Research On Time Series Forecasting Problems Based On Ensemble Echo State Network

Echo state network?ESN?is an efficient neural network technique for time series analysis and forecasting,and has been applied in many fields.However,in the real world,the time series data often presents many characters,such as high-dimension,nonlinearity and instability,and has various sequence patterns,which can lead to poor performance when using traditional ESN approach.Through combining multiple forecasting model,ensemble learning technology can overcome over-fitting and instability problems resulted by traditional approaches,while the ESN forecasting research based on ensemble learning strategies has not been studied in-depth.In this dissertation,with the background of time series forecasting,the ESN forecasting techniques based on ensemble learning are studied and adopted to solve practical forecasting applications,the main works of this dissertation are as follows:Firstly,because the ensemble learning methods need to integrate the basic models at scale,the computation and storage of ensemble models are resource-consuming,and the computational efficiency is rarely considered in current neural network ensemble researches.In this study,an improved Adaboost with sparse structure is proposed to combine ESN and make a sparse ensemble ESN method(Adaboost_sp-ESN).The proposed Adaboost_sp-ESN screen out inactive basic models automatically,and therefore enhances computational efficiency and ensures the forecasting accuracy.The forecasting ability of Adaboost_sp-ESN is evaluated by studying the industrial electricity consuming forecasting in China.Secondly,although ESN has powerful fitting capacity for time series data,while if the data is high-dimensional,ESN method can not extract effective and meaningful predictive features from raw data.In order to solve this problem,in this study,a new neural network ensemble method combined with stacked autoencoder?SAE?and ESN is proposed,which is named SAEN.The SAEN approach uses the deep feature learning ability of SAE and sequence fitting ability of ESN to solve high-dimensional time series forecasting tasks.The tourism demand forecasting applications based on large-scale search query data is studied here to evaluate the performance of SAEN.Lastly,because ESN can not make feature selection,the traditional ESN methods has shortcomings for multi-factor time series forecasting tasks.In this study,a feature grouping based ESN ensemble model inspared by random forest is designed,and is named RF-ESN.RF-ESN is an ensemble ESN approach that integrates feature selection ability.RF-ESN adopts random forest to grouping the external influence factors,and each feature group can best predict its corresponding randomly sampled dataset.RF-ESN trains large amount of ESNs with the selected feature groups and randomly sampled datasets,and thus can enhance its generalization ability.RF-ESN is designed for multivariate time series forecasting problems,and can evaluate the importance of influence factors with the help of the feature selection path of regression trees in the random forest,which remedies the shortcoming of neural network ensemble models on feature selection.In this study,the WTI crude oil price is forecasted based on the proposed RF-ESN approach,and meanwhile,the influence factors impacting the crude oil price is analiyzed based on the findings of RF-ESN.