Evaluation of Artificial Neural Network Surrogate Models for Detention-Based Flood Control in Urban Areas

This paper presents the evaluation of artificial neural network (ANN) surrogate models trained on storm water management model (SWMM) simulations to estimate key hydraulic indicators associated with decentralized detention systems for urban flood control. Previous studies employing ANN surrogate models for detention structure design and optimization have primarily focused on end-of-pipe configurations and often neglected important design variables affecting system performance, such as outlet dimensions and the spatial distribution of storage units. To address these limitations, multilayer perceptron (MLP) models were developed to predict the total flood volume (TFV) and maximum outflow (MOF) while simultaneously accounting for detention capacity, outlet dimensions, storage-unit spatial distribution, and rainfall characteristics. A real urban catchment was adopted as a case study, and multiple decentralized detention scenarios were simulated under different rainfall conditions. The influence of dataset size and network architecture on predictive performance was evaluated, and the developed models were subsequently applied within a surrogate-based multiobjective optimization framework to demonstrate their practical applicability. Results indicate that the MLP models achieved robust predictive performance, with the ANN_MOF model presenting mean absolute percentage error (MAPE) values below 4% for all rainfall scenarios, while the ANN_TFV model achieved MAPE values below 15% for the majority of the evaluated rainfall scenarios. Increasing the dataset size from 20,000 to 50,000 samples yielded only marginal performance gains, whereas deeper architectures generally improved predictive accuracy until overfitting effects became evident. The surrogate-based optimization framework successfully identified nondominated solutions while reducing computational time by approximately 98.4% compared with an equivalent SWMM-based optimization procedure. Overall, the results demonstrate the potential of MLP-based surrogate models to reproduce SWMM-derived hydraulic responses while substantially reducing the computational cost of detention-system design and optimization.