Assessing the efficacy of physics-informed neural networks for enhanced flood routing simulation

Abstract

Flooding has emerged as one of the most destructive natural hazards, causing extensive damage to infrastructure, livelihoods, and ecosystems. The intensifying impacts of climate change, particularly altered rainfall patterns and increased frequency of extreme precipitation events, have further amplified the severity and recurrence of both riverine and localized urban flooding. These conditions often overwhelm existing channel capacities, highlighting the need for accurate flood modeling to support critical decisions related to urban resilience and disaster preparedness. While traditional hydrodynamic models such as HEC-RAS remain widely used for flood simulation, the recent advancement of data-driven methods, particularly Machine Learning (ML), offers promising alternatives. This study explores the application of Physics-Informed Neural Networks (PINNs) as a novel approach for 1D flood routing and peak flow prediction within urban canal networks, especially in data-scarce environments. The PINN framework integrates physical laws, as partial differential equations (PDEs), directly into the model, enhancing the ability to address data scarcities. The 1D Saint-Venant equations were used as the guiding PDE within this particular model. Observed water level data, canal cross sections and rainfall data from the Urban Water Management Division of the Sri Lanka Land Development Corporation (SLLDC) were used as model inputs. The methodology was implemented on three major canal systems in Colombo, Sri Lanka: Madiwela East Diversion, Dehiwala-Wellawatta-Kirulapone Canal System, and Parliament Lake- Kiththampahuwa-Kinda-Dematagoda Canal System. The performance of the PINN model was evaluated by comparing its predicted water levels against observed data at multiple water level monitoring stations along the mentioned canals, using Nash- Sutcliffe-Efficiency (NSE) and Root Mean Square Error (RMSE) as evaluation metrics. In parallel, HEC-RAS 1D hydraulic models were developed for each canal system using the same boundary and terrain data. These models served as a benchmark for comparison with the PINN approach. The results revealed that the PINN model consistently achieved higher accuracy, outperforming HEC-RAS at intermediate water level validation locations. PINN model predictions exhibited NSE values exceeding 0.98 at training locations and above 0.95 at validation points, with corresponding RMSE values below 0.10 m (10 cm). In comparison, while HEC-RAS also demonstrated acceptable performance with NSE values above 0.90 in most cases, its RMSE values were comparatively higher, ranging from 0.10-0.17 m in some locations. These findings highlight higher accuracy of the PINN model, particularly in predicting flood peaks and spatial water level profiles. The integration of physics-based constraints with machine learning allows PINNs to outperform traditional models like HEC-RAS, especially in urban flood scenarios with sparse data availability.