Table_1_Predicting rapid intensification of tropical cyclones in the western North Pacific: a machine learning and net energy gain rate approach.docx

In this study, a machine learning (ML)-based Tropical Cyclones (TCs) Rapid Intensification (RI) prediction model has been developed by using the Net Energy Gain Rate Index (NGR). This index realistically captures the energy exchanges between the ocean and the atmosphere during the intensification of TCs. It does so by incorporating the thermal conditions of the upper ocean and using an accurate parameterization for sea surface roughness. To evaluate the effectiveness of NGR in enhancing prediction accuracy, five distinct ML algorithms were utilized: Decision Tree, Logistic Regression, Support Vector Machine, K-Nearest Neighbors, and Feed-forward Neural Network. Two sets of experiments were performed for each algorithm. The first set used only traditional predictors, while the second set incorporated NGR. The outcomes revealed that models trained with the inclusion of NGR exhibited superior performance compared to those that only used traditional predictors. Additionally, an ensemble model was developed by utilizing a hard-voting method, combining the predictions of all five individual algorithms. This ensemble approach showed a noteworthy improvement of approximately 10% in the skill score of RI prediction when NGR was included. The findings of this study emphasize the potential of NGR in refining TC intensity prediction and underline the effectiveness of ensemble ML models in RI event detection.

本研究借助净能量增益率指数（Net Energy Gain Rate Index, NGR），构建了基于机器学习（Machine Learning, ML）的热带气旋（Tropical Cyclones, TCs）快速增强（Rapid Intensification, RI）预测模型。该指数能够真实捕捉热带气旋增强过程中的海气能量交换过程，具体通过融合上层海洋热状况参数，并针对海面粗糙度采用精准的参数化方案实现。为评估NGR在提升预测精度方面的有效性，本研究选用了5种不同的机器学习算法：决策树（Decision Tree）、逻辑回归（Logistic Regression）、支持向量机（Support Vector Machine）、K近邻（K-Nearest Neighbors）以及前馈神经网络（Feed-forward Neural Network）。针对每一种算法，本研究开展了两组对照实验：第一组仅使用传统预报因子，第二组则引入NGR作为新增预报因子。实验结果表明，引入NGR训练得到的模型，其预测性能优于仅使用传统预报因子的模型。此外，本研究采用硬投票法（hard-voting method）整合上述5种单一算法的预测结果，构建了集成学习模型。当引入NGR后，该集成模型的RI预报技巧评分提升了约10%，效果显著。本研究结果凸显了NGR在优化热带气旋强度预报方面的应用潜力，同时证实了集成机器学习模型在热带气旋快速增强事件检测中的有效性。