Research On The Key Technology Of Sand-dust Storm Warning

Sand-dust storm is a kind of disaster weather, It brings huge disadvantages to our country in both economy and life field, forecast of sand-dust storm causes both internal and foreign researchers’interest. This paper studied for the selection of data sets and predictors to establish a sand-dust storm warning model based on Support Vector Machine, and then optimize the kernel function and the parameters of this model. The results of the study will benefit to achieve the sand-dust storm warning for a single meteorological station, and provide the scientific basis for government and relevant departments.The main research contents are as follows:(1) This paper presented a model of sand-dust storm warning which based on Support Vector Machine by using daily data of meteorological observation station and grid data of NCEP. In order to overcome the limitations of the single kernel function in the model, a Support Vector Machine classifier with combined kernel function is presented. The simulation results show that, the support vector machine model with combined kernel function improves the accuracy of forecasts and the successful limit index exceeds that of the traditional support vector machine model with single kernel function by nearly2.79%.(2) In view of the problem that the imbalanced data sets can lead the classification interface close to the minority class, a hybrid self-adaptive sampling method named SRU-AIBSMOTE algorithm is presented. This method can adaptively adjust neighboring selection strategy based on the internal distribution of the sample sets. It produces virtual minority class instances through randomized interpolation in the sphere space which consists of minority class instances and its neighbors. The random under-sampling is also applied to under-sample the majority class instances for removal of redundant data in the sample sets. Compare to other sampling algorithms, the simulation experiment results on the real datasets from Yanchi district in Ningxia show that the SRU-AIBSMOTE method has better classification performance.(3) In view of the traditional support vector machine’s weaknesses on method of choosing parameters, this paper used grid partition method, genetic algorithm and particle swarm algorithm to optimize the parameters of the model. The experimental results show that, grid partition method, genetic algorithm and particle swarm algorithm is better than manual adjustment optimization method in both of the preferred time and forecast accuracy, and the particle swarm algorithm preferred best in the situation of high precision.