西北太平洋热带气旋生成客观预测模型

该文提出一种西北太平洋热带气旋年生成活动的客观预测模型。研究大尺度环境因子对西北太平洋热带气旋年生成频次的作用，使用最小角回归算法对初始14个预测因子进行选择和降维，将资料集分为训练集（1979—2015年）和验证集（2016—2020年），建立随机森林回归模型预测热带气旋年生成频次。分析环境因子对西北太平洋热带气旋生成位置的作用，使用逐步回归算法筛选影响显著的预测因子，建立局部泊松回归模型预测热带气旋生成空间位置的概率。结果表明:随机森林回归模型可以预测西北热带气旋频次的主要变化和趋势，揭示环境因子对西北太平洋热带气旋年生成频次的影响。局部泊松回归模型对于气旋生成位置概率有一定预测能力。利用随机森林回归模型和局部泊松回归模型模拟1979—2020年西北太平洋热带气旋生成，结果与观测基本一致，可见模型可为热带气旋危险性分析提供参考。

An Objective Prediction Model for Tropical Cyclone Genesis in the Northwest Pacific

At present, the maximum predictable time of tropical cyclone using numerical model is limited to 2 weeks. Statistical forecasting methods have substantial advantages in mining the potential value of massive meteorological and oceanographic observations, surpassing the limit of numerical forecast, and providing a new way to solve the bottlenecks of tropical cyclone forecasts. A novel statistical prediction scheme is proposed for tropical cyclone annual frequency and genesis location in the Northwest Pacific. The effect of large-scale meteorological factors including sea surface temperature, the geopotential height, the humidity, the vorticity, the wind shear, the Nio3.4 index, the QBO index and the SO index on the annual frequency of tropical cyclone in Northwest Pacific are considered. Correlations between the annual frequency of tropical cyclone and the large-scale environmental variables are analyzed and 14 highly correlated predictors are selected to predict tropical cyclone frequency. The least absolute shrinkage and selection operator method is used to select 8 factors from 14 initial predictors. Then, a prediction model based on random forest is established using training samples (1979-2015) for calibration and testing samples (2016-2020) for validation. In addition, the impact of environmental conditions including the vorticity, the wind shear, the humidity, the potential intensity, the sea surface temperature anomaly and the Nio3.4 index on the formation location of tropical cyclone is also investigated. The stepwise regression algorithm is used to choose a set of independent predictive variables by an automatic procedure. The local Poisson regression is performed on training datasets using count data inside data circles whose size is determined by the method of likelihood cross validation maximation. The seasonality of tropical cyclone genesis location is added to Poisson model. Results show that the random forest model presents a major variation and trend of tropical cyclone annual frequency though there are some deviations from the fitted data. The rank importance of influence indicates the primary effect of sea surface temperature and secondary effect of atmospheric variables on tropical cyclone frequency, which further reveals the applicability of the random forest model. The local Poisson regression model predicts where the tropical cyclone is most likely to occur. This model performs well when tropical cyclone occurs in the region of the Philippine and has some deviation in some months when tropical cyclone occurs in the region of the South China Sea. This model has good performance in predicting tropical cyclone genesis location but is weak in predicting abnormal situations. Finally, these two models are used to simulate tropical cyclone genesis activity in 1979-2020. The distribution of simulated tropical cyclone genesis points is consistent with the observations. This new prediction scheme can provide support for tropical cyclone risk analysis.