AI Hydrological Forecasting with Reservoir Operations Integration

Type of project

M.Sc.

Contact

Jakob Luchner (jalu@dhigroup.com)

Topic

Neural hydrology models (e.g., LSTMs) can perform as well as or better than traditional physics-based hydrological models in data-rich basins. However, their effectiveness drops in regulated rivers and data-scarce (ungauged) basins, largely because they lack explicit inputs for human-controlled processes (like dam releases) and are typically designed for hindcasting, not forecasting.

Possible Research Questions:

• Reservoir Data Integration: How can dynamic reservoir operations data (e.g., dam release schedules, storage levels) be effectively integrated into LSTM-based hydrological models to better simulate regulated flow patterns?  
• Forecasting Architectures: What model architectures or strategies (e.g., sequence-to-sequence LSTMs, transformers, or other recurrent models) are most effective for accurate short to medium range (7 day ahead) streamflow forecasting?  
• Generalization to Ungauged Basins: How well do the enhanced AI models (with forecasting and reservoir inputs) generalize to ungauged or data-scarce basins, and can transfer learning (pretraining on data-rich regions and fine-tuning) improve predictions in these regions?        

Relevance to DHI

Accurate hydrological modelling and forecasting is key to many advisory services delivered by DHI around the globe. Neural hydrology models potentially outperform physics based models and require little to no domain knowledge for finding the best models to apply.

What is the student's contribution

Address the posed research questions by applying neural hydrology on selected basins from DHI’s advisory projects in data rich and scarce environments around the globe.