A061-0004
Particulate Matter Forecasting in Los Angeles County with Sparse Ground-Based Sensor Data Analytics

Wednesday, 9 December 2020
Poster
Emmanuel Cocom1, Pratyush Muthukumar1, Jeanne Holm2, Dawn Comer2, Anthony Lyons2, Irene Burga2, Christa A Hasenkopf3, Chisato Calvert3 and Mohammad Pourhomayoun1, (1)California State University Los Angeles, Department of Computer Science, Los Angeles, CA, United States, (2)City of Los Angeles, Los Angeles, CA, United States, (3)OpenAQ, Washington, DC, United States
Abstract:
In recent years, machine learning and deep learning algorithms and models have evolved to help forecast particulate matter 2.5 (PM2.5) and other air pollutants based on a variety of specific datasets. However, some of these models have limitations including an inability to employ the spatial and temporal patterns of the dataset, predicting for a single site or a set of discrete locations, and not incorporating the impact of real-time meteorological conditions as an inseparable part of the data to accurately predict the air pollutants.

Long-term exposure to concentrated PM2.5 has serious adverse health effects including decreased lung function, aggravated asthma, and premature death [1]. It has been reported that PM2.5 contributes to 4.2 million deaths across the globe [2] and the relation to COVID-19 and fatality rates is just starting to be understood. This situation has highlighted the need to accurately monitor and forecast PM2.5, especially for large metropolitan areas.

In this study, which is a partnership between NASA and the City of Los Angeles and the California State University Los Angeles, we collected and used three years of PM2.5 and meteorological ground-based sensor data for Los Angeles County. We reframed PM2.5 forecasting as a spatiotemporal forecasting problem with both spatial and temporal correlations in the data. The sparse sensor data was transformed into dense grid images, where each pixel represents a PM2.5 observation. We created snapshots of the PM2.5 and meteorological conditions of Los Angeles County. Then, we applied a novel deep convolutional recurrent neural networks, particularly deep convolutional long-short-term memory (Deep ConvLSTM) structure to discover and learn the spatial and temporal correlations to forecast PM2.5.

The developed model allows us to predict the PM2.5 level in Los Angeles County (or any other location) 5 days into the future based on the past information of PM2.5 and meteorological data. This work is planned to expand to longer term forecasts and will be open sourced for other cities to use.

<< Previous Abstract | Next Abstract >>