Inspiration
Extreme heat events pose a severe threat to public health, outdoor workers, and urban infrastructure across Pakistan's most vulnerable cities—Jacobabad, Karachi, Multan, and Lahore. Traditional weather apps only provide deterministic point predictions (e.g., "38°C tomorrow"), completely failing to capture forecast uncertainty or provide actionable safety advice. Inspired by the need for reliable climate resilience tools, we built an AI system that not only predicts heat risk with calibrated confidence bounds but also translates those bounds into emergency action plans
What it does
he Heatwave Risk Prediction & Climate Intelligence System is a dual-engine platform designed for proactive heat stress management: Uncertainty-Aware Forecasting: Predicts the next-day maximum Heat Index ($HI$) alongside statistically guaranteed 80% prediction bounds using LightGBM and Split Conformal Calibration ($q = 0.668^\circ\text{C}$). Operational Hazard Mapping: Categorizes predicted heat risk into 5 clear safety tiers (Normal, Caution, Extreme Caution, Danger, Extreme Danger) with recommended work-rest schedules.RAG Climate Agent: Features a Streamlit-integrated AI assistant that queries verified emergency guidelines to generate evidence-cited safety advisories.
How we built it
Data & Feature Engineering: Processed 2,900 station-day observations across 4 Pakistani cities and engineered 13 features, including temporal lags (hi_lag1), 7-day rolling statistics (hi_roll7), and atmospheric pressure/humidity change dynamics (hum_chg, pres_chg). Residual Learning & ML: Modeled day-to-day Heat Index changes ($\Delta y$) relative to a persistence baseline using LightGBM Quantile Regression for 10th, 50th, and 90th conditional quantiles. Conformal Calibration: Applied Split Conformal Prediction using validation non-conformity scores to adjust interval widths, ensuring guaranteed coverage.Application & RAG: Built an interactive Streamlit dashboard backed by a Retrieval-Augmented Generation agent to deliver cited climate guidance.
Challenges we ran into
Quantile Undercoverage: Initial raw quantile regression models achieved only 61.0% empirical coverage on validation data against the 80% nominal target. We solved this by implementing post-hoc Split Conformal Calibration, successfully raising coverage to 80.3% in validation and 88.2% on unseen test data.Lookahead Bias: Preventing data leakage across temporal observations required strict chronological train/validation/test splitting (70% Train, 15% Validation, 15% Out-of-Sample Test).
Accomplishments that we're proud of
High Point Accuracy: Achieved an Out-of-Sample Test MAE of 1.2418 °C across 432 unseen test samples (spanning July 5 to October 20, 2026), outperforming the baseline across all 4 cities.
Robust Interval Coverage: Delivered a calibrated prediction interval averaging 4.63 °C in width while maintaining an 88.2% empirical test coverage.
Multi-City Reliability: Maintained consistent city-level test performance across Jacobabad (1.1568 °C MAE), Karachi (1.1531 °C MAE), Multan (1.2285 °C MAE), and Lahore (1.4286 °C MAE).
What we learned
Conformal Calibration is Essential: Tree-based quantile regressors alone do not guarantee marginal coverage on small climate datasets; post-hoc calibration is critical for reliable risk estimation.Residual Learning Stability: Predicting day-to-day change ($\Delta y$) rather than raw absolute values significantly stabilizes model convergence across changing seasons
What's next for Heatwave Risk Prediction & Climate Intelligence System
Multi-Day Horizon Expansion: Extend the forecasting engine from 1-day ahead to 3-day and 7-day probabilistic heatwave outlooks.
Geospatial Thermal Mapping: Incorporate high-resolution satellite thermal imagery to provide neighborhood-level urban heat island risk maps.
Automated Emergency Alerts: Integrate SMS/WhatsApp alerting pipelines to directly dispatch heat hazard warnings and work-rest advisories to municipal workers and local authorities.
Built With
- machine-learning
- python
- streamlit
