耦合时空特征贡献与深度学习的洪水预报方法

A Flood Forecasting Method Coupling Spatiotemporal Feature Contributions with Deep Learning

摘要: 传统的人工神经网络模型难以对输入特征的时空贡献进行量化。本文通过融合卷积神经网络(CNN)、双向长短期记忆神经网络(BiLSTM)和注意力机制(Attention)，构建了耦合时空特征贡献与深度学习的洪水预报方法(CNN-BiLSTM-Attention)，并探讨其在建溪流域的应用效果。该方法通过捕捉水文序列中的关键时空特征，并将动态权重机制延伸到预报信息中，以达到准确预测的目的。研究结果表明：在模型建模验证过程中，CNN-BiLSTM-Attention的洪峰、洪量、峰现时间合格率均达到甲级标准，满足高精度预报需求。与传统LSTM相比，该方法能有效延长预见期，提高洪水预报精度，为建溪流域防洪决策提供支持。

Abstract: Traditional artificial neural network models struggle to quantify the spatiotemporal contributions of input features. This study proposes a flood forecasting approach that couples spatiotemporal feature contribution analysis with deep learning, termed CNN-BiLSTM-Attention, by integrating Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (BiLSTM), and Attention Mechanism. The method is applied to the Jianxi River Basin to evaluate its forecasting effectiveness. By capturing key spatiotemporal features in hydrological sequences and introducing dynamic weighting into the forecasting process, the model aims to enhance predictive accuracy. Results show that during model evaluation, the CNN-BiLSTM-Attention model meets Grade A standards for peak error, flood volume error, and peak time error, fulfilling high-precision forecasting requirements. Compared with traditional LSTM models, this approach effectively extends the forecast horizon and improves accuracy in predicting flood processes under complex spatiotemporal conditions, providing strong support for flood forecasting in the Jianxi River Basin.