Bhubaneswar: In a breakthrough for weather forecasting, researchers from IIT Bhubaneswar have developed an ensemble-based spatial attention artificial intelligence (AI)-based model that can predict rainfall at the district level far more accurately than existing systems -— a development that could strengthen rainfall warnings in Odisha.The study, published on Wednesday in the quarterly journal of the Royal Meteorological Society, comes at a time when the state is set to face more frequent and intense localised flooding in the upcoming monsoon season due to low-pressure systems, depressions, deep depressions, and cyclonic storms over the Bay of Bengal.At present, local forecasts mainly rely on physics-based dynamical models such as the WRF (Weather Research and Forecasting) model. While these models are useful, they often fail to correctly predict heavy rainfall in specific areas with adequate lead time. Often, they either miss the exact location or overestimate rainfall intensity, making district-level forecasts unreliable.For the first time, researchers from IIT Bhubaneswar’s School of Earth, Ocean, and Climate Sciences have developed an ensemble-based AI spatial attention model to improve these forecasts. They developed deep-learning models that analyse forecast data from multiple WRF simulations and learn from past rainfall patterns. An AI model identifies which type of model works best where and prioritises them based on their performance parameters like rainfall intensity, location, and lead time. For example, a few models may work best for western Odisha, and others for coastal Odisha, depending on lead time and rainfall intensity.“The system was trained using a huge dataset of over 500 simulations from 18 different storm events, including monsoon depressions and post-monsoon systems,” said the study. A key highlight of the study is how the model combines multiple deterministic models as well as ensemble forecasts.The new AI model uses an “attention mechanism” — a deep-learning technique that learns to assign greater weight to better-performing forecasts and less to weaker ones, with these weights adjusting dynamically based on the weather conditions being predicted. At the district level, the system also links large-scale weather patterns with local conditions, making the forecast more precise, the study said.The results show a sharp jump in forecast quality. According to the study, forecast error at the district level dropped from over 70 mm to below 10 mm, with the model accurately predicting different rainfall categories, including heavy and extremely heavy rain. Crucially, the system can produce hourly rainfall predictions for up to three days, capturing the day-night variation that traditional models struggle to resolve.The study noted that Odisha is highly vulnerable to heavy rainfall and floods due to weather systems forming over the Bay of Bengal. Accurate district-level forecasts can help authorities issue timely alerts, plan evacuations, and reduce damage during extreme weather events.With climate change making rainfall more unpredictable, such advanced forecasting tools could play a
