MANY cities in China normally experience , the tiny particulate matter linked to some of the worst health effects from air聽pollution. Now it seems artificial intelligence could help us avoid the stuff.
Most air pollution forecasts are based on models that use maps of annual average air pollution emissions, the application of weather models and assumptions about chemical reactions. These can鈥檛 account for unforeseeable events, such as extra pollution from a traffic jam or an accidental chemical release.
Baihua Li at Loughborough University, UK, and her colleagues took a different approach to see聽whether AI could do a better聽job. They used machine learning to train a model for predicting PM2.5 levels on three聽years鈥� worth of data from Beijing, chosen due to the number of Chinese cities with a聽PM2.5 problem.
Advertisement
Around 26,000 data points聽were used, including average PM2.5 readings from four roadside pollution sensors and historical weather data. Li says the resulting forecasts for the locations near these sensors were highly accurate, based on聽how closely they matched historical observations. One hour ahead, they were 95 per cent accurate, falling to 85 per cent for 6 hours ahead.
Using this approach also helped the researchers tease out聽which factors were the most聽important in predicting dirty air. Sunlight, air pressure, the season and wind speed were聽found to be key. Wind blowing from the north-west often indicated a bad pollution episode was likely, perhaps due to the聽industrial facilities to the north-west of Beijing.
Unlike most air quality forecasts, which usually warn the public of a single likely level of pollution for their area, Li鈥檚 system provides a likely range of聽PM2.5 levels to give an idea of聽the uncertainty. The model could be adapted for other cities affected by PM2.5, such as Delhi, provided it is trained with local data. The developed system is going to be tested in Shenzhen, China, using live data from pollution sensors.
鈥淚f we can warn people of聽impending smog, then they聽can聽take steps to protect themselves,鈥� says Gary Fuller at聽King鈥檚 College London. He says machine learning could be useful where data on pollution sources is lacking. 鈥淎lerting systems around the world, such聽as that run by the UK Met Office, rely on understanding air聽pollution sources in a city. Data-driven models, such as this example for Beijing, can be used where the knowledge of local sources is sparse or uncertain.鈥�
The coronavirus outbreak means that air pollution in China is currently down, as people are staying at home. Satellite analysis reveals that levels of PM2.5 in February 2020聽were down 20 to 30 per cent versus previous years.
A separate study by last Thursday concluded that relatively low-cost air pollution sensors used by individuals and groups can, when used well, provide data meeting the same quality standards required of official roadside monitors. The EEA said聽such citizen science initiatives could 鈥渉elp raise public awareness of air quality issues in communities and trigger behavioural changes to聽reduce emissions鈥�.
