Data-driven emulation of hydrologic and hydraulic modeling for enhanced flood prediction | |
| Author | Maharjan, Saurav |
| Call Number | AIT Thesis no.WM-26-04 |
| Subject(s) | Flood forecasting--Kok River Basin--Thailand Hydrology--Data processing Machine Learning |
| Note | A thesis submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Water Engineering and Management |
| Publisher | Asian Institute of Technology |
| Abstract | Physically based hydrologic and hydraulic models remain powerful yet computationally intensive tools for flood forecasting. Their demanding data needs and slow runtimes hinder operational deployment. This study examines whether AI emulators trained directly on outputs from these models can deliver accurate, near-instantaneous flood predictions for the Kok River Basin in Chiang Rai, northern Thailand. The study first built and calibrated a HEC-HMS rainfall-runoff model and a HEC-RAS 2D hydraulic model for the basin. HEC-HMS reproduced observed streamflow with NSE values of 0.63 to 0.72. HEC-RAS produced stable 2D flood maps, validated against GISTDA satellite flood data from 2024. These two models generated 19 years of daily discharge and flood depth data used to train the AI models. An LSTM network learned to predict discharge from rainfall inputs. A 75-day lookback window produced the best results, with an NSE of 0.931 during training and 0.831 during independent testing. The model captured seasonal patterns well but underestimated extreme peaks. A ConvLSTM network learned to predict spatial flood depth maps from discharge inputs and static terrain layers including elevation, land cover, distance to channel, and a channel mask. NSE values reached 0.94 for low flows, 0.92 for medium flows, and 0.93 for high flows. RMSE ranged from 0.047 m to 0.068 m across flood magnitudes. Sensitivity testing showed the ConvLSTM responds to discharge changes in a way that matches HEC-RAS. Increasing discharge by 30% increased predicted flood depth consistently across all test locations. Adding prior flood depth as an input improved accuracy statistics but removed this physical consistency, making the model less useful for scenario analysis. The ConvLSTM model produces accurate flood predictions at a fraction of the computational cost of physical models. With careful input design, they preserve physically meaningful behavior. These models are ready for integration into real-time flood early warning systems in northern Thailand. |
| Year | 2026 |
| Type | Thesis |
| School | Faculty of Civil and Environmental Engineering (2026) |
| Department | Other Field of Studies (No Department) |
| Academic Program/FoS | Water Engineering and Management (WEM) |
| Chairperson(s) | Natthachet Tangdamrongsub |
| Examination Committee(s) | Shanmugam, Mohana Sundaram;Hassan, Nurul Husna Binti;Khadka, Dibesh |
| Scholarship Donor(s) | AIT Scholarship |
| Degree | Thesis (M. Eng.) - Asian Institute of Technology, 2026 |
 | |