Predicting Extreme Events in Air Quality

In this notebook, we approach the problem of predicting dangerous events in air quality from land-based sensor data in urban centres. We employ continual learning techniques to reintroduce instances of extreme events to the model and mitigate forgetting. Using an example with a sensor in Vinnytsia, this notebook produces results from a baseline LSTM model and our memory replay approach.

A decrease in air quality presents a significant hazard to the sustainability of environmental conditions in the modern world. Its significance in shaping health outcomes and the quality of life in urban areas is projected to escalate over time. Many factors, encompassing anthropogenic emissions and natural phenomena, are recognised as primary influencers contributing to the escalation of air pollution levels. Human health is particularly threatened by high amounts of pollution caused by weather events or disasters. However, these extreme events are difficult to predict using machine learning techniques due to their rapid onset and rarity.
Jack Julian (
Creative Commons Attribution-NoDerivatives 4.0 International (CC BY-ND)
Temporal Extent:
6/11/2019 - 13/2/2020