基于深度学习的四川省洪涝灾害损失预测模型

Deep Learning-Based Prediction Model for Flood Disaster Loss in Sichuan Province

摘要: 现有数据驱动的洪涝灾害损失模型普遍存在模型透明度不足和物理可解释性有限的问题，洪涝灾害损失指标与气象因子均表现出明显的时间序列特性。本文提出一种融合注意力机制–卷积神经网络(CNN)-长短期记忆网络(LSTM)的深度学习预测模型，用于四川省洪涝灾害损失预测。模型首先利用CNN提取多源气象因子及灾损序列的局部动态特征，再通过LSTM捕捉长期时序依赖关系，并引入注意力机制增强关键特征的贡献度，从而提高模型对非线性、非平稳灾损序列的刻画能力。基于1950~2020年四川省气象与灾损数据进行验证，结果表明在使用筛选出的9个强相关影响因子时，该模型在经济损失比与人口损失比上的平均MAPE相比单LSTM模型降低约10.5%，相比SVM模型降低约21.2%，预测精度显著提升。

Abstract: Existing data-driven flood loss prediction models often suffer from limited transparency and insufficient physical interpretability. Flood-related damage indicators and meteorological factors exhibit strong temporal patterns. To address these issues, this study proposes a deep learning model that integrates an attention mechanism, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) network for predicting flood losses in Sichuan Province. The CNN module extracts multidimensional dynamic features from meteorological factors and disaster loss sequences, and the LSTM component captures long-term temporal dependencies, while the attention mechanism further enhances informative features by adaptively learning their importance. Experiments using meteorological and disaster data from 1950 to 2020 show that, when using nine strongly correlated meteorological predictors, the proposed model reduces the mean absolute percentage error (MAPE) of economic and population loss ratios by approximately 10.5% compared with the single-LSTM model and by about 21.2% compared with the SVM model, demonstrating significantly improved prediction performance.