【摘要】 高精度的地震动预测模型有助于提高地震灾害的预警和应对能力。传统回归方法构建地震动预测模型时提前设定了方程的形式，此种方法存在一定局限性，难以反映地震动传播过程中的复杂规律，因此越来越多的学者尝试应用机器学习方法构建地震动预测模型。但采用单一的机器学习算法，难以从数据中捕捉到更多规律，最终导致模型精度难以提升。本文基于日本KiK-net和K-Net强震台网收集到的俯冲带板缘地震动记录，使用Stacking模型融合策略，以LightGBM、XGBoost和CatBoost算法作为基学习器，线性回归算法作为元学习器，引入客观且高效的贝叶斯优化算法对模型进行超参数优化，最终训练并提出了一种适用于日本俯冲带板缘地震动预测的融合模型Stacking-Interface。对比分析所提出模型、单一机器学习模型和传统模型，发现机器学习模型的精度普遍高于传统模型，且相较于单一的机器学习模型，融合模型的预测能力有一定的提升；通过与实际地震动记录的对比和特征参数敏感性分析，验证了所提模型的可靠性和泛化能力。研究方法和结果能够为地震风险分析提供参考。更多还原

【Abstract】 High-precision earthquake motion prediction models contribute to the improvement of earthquake disaster early warning and response capabilities. Traditional regression methods for constructing earthquake motion prediction models rely on predefined equations, which have certain limitations and struggle to capture the complex patterns of seismic wave propagation. As a result, more and more scholars have been exploring the application of machine learning methods in constructing earthquake motion prediction models. However, using a single machine learning algorithm makes it difficult to capture more patterns from the data, leading to limited improvement in model accuracy. In this study, we utilized ground motion records collected from the KiK-net and K-Net strong motion networks in Japan. This study employed a Stacking model fusion strategy, with LightGBM, XGBoost, and CatBoost algorithms as base estimators and linear regression as the meta-estimator. Additionally, an objective and efficient Bayesian optimization algorithm was introduced to optimize the model’s hyperparameters. The proposed approach resulted in a ground motion prediction fusion model called Stacking-Interface, specifically designed for subduction interface earthquakes in Japan. Comparative analysis of the proposed model, single machine learning models, and traditional models revealed that machine learning models generally outperformed traditional models in terms of accuracy. Furthermore, the fusion model demonstrated improved predictive capabilities compared to single machine learning models. The reliability and generalization ability of the proposed model was validated through comparisons with actual ground motion records and sensitivity analysis of feature parameters. The research methodology and results presented in this paper can serve as a reference for earthquake risk analysis.更多还原