Research On Time Series Forecasting Based On ELM Improved Layer Ensemble Architecture

A time series is a random and interrelated dynamic data sequence that changes over time.Time series forecasting(TSF)is the use of a model or a technique to predict the future values of the time series based on observed values.In recent years,TSF has been widely used in many application areas such as communications,finance,energy and so on.In the past few decades,a large number of time series prediction models and methods have been proposed.Classical prediction methods are mainly relying on linear statistical models and their improved models.However,time series tend to be characterized by nonlinear and non-stationary.Therefore,nonlinear models are more reasonable choices.The artificial neural network has strong ability of nonlinear fitting,which makes it become one of the hot topics in the nonlinear time series forecasting problem.The introduction of a series of neural network algorithms provides a powerful tool for nonlinear time series prediction implementation.But many time series in real-world are so complexity and diversity that a single system cannot deal with it alone.Also,the prediction of the diversity problem is limited by the single nonlinear model,resulting in poor generalization performance of the model.In order to improve the generalization performance of the model,there have been some attempts in developing neural network ensembles time series forecasting problems.A ensemble brings together several individual networks to improve the generalization performance of the learning system.The main issue with the ensemble approaches is to consider the accuracy and diversity of the individual networks used to build the ensemble.Unfortunately,existing ensemble algorithms consider only accuracy or diversity but not both to solve TSF problems.Aiming at these drawbacks,Recently,some scholars have proposed a layered ensemble learning structure based on MLP,which considers both accuracy and diversity of basic learners when constructing a ensemble.But there are several shortcomings associated with this layered ensemble learning structure:(1)MLP network is easy to fall into local optimum rather than global optimal,and the convergence speed of MLP network is slow witch means we need more time to training networks;(1)It is necessary to setthe K value in K-means algorithm used by the architecture,and the K-means algorithm is extremely sensitive to the selection of the initial cluster center.All of these mentioned above greatly affects LEA's application in practice.In order to overcome the two deficiencies of the above LEA,this paper introduces the extreme learning machine and the improved density peak clustering algorithm into the layered ensemble network,and then proposes an extreme learning machine(ELM)based improved layering ensemble architecture(EILEA).The first is to use ELM as a base learner,using its extremely fast learning speed and good generalization ability to improve the performance of the single learner in the ensemble network,thereby improving the performance of the entire ensemble network;secondly,we choose a new clustering algorithm-The density peak clustering algorithm(DPC)replaces the K-means algorithm used by LEA,which does not need to specify the cluster number K of clusters in advance,and can identify some sample points with complex distribution.At the same time,for the DPC algorithm,the defect of cluster center needs to be manually selected.An improved density peak clustering algorithm based on inflection point estimation(IDPC)is proposed,which can determine the cluster number and cluster center autonomously,thus improving the level of automation in model.