Tropical cyclone (TC), one of the worst natural disasters in China, has garnered a lot of interests for both its activity characteristics and disaster loss assessment, especially in the context of global warming. In this paper, the combined weighting and k-means clustering methods are used to analyze the spatial and temporal characteristics of TC and its disaster loss in China since 2000. In addition, the disaster grade assessment model of TC based on machine learning algorithm is also constructed. The results show that the frequency of TC landing in China is in a trend of decreasing year by year, but the maximum landing wind speed has been slowly strengthening. Guangdong, Zhejiang, Fujian and Guangxi provinces are seriously affected by TC, but overall, the comprehensive disaster index shows a downward trend. Compared with the classic RF, SVM and NB algorithms, LightGBM (Light Gradient Boosting Machine) has the best performance in assessing the TC disaster loss, and the accuracy can reach 0.91. Moreover, the disaster-inducing factor is the most critical factor in the assessment model, followed by the disaster prevention and mitigation, exposure and vulnerability indicators.