A Spatio-Temporal Prediction Model for Black Carbon Based on Automated Machine Learning

Abstract

There is increasing evidence for the health effects of particles from transportation sources, with black carbon (BC) in particular gaining attention. The high spatial and temporal variation of exposure on a small scale can make the effects of this pollutant difficult to study. We previously fit a model using the machine learning (ML) methods support vector regression and simple gradient boosting to predict daily BC concentrations in Massachusetts, Rhode Island and southern New Hampshire. However, this model only went up to 2012 and while it did use ML methods, there have been significant advances in the ML field that can produce even better prediction models.We used automated machine learning (AutoML) which is a powerful and fast tool that can rescale data and then blend up to 15 ML algorithms in order to generate predictions. AutoML also carries out Bayesian hyperparameter tuning and iteratively improves with longer runtime. A total of 49,263 BC measurements from 371 monitors over a 16 year period (2000 to 2015) were obtained from various sources and calibrated for consistency. Both land use and temporal predictors were deployed, which allowed us to capture changes in spatial patterns of BC over time. We then input our data into the auto-sklearn package in python 3 and allowed it to run over 7 hours. The model showed good accuracy with an R2 of 0.80 in the held-out test data. We have successfully built a model that can be used to estimate long-term and daily exposures to ambient BC and will be useful for research looking at various health outcomes in MA, RI and Southern NH.