Research On Time Series Forecasting Approach Based On Decomposition-ensemble Learning

Time series refers to the actual observed values of the controlled objects in a complex dynamic system at different time points.The predictive modeling of time series data is based on the known data,constructing a mathematical model that can reflect a certain dynamic relationship implicit in the data,revealing its changing laws,and predicting the future trend.Time series data forecasting is used in many fields,such as air passenger demand,exchange rate,wind speed,electricity price and carbon price,etc.However,the time series data in the complex dynamic market is often affected by many factors,which often show complex characteristics such as volatility,irregularity,and non-stationary,making accurate prediction of time series data the focus and difficulty of current research.Therefore,how to scientifically build forecasting methods and improve the accuracy of data forecasting,whether it is to dig out the development rules of dynamic systems,supplement existing theoretical knowledge,deeply understand the scientific level of dynamic system changes,or gain commercial benefits,It is of great significance to provide valuable reference opinions or strong data to support such a realistic level.The existing forecasting approaches can be summarized as econometric approaches,artificial intelligence techniques and hybrid approaches.Econometric methods mainly use stochastic equations to describe the quantitative characteristics of practical problems in a concise manner,but simple mathematical formulas cannot accurately express and deal with irregular and nonlinear complex characteristics of data in actual situations.Artificial intelligence technologies can establish a mathematical model of nonlinear representation according to the needs of practical applications,so as to solve specific problems by designing corresponding learning algorithms,but it has problems such as overfitting and unstable parameter fluctuations.In terms of accuracy and stability,a single econometric approach and artificial intelligence technology cannot meet the high requirements and high standards of nonlinear time series data forecasting.To increase the ability of forecasting approaches,a hybrid approaches based on decomposition-ensemble learning can creatively combine data decomposition,neural networks,optimization algorithms and comprehensive ensemble approaches to further improve forecasting accuracy.Earlier research results focused more on the decomposition of the original time series data to reduce its inherent complexity,while ignoring the characteristics of each decomposition component's own frequency range,stability,and complexity.In view of the above problems,this paper starts from the three aspects of raw data decomposition,component feature analysis,and predictive modeling,and constructs time series data forecasting approaches in specific fields: based on decomposition,stationarity test,andintegrated ensemble air passenger demand hybrid forecasting approach and secondary decomposition,entropy calculation,optimization algorithm and neural network hybrid forecasting approach of EU carbon prices.Empirical analysis of the established approaches using different error standards and statistical tests,the results clearly show that,in specific data areas,the two approaches established are superior to other benchmark approaches,and can provide theoretical guidance and reference opinions for scientific decision-making of individuals,enterprises and government departments.