AISEHack Phase 2 - PM2.5 Forecasting

Problem

Short-term PM2.5 forecasting over India is difficult because pollution dynamics are strongly nonlinear and performance on average conditions often breaks down during severe episodes. Phase 2 explicitly evaluates both global quality and episodic behavior while restricting inference to a 10-hour lookback window.

Approach

Built an episode-aware spatiotemporal forecasting pipeline over a fixed 140x124 India grid, predicting 16 future steps from 10-hour history. Trained on representative seasonal months from 2016 and inferred on unseen 2017 months. Used a hybrid deep-learning setup inspired by operator-learning and recurrent spatiotemporal modeling, with validation geared toward both global and episode metrics.

At a glance

Tech decisions

• Episode-aware objective design

  Phase 2 scoring emphasizes both global SMAPE and behavior on episodic grid points, so training/validation targeted this trade-off directly.

• Spatiotemporal modeling over fixed grid

  The task is naturally a sequence-of-fields problem; learning spatial transport and temporal evolution jointly improves stability.

• Reproducibility via W&B benchmarks

  Baseline and final experiments were tracked as separate runs to compare improvements and avoid overfitting to isolated submissions.

• Notebook-first end-to-end pipeline

  Competition requires a single executable submission notebook producing `/kaggle/working/preds.npy` with exact shape constraints.