Deep Learning–Based Multiobjective Optimization Framework for Stormwater Management Models

Global climate change and rapid urbanization have intensified the challenges of urban flood hazard prevention and stormwater reuse. Traditional large-scale stormwater management models (SWMMs) require innovative enhancements to achieve accurate calibration at the community scale and to enable the intelligent design of low-impact development (LID) solutions. This study proposes an intelligent framework that integrates an improved location update strategy sparrow search algorithm with a deep backpropagation neural network (AS-SSA-BPNN) for calibrating SWMM’s multiprocess parameters. Additionally, it couples a deep neural network (DNN) with an NSGA-II algorithm featuring an enhanced mutation strategy (DNN-DM-NSGA-II) to achieve multiobjective optimization of LID for construction costs, runoff volume, and pollutant load. A residential-office mixed-use community (Community A) in Xi’an, China, was selected as the study area. The intelligent framework was applied to perform multiprocess calibration, validation, LID multiobjective optimization, and design of SWMM models for Community A. Results demonstrate that the AS-SSA-BPNN model accurately establishes high-dimensional nonlinear mappings between SWMM multiprocess parameters and runoff responses. Validation confirms that the calibrated SWMM exhibits high simulation accuracy. The DNN-DM-NSGA-II model enables iterative LID layout decisions, minimizing construction costs, runoff volume, and pollutant load while efficiently computing Pareto-optimal solutions. The DNN module effectively replaces SWMM in optimization tasks to reduce computational costs, with the mean absolute error for runoff and total suspended solids (TSS) ranging from 0.231 to 0.559. Based on the Pareto frontier, a LID combination scheme balancing cost and water-quality optimization preferences was selected. This scheme achieved total runoff and TSS reduction rates of 85.9% and 86.1%, respectively.