Research On The Theory And Application Of The Combination Forecasting Methods Based On Model Selection And Decomposition And Reconstruction Techniques

Combination forecasting can efficiently improve forecasting precision and reduce risk by absorbing the advantages of multiple individual models and overcoming the limitations of single models.In this dissertation,we launch the discussion on the theory of model selection,decomposition and reconstruction techniques and so on.And three combination forecasting models are constructed,two based on model selection methods,and one integrated decomposition and reconstruction techniques.This dissertation aims at building combined forecasting models that take both the performance and the efficiency into account,so that the proposed combined forecasting models achieve higher accuracy and stronger stability in different fields.Firstly,there lacks a method to determine an optimal subgroup from multiple individual models by taking the forecasting accuracy and the diversity into consideration simultaneously.In order to cope with this problem,the combination forecasting model based on the cooperative improvement rate is developed.The cooperative improvement rate is defined to reflect the performance of improvement when combining two cooperative groups.This index not only takes the precision of constitute models into consideration,but also relies on the correlation of forecasting errors among different models.Accordingly,the combined model based on the cooperative improvement rate is valid to improve the combination performance by picking a forecasting subgroup from multiple alternative models.The experimental results in air quality data analyzing and forecasting demonstrate that the combination of individual models from the selected subgroup based on the cooperative improvement rate outperforms both the best individual model and the combination of all available models.Secondly,the theory of cooperative game is applied into constructing combination forecasting methods.By regarding individual models in a forecasting group as players in a cooperative game,the contributions of different models to combination are evaluated using the Shapley value method.On this basis,the combination forecasting model based on the Shapley value is devised to identify a superior subgroup from multiple individual forecasts.It is proved that the Shapley value of a forecast is expressed as the function of four factors: the mean squared error of the forecast,the average mean squared error of other forecasts in the combination,the average mean squared error for coherence between the forecast and other forecasts in the combination,and the average mean squared error for coherence between any other two forecasts in the combination.Thus,the devised combination method integrates the information provided by the forecasting precision and the correlation of forecasting errors.The effectiveness,superiority and robustness of the combination forecasting model based on the Shapley value have been verified via three experiments.The empirical results in crude oil price,carbon price,and air quality data forecasting indicate that the proposed combined model always outperforms compared individual models and combination forecasting models.Thirdly,for solving the problem that existing methods mechanically reconstruct sub-sequences extracted from different complexity series into a preset number of items,the KMC-RLN reconstruction method,an improved version of run-length judgement(RLJ)method,is proposed.On the one hand,the KMC-RLN method preserves the advantage of the RLJ method that allows the reconstruction process to completely depend on the characteristics of data fluctuation.On the other hand,the number of components reconstructed by the KMC-RLN method is not restricted to a constant value.Thus the KMC-RLN method deals with decomposed sub-sequences more flexibly than other reconstruction methods.Lastly,the CEEMDAN decomposition method and KMC-RLN reconstruction method are applied into constructing the combination forecasting model based on the CEEMDAN-KMC-RLN method.In the devised model,the CEEMDAN algorithm is employed to decompose the original time series into several sub-sequences.And the KMC-RLN method is utilized to reconstruct the obtained sub-sequences into multiple components from the lowest frequency to the highest frequency.Then various individual models are adopted to produce forecasts for different frequency components,and the combination forecasting results are integrated by the predicted values of different components.In addition,the point forecasts are extended to interval forecasts by the quantile regression method.The characteristic of the devised model that differs from other models is making a tradeoff between a higher forecasting accuracy and a lower workload by optimizing the number of reconstructed components.The effectiveness and superiority of the devised model have been verified with the experimental results.The results in air quality data forecasting indicate that the established combined model outperforms three individual models and five hybrid models.