Operational digital twin for multi-hazard coastal flood prediction with adaptive learning: Real-time performance in the Chesapeake Bay

Coastal flooding threatens growing populations where compound hazards amplify risks. This study presents a proof-of-concept operational digital twin for single-location flood prediction at Annapolis in the Chesapeake Bay, integrating real-time NOAA and USGS data. An ensemble of Random Forest, XGBoost, Gradient Boosting, and LSTM models achieves RMSE = 0.043 ft with R 2 = 0.997 for short-term predictions, validated across 508,000 records and 369 extreme events over six years. Strong short-term accuracy largely reflects tidal autocorrelation in this semi-enclosed estuarine system. Feature importance shows 98.9% of predictive power derives from three water-level persistence variables, enabling efficient deployment. An empirical correction factor (0.87) calibrates predictions to local conditions. Key limitations include single-site validation without spatial inundation capability and no major hurricane landfall during the study period. The system complements physics-based models such as NOAA’s STOFS, which provide essential spatial detail and process understanding. The open-source implementation enables replication and community evaluation. • Single-location digital twin achieves RMSE = 0.043 ft for short-term flood prediction. • Feature analysis shows 98.9% predictive power from three water level persistence variables. • Multi-hazard integration captures 15%–35% compound flooding amplification effects. • Lightweight deployment reduces computational requirements by 95% with minimal accuracy loss. • Six-year validation demonstrates operational reliability at a single estuarine site.