Forecasting Dangerous Particulate Matter Using Deep Learning

Particulate matter less than 2.5 microns in diameter is dangerous to humans, as those particles can travel deep into the respiratory track. These particles increase the likelihood and severity of respiratory and cardiovascular complications, like asthma and heart disease. In the winter, wood burning stoves, a known source of 2.5 micron particulate matter (PM2.5), are a commonly used in Keene, New Hampshire. Typically, the wood smoke, and therefore particulate matter, rises and dissipates into the upper atmosphere so PM2.5 levels don’t drastically increase at ground level.

However, Keene NH is located in a shallow valley surrounded on three sides by hills and is subject to temperature inversions where a dense layer of cold air is stagnant under a layer of warm air. This cold layer traps PM2.5 produced in the area. Thus, predicting these inversions becomes vital for air quality forecasting. Unfortunately, the models that the EPA uses looks at large scale inversions at higher altitudes and can’t capture the temperature inversions that occur at lower altitudes like in the Keene valley.

We will develop a deep learning regressor to make predictions about the level of PM2.5 based on current weather conditions and time of day. Next we will predict future levels of PM2.5 to determine when inversions are likely to occur. Our research aims to predict air inversions during the winter months in Keene so that we can ask residents to stop or reduce the use of wood burning stoves at those times.