IIT Delhi uses AI to improve streamflow predictions for India's rivers

Researchers from the Indian Institute of Technology (IIT) Delhi have combined traditional models used to predict streamflow in India's rivers with artificial intelligence, finding that the new approach significantly improved prediction accuracy in 208 of the 220 rivers tested.

Accurate information of river flow is critical for water resources management, including irrigation scheduling, reducing flood risk, and reservoir operations.

The team, Bhanu Magotra and Manabendra Saharia, said that large-scale hydrological models often produce significant uncertainties in streamflow estimates at local scales unless extensive basin-specific calibration is performed.

Such calibration is computationally expensive and challenging to implement across a country as hydrologically diverse as India, they said.

 

Calibration refers to the technique of adjusting a model's output to better align with observed real-world data.

The AI-integrated approach, described in a paper in the journal Water Resources Research, shows how integrating artificial intelligence with traditional hydrological models can help overcome longstanding challenges in water cycle prediction, the researchers said in a statement to PTI.

They employed long short-term memory (LSTM) neural networks -- a type of AI particularly effective at recognising patterns over time -- to systematically correct river streamflow from the Indian Land Data Assimilation System (ILDAS).

The ILDAS is aimed at producing high-quality, long-term estimates over India for land surface conditions such as evapotranspiration, soil moisture, runoff, and streamflow.

The AI-integrated model was trained on at least two decades of streamflow data from 220 river gauge stations across India maintained by the Central Water Commission (CWC) under the Ministry of Jal Shakti, the researchers said.

"The framework is showcased on a national scale using a multi-model hydrologic ensemble from the Indian Land Data Assimilation System (ILDAS)," the authors wrote.

"Trained on multi-decadal data from 220 catchments across India, the framework improves Kling-Gupta Efficiency in 208 catchments, raising the national median (typical) from 0.18 (uncalibrated) to 0.60," they said.

The Kling-Gupta Efficiency is a widely used measure of a hydrological model's performance.

"It (the AI-integrated model) also reduced peak flow timing error and peak mean absolute percentage error by 25 per cent in 135 catchments," the team said.

"Our research combines LSMs (land surface models) with deep learning to improve daily streamflow predictions without needing complex adjustments for each area. We tested this method across 220 rivers in India and found it improved streamflow accuracy in 208 of them," the authors wrote.

The study presents a significant step forward in integrating traditional hydrological science with modern artificial intelligence, the researchers said.

They added that the technology can be used to develop river basin digital twins, supporting informed decision making in India.