Research on similar flood recommendation model based on deep learning

Identifying hydrologically similar historical flood events is crucial for improving flood forecasting accuracy and optimizing flood control decision-making. Traditional methods often suffer from insufficient feature extraction and poor robustness of similarity measurement when dealing with high-dimensional, nonlinear flood series. To address these limitations, this paper proposes a hierarchical flood similarity recommendation model that integrates clustering analysis and a Siamese Network. The model follows a "coarse clustering-fine matching" strategy. First, based on 27 flood characteristic indicators, correlation analysis and principal component analysis (PCA) are applied for dimensionality reduction, followed by K-means clustering to classify historical floods into six typical flood patterns. Subsequently, an improved Siamese Network is constructed, combining a bidirectional long short-term memory (BiLSTM) network and fully connected layers to extract deep features, and employing a contrastive loss function to achieve refined sequence matching and similarity calculation within each flood class. The model is validated using typical flood events at hydrological stations in the upper Changjiang River Basin. Results show that the correlation coefficient between the 20 target flood events and their similar floods exceeds 0.90, the relative error of peak discharge is 1.79%, the mean Nash-Sutcliffe efficiency (NSE) coefficient is 0.85, the goodness of fit (R²) is above 0.90, and the matching error of total flood duration is zero. The proposed method effectively captures key morphological and dynamic characteristics of flood processes, demonstrating strong adaptability and robustness under multi-station conditions. This study provides a reliable and efficient tool for similar flood retrieval in real-time flood forecasting and risk management.