Machine learning-based ozone and PM2.5 forecasting: Application to multiple AQS sites in the Pacific Northwest.

Air quality in the Pacific Northwest (PNW) of the U.S has generally been good in recent years, but unhealthy events were observed due to wildfires in summer or wood burning in winter. The current air quality forecasting system, which uses chemical transport models (CTMs), has had difficulty forecasting these unhealthy air quality events in the PNW. We developed a machine learning (ML) based forecasting system, which consists of two components, ML1 (random forecast classifiers and multiple linear regression models) and ML2 (two-phase random forest regression model). Our previous study showed that the ML system provides reliable forecasts of O3 at a single monitoring site in Kennewick, WA. In this paper, we expand the ML forecasting system to predict both O3 in the wildfire season and PM2.5 in wildfire and cold seasons at all available monitoring sites in the PNW during 2017-2020, and evaluate our ML forecasts against the existing operational CTM-based forecasts. For O3, both ML1 and ML2 are used to achieve the best forecasts, which was the case in our previous study: ML2 performs better overall (R2 = 0.79), especially for low-O3 events, while ML1 correctly captures more high-O3 events. Compared to the CTM-based forecast, our O3 ML forecasts reduce the normalized mean bias (NMB) from 7.6 to 2.6% and normalized mean error (NME) from 18 to 12% when evaluating against the observation. For PM2.5, ML2 performs the best and thus is used for the final forecasts. Compared to the CTM-based PM2.5, ML2 clearly improves PM2.5 forecasts for both wildfire season (May to September) and cold season (November to February): ML2 reduces NMB (-27 to 7.9% for wildfire season; 3.4 to 2.2% for cold season) and NME (59 to 41% for wildfires season; 67 to 28% for cold season) significantly and captures more high-PM2.5 events correctly. Our ML air quality forecast system requires fewer computing resources and fewer input datasets, yet it provides more reliable forecasts than (if not, comparable to) the CTM-based forecast. It demonstrates that our ML system is a low-cost, reliable air quality forecasting system that can support regional/local air quality management.

Development of a Machine Learning Approach for Local-Scale Ozone Forecasting: Application to Kennewick, WA.

Chemical transport models (CTMs) are widely used for air quality forecasts, but these models require large computational resources and often suffer from a systematic bias that leads to missed poor air pollution events. For example, a CTM-based operational forecasting system for air quality over the Pacific Northwest, called AIRPACT, uses over 100 processors for several hours to provide 48-h forecasts daily, but struggles to capture unhealthy O3 episodes during the summer and early fall, especially over Kennewick, WA. This research developed machine learning (ML) based O3 forecasts for Kennewick, WA to demonstrate an improved forecast capability. We used the 2017-2020 simulated meteorology and O3 observation data from Kennewick as training datasets. The meteorology datasets are from the Weather Research and Forecasting (WRF) meteorological model forecasts produced daily by the University of Washington. Our ozone forecasting system consists of two ML models, ML1 and ML2, to improve predictability: ML1 uses the random forest (RF) classifier and multiple linear regression (MLR) models, and ML2 uses a two-phase RF regression model with best-fit weighting factors. To avoid overfitting, we evaluate the ML forecasting system with the 10-time, 10-fold, and walk-forward cross-validation analysis. Compared to AIRPACT, ML1 improved forecast skill for high-O3 events and captured 5 out of 10 unhealthy O3 events, while AIRPACT and ML2 missed all the unhealthy events. ML2 showed better forecast skill for less elevated-O3 events. Based on this result, we set up our ML modeling framework to use ML1 for high-O3 events and ML2 for less elevated O3 events. Since May 2019, the ML modeling framework has been used to produce daily 72-h O3 forecasts and has provided forecasts via the web for clean air agency and public use: http://ozonematters.com/. Compared to the testing period, the operational forecasting period has not had unhealthy O3 events. Nevertheless, the ML modeling framework demonstrated a reliable forecasting capability at a selected location with much less computational resources. The ML system uses a single processor for minutes compared to the CTM-based forecasting system using more than 100 processors for hours.

Mortality associated with wildfire smoke exposure in Washington state, 2006-2017: a case-crossover study.

BACKGROUND: Wildfire events are increasing in prevalence in the western United States. Research has found mixed results on the degree to which exposure to wildfire smoke is associated with an increased risk of mortality. METHODS: We tested for an association between exposure to wildfire smoke and non-traumatic mortality in Washington State, USA. We characterized wildfire smoke days as binary for grid cells based on daily average PM2.5 concentrations, from June 1 through September 30, 2006-2017. Wildfire smoke days were defined as all days with assigned monitor concentration above a PM2.5 value of 20.4 µg/m3, with an additional set of criteria applied to days between 9 and 20.4 µg/m3. We employed a case-crossover study design using conditional logistic regression and time-stratified referent sampling, controlling for humidex. RESULTS: The odds of all-ages non-traumatic mortality with same-day exposure was 1.0% (95% CI: - 1.0 - 4.0%) greater on wildfire smoke days compared to non-wildfire smoke days, and the previous day's exposure was associated with a 2.0% (95% CI: 0.0-5.0%) increase. When stratified by cause of mortality, odds of same-day respiratory mortality increased by 9.0% (95% CI: 0.0-18.0%), while the odds of same-day COPD mortality increased by 14.0% (95% CI: 2.0-26.0%). In subgroup analyses, we observed a 35.0% (95% CI: 9.0-67.0%) increase in the odds of same-day respiratory mortality for adults ages 45-64. CONCLUSIONS: This study suggests increased odds of mortality in the first few days following wildfire smoke exposure. It is the first to examine this relationship in Washington State and will help inform local and state risk communication efforts and decision-making during future wildfire smoke events.

A Monte Carlo method for summing modeled and background pollutant concentrations.

Air quality analyses for permitting new pollution sources often involve modeling dispersion of pollutants using models such as AERMOD (American Meteorological Society/U.S. Environmental Protection Agency Regulatory Model). Representative background pollutant concentrations must be added to modeled concentrations to determine compliance with air quality standards. Summing 98th (or 99th) percentiles of two independent distributions that are unpaired in time overestimates air quality impacts and could needlessly burden sources with restrictive permit conditions. This problem is exacerbated when emissions and background concentrations peak during different seasons. Existing methods addressing this matter either require much input data, or disregard source and background seasonality, or disregard the variability of the background by utilizing a single concentration for each season, month, hour-of-day, day-of-week, or wind direction. Availability of representative background concentrations are another limitation. Here the authors report on work to improve permitting analyses, with the development of (1) daily gridded, background concentrations interpolated from 12-km CMAQ (Community Multiscale Air Quality Model) forecasts and monitored data. A two-step interpolation reproduced measured background concentrations to within 6.2%; and (2) a Monte Carlo (MC) method to combine AERMOD output and background concentrations while respecting their seasonality. The MC method randomly combines, with replacement, data from the same months and calculates 1000 estimates of the 98th or 99th percentiles. The design concentration of background + new source is the median of these 1000 estimates. It was found that the AERMOD design value (DV) + background DV lay at the upper end of the distribution of these one thousand 99th percentiles, whereas measured DVs were at the lower end. This MC method sits between these two metrics and is sufficiently protective of public health in that it overestimates design concentrations somewhat. The authors also calculated probabilities of exceeding specified thresholds at each receptor, better informing decision makers of new source air quality impacts. The MC method is executed with an R script, which is available freely upon request. IMPLICATIONS: Summing representative background pollutant concentrations with air dispersion model output using a Monte Carlo method that respects the seasonality of each provides for more robust and scientifically defensible air quality analyses in support of permit applications. This work provides applicants a method to demonstrate compliance with National Ambient Air Quality Standards and avoid emission controls that might be based on overly conservative analyses. It also calculates the probability of exceeding the standard, allowing regulators to make more informed permitting decisions.

Methoxyphenols and levoglucosan ratios in PM2.5 from wheat and Kentucky bluegrass stubble burning in eastern Washington and northern Idaho.

Eastern Washington is compromised by various pollution sources, of which agricultural burning is a particular burden. Smoke from field burning is a nuisance to nearby communities and is a concern for health. This study evaluates levoglucosan (LG) and methoxyphenols (MPs) as potential tracers for apportioning field burning smoke. PM2.5 (particulate matter < 2.5 microM m in aerodynamic diameter) samples from wheat and Kentucky bluegrass (KBG) stubble smoke were collected from chamber and field burns. The samples were analyzed for inorganic and organic tracers, including LG and 19 MPs. For the chamber experiments, the amount of LG, approximately 23 microg mg(-1) PM2.5, found in wheat and KBG stubble smoke was similar, while the total MPs was higher in wheat. Trace elements associated with soil were found in smoke samples in the field. Syringaldehyde, acetosyringone, and coniferylaldehyde were found to be the most prominent particle-phase MPs in wheat smoke, and these compounds were not always present in detectable amounts in KBG smoke. The ratio of LG/ syringaldehyde found in wheat (78 +/- 27) was higher than the same ratio reported for softwoods (22 +/- 3) and hardwoods (approximately 5). Similarly, the ratio of LG/coniferylaidehyde was higher in wheat stubble smoke (180 +/- 39) compared to that in softwoods (approximately 7) and hardwoods (approximately 8).