Non-linear models for black carbon exposure modelling using air pollution datasets.

Black carbon (BC) is a product of incomplete combustion, present in urban aerosols and sourcing mainly from road traffic. Epidemiological evidence reports positive associations between BC and cardiovascular and respiratory disease. Despite this, BC is currently not regulated by the EU Air Quality Directive, and as a result BC data are not available in urban areas from reference air quality monitoring networks in many countries. To fill this gap, a machine learning approach is proposed to develop a BC proxy using air pollution datasets as an input. The proposed BC proxy is based on two machine learning models, support vector regression (SVR) and random forest (RF), using observations of particle mass and number concentrations (N), gaseous pollutants and meteorological variables as the input. Experimental data were collected from a reference station in Barcelona (Spain) over a 2-year period (2018-2019). Two months of additional data were available from a second urban site in Barcelona, for model validation. BC concentrations estimated by SVR showed a high degree of correlation with the measured BC concentrations (R2 = 0.828) with a relatively low error (RMSE = 0.48 µg/m3). Model performance was dependent on seasonality and time of the day, due to the influence of new particle formation events. When validated at the second station, performance indicators decreased (R2 = 0.633; RMSE = 1.19 µg/m3) due to the lack of N data and PM2.5 and the smaller size of the dataset (2 months). New particle formation events critically impacted model performance, suggesting that its application would be optimal in environments where traffic is the main source of ultrafine particles. Due to its flexibility, it is concluded that the model can act as a BC proxy, even based on EU-regulatory air quality parameters only, to complement experimental measurements for exposure assessment in urban areas.

Input-Adaptive Proxy for Black Carbon as a Virtual Sensor.

Missing data has been a challenge in air quality measurement. In this study, we develop an input-adaptive proxy, which selects input variables of other air quality variables based on their correlation coefficients with the output variable. The proxy uses ordinary least squares regression model with robust optimization and limits the input variables to a maximum of three to avoid overfitting. The adaptive proxy learns from the data set and generates the best model evaluated by adjusted coefficient of determination (adjR2). In case of missing data in the input variables, the proposed adaptive proxy then uses the second-best model until all the missing data gaps are filled up. We estimated black carbon (BC) concentration by using the input-adaptive proxy in two sites in Helsinki, which respectively represent street canyon and urban background scenario, as a case study. Accumulation mode, traffic counts, nitrogen dioxide and lung deposited surface area are found as input variables in models with the top rank. In contrast to traditional proxy, which gives 20-80% of data, the input-adaptive proxy manages to give full continuous BC estimation. The newly developed adaptive proxy also gives generally accurate BC (street canyon: adjR2 = 0.86-0.94; urban background: adjR2 = 0.74-0.91) depending on different seasons and day of the week. Due to its flexibility and reliability, the adaptive proxy can be further extend to estimate other air quality parameters. It can also act as an air quality virtual sensor in support with on-site measurements in the future.