The Research And Applications Of The Ensemble Model For Time-series Forecasting

With the arrival of the era of big data,massive time-series data have been produced in all walks of life.As important and complex data objects,large-scale time-series datasets has brought both opportunities and challenges.On the one hand,abundant data mining and analysis technologies have been enriched data sources by reducing the cost of data generation and collection.On the other hand,the large scale of time-series datasets also posed serious problems and challenges.Therefore,it is the hot topic in the academic and industrial fields to seek high-performance approaches to handle the large-scale of time-series data.In general,two main ways can be improved the accuracy and robustness of forecasting method: proposing new forecasting models and integrations of existing forecasting models.Among them,ensemble models are generally performed aiming to take the strength of individual models in patterns modeling and recognition of time series forecasting.The main work of this dissertation is that four ensemble models for multi-step forward prediction have been proposed based on the existing forecasting models.The first ensemble model is the NCFM model,which uses ESN as a composite function to combine four hybrid algorithms,SSA-QPSO-ESN,SSA-NARX,SSA-CBP and SSA-QPSOLSSVM.In addition,SSA and QPSO algorithms are used to preprocess time-series data and optimize ESN parameters,respectively.The accuracy and effectiveness of the NCFM model are verified by simulating the wind speed time-series data and comparing the predicted results with those of other seven methods.The second ensemble model is the MOPSO-Combined model,which uses the MOPSO algorithm to optimize ESN parameters in order to improve both accuracy and stability of the prediction.Then,the MOPSO-ESN algorithm combines three hybrid algorithms(SA-GRNN,SAElman and SA-CBP).In addition,the sliding window mechanism is used to build samples for multistep forward predictions using time-series data.To verify the performance of MOPSO-Combined model,three sets of experiments were conducted on wind speed time-series data.The results show that this model is superior to other ten models in accuracy and stability.The third ensemble model is the CE-SOM-MOG-RELM model based on multivariate timeseries data.In order to make the noise reduction process more reasonable,this algorithm uses CEACF mechanism combining CEEMD and ACF to preprocess time-series data.In addition,the concept of clustering is also introduced in the ensemble model,and several sample clusters are obtained by clustering the sample space through SOM methods.Then,the RELM network optimized by MOGWO algorithm is used to model and identify each sample cluster.In order to verify the performance of this model,four sets of experiments have been carried out on wind speed multivariate time-series.The experimental results show that this model is superior to the comparison model both in prediction accuracy and stability.The fourth ensemble model is DLR-MogR model based on deep learning.It has become a trend for deep learning to be applied in time-series forecasting.RELM algorithm,as a combination function,optimized by MOGWO(MogR)combines three deep learning algorithms.In addition,in this ensemble model,the grid search method and the dropout mechanism are respectively used to optimize the structure of the three deep networks and to prevent overfitting.Experimental results on wind speed time-series data show that the DLR-MogR model has higher precision and more stable performance than other seven comparison models.Finally,to more quantitatively compare the performance of the four ensemble model,four types time-series data are used.The experimental results of the four models on time-series data are compared and analyzed.The main achievements and contributions of this dissertation are as follows:(1)Four ensemble algorithms for multi-step forward forecasting are proposed.(2)Take ESN as a combination function optimized by two algorithms,QPSO and MOPSO.(3)Take RELM as the combination function optimized by the MOGWO algorithm.(4)The concepts of multivariate time-series and sample clustering are introduced.(5)The CEACF mechanism is proposed for time-series preprocessing to make the noise reduction operation more reasonable.(6)Three deep learning algorithms are combined and the dropout mechanism is used to prevent overfitting in the fourth ensemble model.