Improving Air Quality Prediction with Machine Learning

2023-08-14

Machine learning models can show the locations and times where high pollution levels are most likely to occur, but they often struggle to handle additional random events.

The wildfire smoke generated during Canada's extreme fire season has prompted many people to pay attention to air quality issues and wonder what will happen in the coming days.

All air contains gaseous compounds and small particles. These gases and particles enter the nasal cavity, throat, and lungs, and even circulate in the bloodstream. As air quality worsens, they are more likely to trigger asthma and exacerbate cardiovascular and respiratory problems. In early June 2023, when wildfire smoke turned the sky over New York City orange, the number of people visiting the emergency room for asthma doubled.

In most cities, it is easy to find the daily air quality index score, which tells you when the air is considered unhealthy or even hazardous. However, predicting air quality for the next few days is not that simple.

Artificial intelligence can improve the level of air quality prediction, but research shows that combining artificial intelligence with traditional techniques can yield better results. Here's why:

How scientists predict air quality

To predict air quality for a future period of time, scientists typically rely on two main methods: chemical transport models or machine learning models. These two models produce results in completely different ways.

Chemical transport models use a large number of known chemical and physical formulas to calculate the presence and generation of air pollutants. They use data from emission inventories reported by local agencies, which list pollutants from known sources such as wildfires, traffic, or factories, as well as meteorological data that provide atmospheric information such as wind, precipitation, temperature, and solar radiation.

These models simulate the flow and chemical reactions of air pollutants. However, their simulations involve multiple variables and have a high degree of uncertainty. For example, cloud cover can alter incoming solar radiation, thereby changing photochemistry. This reduces the accuracy of the results.

In contrast, machine learning models can learn patterns within historical data to predict future air quality in a given area and then apply this knowledge to current conditions to forecast air quality for the next few days.

The drawback of machine learning models is that they do not consider chemical and physical mechanisms like chemical transport models do. Additionally, under extreme conditions such as heatwaves or wildfire events, the accuracy of machine learning predictions can be biased if the models have not been trained on such data. Therefore, although machine learning models can show when and where high pollution levels are most likely to occur, such as during peak hours near highways, they often struggle to handle additional random events, such as wildfire smoke blowing in from Canada.

Which one is better?

Scientists have determined that neither of these models alone is accurate enough, but combining the best attributes of both models can help better predict the air quality we breathe.

This combined approach is called ML-MMF and can provide scientifically-based predictions with an accuracy exceeding 90%. It is based on known physical and chemical mechanisms and can simulate the entire process from air pollution sources to your nose. Adding satellite data can help scientists inform the public more accurately about the level of air quality safety and the movement of pollutants.

Recently, we compared the predictions of these three models with actual pollution measurements. The results were astonishing: the composite model's accuracy was 66% higher than that of the chemical transport model and 12% higher than that of the standalone machine learning model.

Chemical transport models are still the most commonly used method for air quality prediction, but the application of machine learning models is gradually increasing. The conventional prediction method used by the U.S. Environmental Protection Agency's AirNow.gov website relies on machine learning. The website also compiles air quality predictions from state and local agencies, most of which use chemical transport models.

As information sources become more reliable, the composite model will become a more accurate method for predicting harmful air quality, especially during unpredictable events such as wildfire smoke.

RECENT AI NEWS

RECENT AI TOOLS