A Novel Combinational Forecasting Model of Dust Storms Based on Rare Classes Classification Algorithm

It is very important for people, who are facing the dust storm disaster, to forecast dust storm accurately. Traditional prediction algorithms tend to be suitable for discovering the rules of majority classes instead of that of minority classes or rare classes. In this paper, according to the monthly data of dust storm observation and the data of occurrence regularity of dust-storm provided by observation points in China, we have discovered that the dust storm occurrence data are merely the rare classes, while the data of non-occurrence of dust storm are the majority classes. Considering that current adopted methods are only suitable for excavating the time period of non-occurrence of dust storm rather than the regularity of dust storm occurrence, we find that the current algorithms are defective in studying rare classes, thus their accuracy is relatively low and is difficult to be improved. Taking this into account, according to the principles of rare classes algorithms, we combine SMOTE algorithm with adaboost algorithm as well as the random forest algorithm and propose a combination machine learning method applicable for the study of rare classes regularities. In this combination algorithm, we first balance the samples of different classes according to the idea of SMOTE algorithm, and then we make predictions utilizing random forest algorithm according to the adaboost system. This new combination algorithm possesses a total predictive accuracy which reaches 96.51%, a false alarm rate of zero, and a missing report rate of merely 0.28%. In general, this combination algorithm is one with practicability, effectiveness and simple feasibility, thus it can be applied and popularized to realistic dust storm forecasting.