Phase II · AI-driven analytics

AI forecasting and heatwave analytics

A sequence model captures the temporal build-up of a heat spell while a gradient-boosted learner captures the spatial and calendar covariates. The two are blended into a seven-day region-wise maximum temperature forecast.

Select IMD region

Seven-day maximum temperature forecast

Solid line is the ensemble forecast; dashed line is the long-period normal

Ensemble forecast Tmax Long-period normal

Interior Peninsula

—

Peak Tmax

—°C

Normal

—°C

Departure

—°C

Classified as

—

Mean seasonal maximum temperature

Seven IMD homogeneous regions against the four IMD seasons, in °C

Warming trend, pre-monsoon Tmax

Change per decade over the 1995–2024 record

Mean heatwave days per year

Four most heat-affected regions, by decade

North West India North Central India Interior Peninsula East Coast

Model evaluation on the held-out test split Trained on 1951–2019, evaluated on 2020–2024 IMD gridded observations
Model	Task	Metric	Score	Note

Feature pipeline

Each GRD grid cell contributes a 30-day lagged temperature window, day-of-year harmonics, a region one-hot, elevation, and the rolling departure from that cell's own long-period normal.

tmax_grd  ->  quality flags  ->  regrid to region
          ->  lag window (30 d)
          ->  seasonal harmonics (sin, cos)
          ->  departure from LPA
          ->  LSTM + XGBoost ensemble
          ->  severity classifier

Retraining runs monthly on newly released observations, with the previous champion kept until the challenger beats it on the rolling test window.

Uncertainty and review

The ensemble reports a prediction interval alongside the point forecast. Where the interval crosses a severity boundary the case is routed to a duty meteorologist rather than being auto-published, which is the human-in-the-loop checkpoint required by the use case governance table.

Forecasts are additionally scored every day against the co-located IoT weather stations described on the IoT network page, so that degradation is caught before it reaches an advisory.

See how a forecast becomes an advisory