combines the ideas of data pre-processing,feature selection,machine learning and swarm intelligence optimization algorithms to propose a series of novel combination forecasting models,which enriched the research in the field of time series forecasting.The forecasting performance of the model is reasonably tested and verified by applying the developed models to time series forecasting in the energy field,which is rich in research value and significance.The empirical study showed that the proposed forecasting models can effectively improve its forecasting performance and expand the scope of application.The research content of this paper is divided into seven parts.The first chapter is the introductory part of this paper,which introduces the research background,the research ideas and contents,the significance of the research,and the main innovations and shortcomings of this paper,respectively.Chapter 2 takes time series forecasting methods as a starting point to classify and sort out forecasting models based on different modeling strategies.Then,a specific application scenario is used as a foothold to summarize the current research on time series forecasting of representative indicators within the energy field and to lay the theoretical foundation for conducting subsequent studies.Chapter 3 presents the evaluation system of the time series forecasting model developed in this paper.The main part of the study is presented in Chapters 4 to 6,where a series of combination forecasting models are developed based on different strategies,respectively.Chapter 4 proposes a combination forecasting model based on data preprocessing and induced ordered weighted averaging operator to alleviate the negative impact of data noise by decomposing the time series,meanwhile,a combined strategy based on induced ordered weighted averaging operator is used to effectively improve the prediction performance of the combined model,and the experimental results verified the significance and applicability of the model that applied to wind speed prediction.Chapter 5 proposes a combination forecasting model based on chaos theory and multi-objective optimization algorithm to improve the prediction performance of the model,the quality of the input set is improved by analyzing and discriminating the chaotic characteristics of the data,and by combining phase space reconstruction for the feature selection of input variables.Following that,the MO Jay a algorithm based on the mean square error-standard deviation minimization framework is constructed for model combination,which improves the accuracy and stability of the model,and the experimental results showed that the model has achieved excellent performance in electricity price prediction.Chapter 6 proposes a combination forecasting model based on a two-stage feature selection and model screening strategy,based on the preprocessing of time series,the proposed two-stage feature selection algorithm is utilized to select the optimal input variables of forecasting model.Then the screening of sub-models is performed based on a weighted information criterion strategy,and the final forecasting results are obtained by combining the improved multi-objective optimization algorithm.In addition,the interval forecasting model is constructed to analyze the uncertainty of time series forecasting by combining the probability distribution of deterministic forecasting errors,and the empirical results showed that the model has achieved the desired results in the forecasting scenarios of electricity demand.Chapter 7 summarizes and outlooks the research work of this paper.In summary,the main findings of this paper are summarized as follows:(1)In response to the problem that time series are affected by random factors to exhibit typical irregularities and stochasticity that lead to low forecasting accuracy,the data pre-processing technique is introduced into the combined forecasting model and a combination forecasting model based on data preprocessing and induced ordered weighted averaging operator is proposed.The model combines the advantages of data noise preprocessing techniques,machine learning models,artificial intelligence optimization algorithms and induced ordered weighted averaging operator.Firstly,the complete ensemble empirical mode decomposition of adaptive noise method is utilized to preprocess the original time series and eliminate the influence of data noise on the model prediction process.Then,the pre-processed data are input into the extreme learning machine,Elman neural network and support vector machine for single prediction,subsequently,the prediction model is assigned with the induced ordered weighted averaging operator and the weights are solved with the multi-tracker optimization algorithm to achieve the combination predictions.The empirical results showed that this model effectively integrates the advantages of the single models and obtains accurate multi-step forward forecasting results when applied to wind speed forecasting,which provides technical support for the rational operation and scheduling of wind power systems.(2)In response to the lack of analyzing the chaotic characterization of time series,the feature selection of series,and the stability of the model in existing portfolio forecasting studies,a combination forecasting model based on chaos theory and multi-objective optimization algorithm is proposed.The model combines the advantages of data pre-processing,chaos analysis,machine learning models and multi-objective optimization algorithm,first analyzing the chaotic characteristics of the original time series,then combining data pre-processing algorithms and phase space reconstruction to decompose and reconstruct the time series,thereby eliminate the negative interference of data noise,and chaotic characteristics on the prediction model,so as to obtain a high-quality data input set.Then,sub-model prediction is made based on back propagation neural network,extreme learning machine,Elman neural network and long and short-term memory artificial neural network and a combination strategy based on MOJaya algorithm is constructed.In the process of assigning weights to the combined model,the multi-objective Jaya algorithm based on the mean square error-standard deviation minimization framework is set up to optimize the weight coefficients and improve the prediction accuracy and stability of the model simultaneously.The empirical results showed that the established combination forecasting model can effectively analyze and extract the chaotic characteristics of the data,and obtain accurate and stable predictions when applied to the forecasting of electricity price time series,which is of practical significance for guiding the safe operation of the electricity market.(3)In response to the lack of consideration of model screening and uncertainty prediction analysis in existing portfolio studies,this paper proposes a new combination forecasting model based on a two-stage feature selection and model screening strategy.The model incorporates the advantages of data preprocessing,feature selection,improved multi-objective optimization algorithms and uncertainty prediction analysis,and the original time series is first decomposed based on an improved complete ensemble empirical mode decomposition algorithm with adaptive noise,then the subseries is characterized and reconstructed using adaptive Lempel-Ziv complexity.Following that,the optimal input variables are screened according to the proposed two-stage feature selection algorithm,and the optimal prediction sub-model is selected according to the weighted information criterion strategy.Finally,the forecasting results are combined with the improved multi-objective salp swarm algorithm to obtain multi-step forward deterministic prediction results.Meanwhile,an interval prediction model with the improved multi-objective salp swarm algorithm is constructed based on the probability distribution of deterministic prediction errors to quantify the uncertainty of time series forecasting.The empirical results showed that the proposed model achieves effective deterministic and interval predictions when applied to power demand time series forecasting,which provides rich and comprehensive forecasting information for the stable operation of power systems and the formulation of power system regulation schemes.This research develops a series of combined forecasting models based on the modular modeling idea,which enriches the relevant research in the field of time series forecasting,while the accurate and efficient forecasting results provide a scientific reference basis for decision makers,thus helping to solve practical problems.The innovations of this paper are summarized as follows:(1)Dislike previous studies that only build a combination forecasting model,this paper presents the modular modeling idea based on different application scenarios and data characteristics.Then a series of combination forecasting models is proposed and their significance and validity that applied to specific scenarios are tested empirically.(2)This paper incorporates the data pre-processing,feature selection and optimization algorithms into the combination forecasting models.To address the existing shortcomings of the combination models,new forecasting models based on different combinatorial strategies are proposed to meet the needs of different types of data prediction,thereby improving the performance of combined models that applied to different forecasting scenarios.(3)For the proposed combination forecasting models,this paper focuses on the strategy study of combination forecasting,the proposed combination models are progressive and independent from each other,and the corresponding strategies can be selected to build combinatorial forecasting models for different application scenarios,thus providing scientific guidance for solving practical problems.