Variable Weight Combination Prediction Methods And Their Applications Based On K Nearest Neighbor And Relative Entropy

A prediction method accurately can not only provide scientific theoretical basis for the decision of the problem,but also help us to judge the development direction and trend of things in advance.A large number of theoretical studies show that the combined forecasting model can give full play to the combined advantages of single forecasting models,and has been widely used in weather,traffic,price and other specific areas of life.In order to further improve the performance and adaptability of the prediction,it is necessary to explore the variable weight combination prediction model.In this dissertation,from the perspective of K-nearest neighbor algorithm and relative entropy,the combined prediction models of two kinds of new variable weights are constructed,and the determining method of model weights is discussed.Therefore,this topic has important theoretical and application value.The first kind of model is based on the k-nearest neighbor algorithm to screen the most relevant time points with the predicted period,so as to solve the problem of information redundancy when the traditional simple average method is used to calculate the weight.The second kind of model adjusts the weight distribution of several different combination prediction models through relative entropy to obtain more ideal prediction results.The empirical analysis shows that the two models proposed in this dissertation have certain practical significance.And the main research contents of this dissertation are summarized as the following two aspects:(1)Considering the nature of similarity measurement in K-nearest neighbor pattern recognition,a variable weight combination prediction model based on K-nearest neighbor is proposed.The k nearest neighbors similar to the predicted time point are found through the integrated vector of the predicted value of the single item model,and the weight vectors corresponding to the K nearest neighbors are used to calculate the weight coefficient of the predicted time point.For the purpose of verifying the effect of the model,this dissertation builds weight combination forecasting model,the variable weight combination forecasting model based on the simple average method into the contrast experiment,the results show that the variable weight combination prediction based on the K nearest neighbor outperforms other contrast models in both prediction accuracy,fitting effect and relative error evaluation.(2)Introducing the concepts of relative entropy and probability distribution,the variable weight combination prediction model based on relative entropy is proposed.The relative entropy can be used to measure the difference between different probability distributions,and the weight coefficients of different models can be given by the relative entropy,then an ideal weight distribution of the predicted time point can be obtained.In the model structure,the combination of various optimization models ensures the reliability and robustness of the calculation results.In terms of model effect evaluation,the variable weight combination prediction model based on relative entropy in this dissertation can better guarantee the prediction accuracy of the model compared with other comparative models.Changing the size of the sliding window or eliminating the weight distribution of some relatively inferior properties does not have a significant impact on the experimental results,which further demonstrates that the system adjustment of relative entropy guarantees the robustness of the model.