Abstract: In the research of tropical cyclone forecast model, most of the conventional forecast methods adopted single model to do research. And the traditional classification methods are based on the assumption that the misclassification costs are equal which means the risk of missing report rate is equal to that of false alarm rate. Taking this into account and considering the lack of combination model with high prediction accuracy, we propose a novel combinational classification method of cost-sensitive analysis. First, the concept of the cost coefficient is presented. And then we apply SVM, GRNN, PNN and 3 decision tree algorithms to build classification models and compare the forecast accuracy as well as the cost coefficient of these models. Lastly, based on cost-sensitive analysis, three models with higher forecast accuracy and lower cost coefficient named GRNN, PNN and C5.0, are selected to build a new combination forecast model for a complex system. 2117 tropical cyclones’ meteorological indexing from 1949 to 2012 are applied to create a complex forecasting system and all the tropical cyclones are classified into two groups: Whether they will land on China or not. The final result is satisfactory that the overall accuracy is 81.81%. More importantly, the accuracy in identifying the tropical cyclones which have landed on China is 94.76%. The combination model has reduced the possibility of omitting landed tropical cyclones significantly and performed better than any single model. Therefore, the combination model is an important reference for emergency management of disaster.