Research On Periodic Time Series Clustering Analysis And Forecasting Method Based On Density Measure

As a representative form of data,the time series rich mining significance and mining dimensions.Different from traditional multidimensional data,time series has the following potential characteristics:periodicity in time attributes,changing monotonicity and correlation between moments.How to effectively use these common characteristics to analyze and predict the time series is of great significance.For large sample periodic time series problems,this paper discusses how to build an effective time series forecasting model based on sequence similarity,and introduces the similarity measure of time series and error adjustment mechanism based on LSSVR at both levels of preprocessing and regression analysis.Improve time series prediction accuracy.The research focuses on the problem of dynamic time series prediction.It focuses on the availability of sample subset selection and density clustering methods.It forms two research routes based on preprocessing and kernel adjustment methods to improve the accuracy of large-sample time series sequence prediction.In the preprocessing stage,the density clustering problem facing the Euclidean distance of isometric sequences is introduced and studied in depth.Based on the mutual information-entropy measure the similarity of the sample condition dimensions,a density clustering model based on the sample conditional dimension Euclidean distance was constructed.The study is based on a core view:the number of categories of periodic time series can be obtained through visualization,and the density of sample neighborhoods at each cluster center should be significantly higher than the neighborhood of its neighborhood,and the distance between high-density points as the core of clustering Not affected by nearby neighborhoods,and thus their distance from one another to the lower density points is significantly greater.Based on this idea,the presented clustering algorithm ATD is studied:the algorithm automatically discovers the density path,performs local consolidation based on the distance,forms several candidate nucleus,discovers the clustering kernel by heuristic method,and realizes conditional dimensional clustering based on Euclidean distance.The conditional dimension similarity determination method based on the DTW distance metric is introduced,and the nuclear sequence adjustment is used to achieve efficient prediction of the time series.The study first analyzes the causes of noise points and builds a kernel adjustment method based on support vector regression.By constructing the similarity measure on the time-series segments in segments,similar sample subsets for a given prediction condition are determined in the corresponding framework:If the preprocessing stage performs a similarity search on all historical samples based on the entropy measure,each update is predicted.Conditions need to complete the search process again;if the conditional dimension clustering based on Euclidean distance is adopted,each time the predicted sequence is updated,only the similarity of the current prediction condition and the set of cluster center sequences need to be measured to select the training set.For any prediction condition,the model adjusts the kernel function of the LSSVR through the error tuning function,and the prediction accuracy is improved based on the temporal distance measurement between the selected sample condition dimension and the prediction condition.The study validated the effectiveness of the experiment by analyzing prediction accuracy and parameter sensitivity.The experimental section selected the simulation data(which can be generated according to the specified number and proportion)and the real data(the number of samples varies from sample to sample)for analysis.At the same time,the data also differs in sample dimensions.The experimental results show that under the same data,the forecasting error based on the similarity judgment of the conditional dimension is smaller,and the training time is improved by an average of 5 to 10 times;while the forecasting framework based on the clustering method only needs to be trained once,compared with the existing ones.The accuracy of the full-sample prediction method has improved significantly.