A Study Of Intelligent Dust Storm Forecast System Based On Data Mining

The dust storm historical data set has the characteristics of field distribution,high dimensionality, and huge data volume. Data mining is a good method used todiscover knowledge of interest right from large amount of data to supportdecision-making. So the main content of research in this paper is how to use datamining technology to intelligently forecast dust storm. The works evolved in thiscontent are listed below:1 Dimensionality of data samples is reduced using principal component analysis(PCA) and the improved clustering method. The performance of both methods isenhanced much with respect to the original dimensionality reduction (DR) method.Under the condition of guaranteed performance, the operations of PCA are simple,and processing can be finished in short time. The improved clustering method is aDR method proposed in this paper, which combines clustering and ensemble forecastmodel. Among these DR methods, its performance is best.2 The dust storm forecast models are created using the improvedbackpropagation neural network (BPNN), k-nearest neighbour, and support vectormachine (SVM). The improved BPNN uses Levenberg-Marquardt (LM) algorithm,where the weights training process can converge rapidly, and the best performance isachieved. Under the condition of keeping relatively high forecast performance,k-nearest neighbour has the best stability. SVM is implemented preliminarily.Although the result is not ideal, it shows the better ability of avoiding overfitting.3 LM network is generalized using Bayesian regularization. It can remove theinstability effect of the random initial weights on performance, and successfullymake the CSI value of the LM network stable in a small range.In the end, the deficiencies of the obtained system, such as the selection of theSVM kernel function and the adaption of its parameters, are summarized. Andfurther research scheme is given, i.e. the method of editing the typical andnon-typical samples.