Author(s): Kaige Chen; Kexuan Liu; Yan Long; Xiaohui Lei; Zhifeng Huang; Jiaolong Zhang; Wentao Wei

Linked Author(s): Yan Long, Xiaohui Lei

Keywords: Deep learning; Flood control; Model predictive control; Surrogate model

Abstract: Real-time application of model predictive control (MPC) for urban flood mitigation is constrained by the computational intensity of physical models. This study introduces an integrated framework incorporating a bidirectional long short-term memory (BiLSTM) surrogate into the MPC approach. This surrogate model captures bidirectional hydraulic coupling in flat river networks, serving as a rapid prediction model for global optimization within the Baishichong catchment case study. The results indicated that the BiLSTM surrogate replicates physical dynamics with a Nash-Sutcliffe efficiency exceeding 0.997, achieving a 94.3% computational speedup. The MPC strategy outperforms rule-based control by executing preemptive drawdown, effectively converting energy consumption into hydraulic buffering capacity. This mechanism eliminates water level violations at critical nodes and triples the duration of ecological water level maintenance. Such findings suggest that the deep learning-driven MPC is a robust alternative to extensive infrastructure expansion, which significantly enhances system resilience without substantial additional construction.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.55

Year: 2026