Skip to content
heiSay hello

SmokeNet - Proactive Air Quality Forecasting

Next-day PM2.5 forecasting for Melbourne, with chronological evaluation and methods for inspecting model predictions. Published at AJCAI 2025.

Published · AJCAI 2025

Lead researcher · First & corresponding author

Read the published paper Source code & documentation
SmokeNet architecture
Original project architecture · Open to inspect
Input history
48 h
Mean PM₂.₅ forecast
Next day
Held-out test year
2024

Overview

SmokeNet is a transformer-based forecasting study using Melbourne air-quality data from the OpenWeather API. It uses 48 hours of history to predict next-day mean PM2.5, then connects forecasts to an interpretable alerting prototype.

Publication and evaluation

I am the first and corresponding author of the paper published at AJCAI 2025, Springer LNAI. The published chapter contains the numerical results and their experimental setup.

Training and validation precede a held-out 2024 test year. The May 2024 smoke-event backtest is a separate evaluation, not an extra independent test year. Metrics in the paper are reported on the original PM2.5 concentration scale after inverse transforms.

Architecture

  • Custom transformer encoder with seasonal embeddings and a reconstruction loss that stabilises long-horizon error.
  • Chronological cross-validation keeps training and evaluation ordered in time.
  • Data from the OpenWeather Air Pollution API, spanning late 2020 to early 2025.

Explainability

Transparency is central to the project:

  • Integrated SHAP, Integrated Gradients, and DiCE to audit forecasts both globally and per prediction.
  • Used attribution and counterfactual analysis to inspect forecast drivers and possible changes to model inputs.

Product Surface

  • Dashboard prototype for next-day PM2.5 forecasts and explanations.
  • Proposed alert tiers mapped to the U.S. EPA AQI categories documented in the paper.

Lessons

  • The model comparison depends on the data split and evaluation protocol. I report the held-out and smoke-event results separately.
  • The original evaluation logs are not linked here. A consistency check on the paper's baseline table is pending; its numbers should not be treated as an independently reproduced benchmark.
  • Attribution and counterfactual analysis help inspect a forecast. They do not establish that the system is ready for public-safety use.