The Study Of Time Series Interval Prediction Theory Based On Computational Intelligence

At present,most of the researches on prediction problems focus on how to improve the accuracy of point prediction methods.However,due to the uncertainty of variables and environment,prediction errors always exist and are difficult to eliminate.Therefore,in view of the fact that point prediction can only predict a certain value,interval prediction,that is,to predict the interval of future fluctuations of variables,has become a hot research topic in recent years.Interval prediction can quantify the uncertainty of the target,and get the range and probability distribution of the future fluctuation of the target.Compared with point prediction,interval prediction can provide more information for decision makers,so as to realize the security and stability control of the system.In this paper,a general interval prediction method,i.e.Knee point based multi-objective upper and lower boundary estimation method,is proposed.This method uses multi-objective optimization algorithm to optimize the parameters of the interval prediction neural network,so as to obtain the Pareto frontier of the problem.According to the idea of Knee point,the points in the Knee point region are selected as the final parameters of the neural network,and the interval prediction is carried out to obtain the most satisfactory results for the decision maker.In this paper,it is proved mathematically that there must be Knee point in the Pareto frontier for interval prediction,which lays an important theoretical foundation for this kind of algorithm.A parameter migration strategy is proposed to improve the convergence speed of the algorithm and a data sampling training strategy is proposed to improve the running speed of the algorithm.Furthermore,a multi-objective optimization algorithm,namely NSGA-II algorithm based on prediction interval preference,is proposed to effectively obtain the Pareto front of uniform distribution with interval coverage greater than the decision maker's preference threshold.Secondly,a non-parametric bootstrap probability prediction method based on envelope clustering is proposed for complex data with multi-pattern characteristics.Firstly,an envelope-based clustering algorithm is designed to cluster the data according to the fluctuation characteristics represented by the second-order envelope,and several kinds of data with different statistical characteristics are obtained.Make the predicted interval more accurate.In this paper,a large number of experiments and a number of data sets have been carried out to compare the proposed algorithm,which verifies that the proposed method can effectively achieve the construction of prediction interval.