Graph neural network surrogates for simulation of urban drainage systems containing orifice actuators

Abstract

Study region The urban drainage system (UDS) in central Yueyang city, China. Study focus Surrogate models for UDSs can reduce computational costs. Graph neural networks (GNNs), which capture the topological structure and intrinsic relationships within UDSs, are natural candidates for constructing surrogate models. However, existing GNN surrogate models fail to incorporate orifice actuators, which are critical components in UDS management. To address this, the study develops physics-guided surrogate models based on GNN. For complex hydraulic relationships caused by orifice actuators, physics-guided surrogate models are established to enhance the hydraulic simulation by incorporating conduit capacity. Thus, two model architectures are developed: graph convolution with seq2seq (GC-S2S) and graph attention with transformer (GAT-Trans). These models are compared with the non-physics-guided GC-S2S and long short-term memory network with seq2seq (LSTM-S2S). New hydrological insights for the region (1) Compared with the non-physics-guided GC-S2S, the physics-guided model increases the NSE of conduit and orifice flow from 0.85 to 0.90 and 0.75 to 0.85, respectively. (2) The proposed GAT-Trans model demonstrates superior performance, achieving NSE of 0.94 for conduit flow, compared with 0.90 and 0.85 for GC-S2S and LSTM-S2S, respectively. (3) Compared to the Storm Water Management Model, the GAT-Trans model completes each simulation in 124 ms, corresponding to an 83-fold speedup. The proposed method can be utilized for orifice actuator application of UDS surrogate model, enhancing real-time prediction and control efficiency.