IIT Bhubaneswar has developed an AI model to improve the accuracy of prediction of heavy rainfall events with an adequate lead time. They developed a hybrid technology integrating the Weather Research and Forecasting (WRF) output into a deep learning (DL) model.

The studies were carried out in June 2023 over the complex terrain of Assam (highly vulnerable to severe flooding) and over the state of Odisha, where heavy rainfall events are highly dynamic in nature due to the landfall of multiple intense rain-bearing monsoon low-pressure systems. In Assam, the hybrid model displays prediction accuracy nearly double that of traditional ensemble models at a district level with a lead time of up to 96 hours, showcasing its remarkable performance. These innovative studies have been carried out using retrospective cases. 

To demonstrate the robustness of the technology for real-time situations over the complex terrain of Assam, in another ground-breaking study, researchers from the IIT Bhubaneswar have demonstrated a significant leap in accurately predicting heavy rainfall events over the region in real-time, using deep learning techniques. The study titled "Minimization of Forecast Error Using Deep Learning for Real-Time Heavy Rainfall Events Over Assam", published in IEEE Xplore, has revealed that integrating DL with the traditional WRF model dramatically improves forecast accuracy for heavy rainfall events in real-time, a critical advancement for this flood-prone mountainous region like Assam.

Between June 13 and 17, 2023, Assam experienced severe flooding due to heavy rainfall. The DL model could more accurately predict the spatial distribution and intensity of rains across districts scale. The research employed the WRF model to generate initial weather forecasts in real-time, which were then refined using the DL model. This method allowed for a more detailed analysis of rainfall patterns, incorporating a spatio-attention module to better capture the intricate spatial dependencies in the data. As discussed, the model was trained using data from past heavy rainfall events from multiple ensemble outputs as well as observations from the India Meteorological Department (IMD) to improve its accuracy.

Key findings 

District-Level Precision: First of its kind in real-time to improve forecast skills on a district scale.

Enhanced Prediction Accuracy: The DL model demonstrated a notable improvement in forecast accuracy, capturing 54.4% of HREs compared to the WRF model's 22.8%. The DL model also achieved a mean absolute error (MAE) of under 30 mm, significantly lower than WRF's over 50 mm MAE for days 2–4 of the forecast period.

Technological Innovation: The research introduces a U-Net model with a spatio-attention (SA) module that captures intricate spatial dependencies of rainfall features at the district scale.

Conclusion

The findings of these pioneering studies clearly demonstrate the immense potential of artificial intelligence in improving real-time weather forecasting, particularly for heavy rainfall events in complex terrains over the Indian region. This advancement is crucial for mitigating the impacts of natural disasters and public safety. These pioneering works will also serve as a guiding light in creating analogous hybrid models for other intricate topographical terrain areas such as the Western Himalayas and Western Ghats regions of India.

IIT Bhubaneswar