Construction and verification of multi-model fusion computational framework for bridge backwater prediction

To address the issues of poor error controllability and low computational efficiency in single-bridge backwater calculation methods under complex topography, this paper proposes a multi-model fusion computational framework based on a “division-feedback-integration” architecture. This framework integrates the HEC-RAS one-dimensional model, the MIKE 21 two-dimensional hydrodynamic model, and three empirical backwater calculation formulas. It employs a dynamic data assimilation mechanism to achieve parameter interaction and error suppression among models, while incorporating Bayesian Model Averaging (BMA) for adaptive weight optimization. This approach enhances backwater prediction accuracy while ensuring high computational efficiency. Using six bridges along the Zhuoni section of the Tao River as case studies, the computational framework was validated under flood conditions of varying frequencies. Results indicate: The integrated framework avoids the blindness of empirical formulas and the limitations of single numerical models, providing more reasonable and mechanistically clear backwater length predictions. This approach prevents both engineering waste from overestimation and flood safety hazards from underestimation. Robustness verification indicates that when key river parameters are disturbed, the integrated framework exhibits the lowest maximum relative error and demonstrates excellent adaptability to heterogeneous pier structures. This integrated framework delivers a high-precision, high-efficiency computational method for bridge flood safety assessment, providing technical support for floodwater backwater analysis and optimized engineering design in complex river environments.