We developed an ML-based air quality forecast modeling framework that consists of two independent ML models, in order to predict O3 at Kennewick, WA (Fan et al., 2022 (link)). The first ML model (ML1; Supplementary Figure 1a ) consists of a random forest classifier and a multiple linear regression model: the RandomForestClassifier and RFE functions in the Python library scikit-learn are used (Pedregosa et al., 2011 ). The second ML model (ML2; Supplementary Figure 1b ) is based on a two-phase random forest regression model: the RandomForestRegressor function in the Python library scikit-learn is used (Pedregosa et al., 2011 ). More details of our ML modeling framework are available in Dataset and Modeling Framework section in Fan et al. (2022 (link)).
In this study, we use the same ML models to predict the O3 and PM2.5 at various AQS sites in the PNW. To better fit the local conditions, the model is trained at each individual site. Hourly O3 and PM2.5 predictions are used to compute maximum daily 8-h running average (MDA8) O3 mixing ratio and 24-h averaged PM2.5 concentrations, as these are the requirements of the National Ambient Air Quality Standards (NAAQS). Due to the different sources of PM2.5 during wildfire and cold seasons in the PNW, the model is trained separately for two seasons at each site. The feature-selection module from the functions listed above are used to select the features at each site to train the models. For ML2, the weighting factors vary at each site, which are computed based on the local input data.
Given ML models can be subject to overfitting and can be sensitive to issues in the training dataset, we account for these issues in our modeling setup. To avoid overfitting, we limit five features in the model training, and use 10-time 10-fold cross-validation to evaluate our model. Our training datasets are air quality observation, which are generally imbalanced: a highly polluted event or an extremely clean event is a rare event. Haixiang et al. (2017 (link)) shows that imbalanced training data may lead a bias toward commonly observed events. To alleviate the imbalance problem, we apply several methods such as turning on the balanced_subsample option in the function of the random forest model and using multiple linear regression and second phase random forest regression in the modeling system.
In this study, we use the same ML models to predict the O3 and PM2.5 at various AQS sites in the PNW. To better fit the local conditions, the model is trained at each individual site. Hourly O3 and PM2.5 predictions are used to compute maximum daily 8-h running average (MDA8) O3 mixing ratio and 24-h averaged PM2.5 concentrations, as these are the requirements of the National Ambient Air Quality Standards (NAAQS). Due to the different sources of PM2.5 during wildfire and cold seasons in the PNW, the model is trained separately for two seasons at each site. The feature-selection module from the functions listed above are used to select the features at each site to train the models. For ML2, the weighting factors vary at each site, which are computed based on the local input data.
Given ML models can be subject to overfitting and can be sensitive to issues in the training dataset, we account for these issues in our modeling setup. To avoid overfitting, we limit five features in the model training, and use 10-time 10-fold cross-validation to evaluate our model. Our training datasets are air quality observation, which are generally imbalanced: a highly polluted event or an extremely clean event is a rare event. Haixiang et al. (2017 (link)) shows that imbalanced training data may lead a bias toward commonly observed events. To alleviate the imbalance problem, we apply several methods such as turning on the balanced_subsample option in the function of the random forest model and using multiple linear regression and second phase random forest regression in the modeling system.
Free full text: Click here
