Sulfur dioxide emissions in Portugal: Prediction, estimation and air quality regulation using machine learning

Latest reports of the European Environment Agency and Agencia Portuguesa do Ambiente raise a reasonable doubt on the satisfaction of 2030 targets imposed by supranational regulation for sulfur dioxide emissions in Portugal. As such, efforts to predict the evolution and estimate statistically significant effects of covariates related to this air pollutant are recommended. Bayesian, econometrics and machine learning models are applied to predict future values of sulfur dioxide emissions in the vicinity of the most relevant thermoelectric power plant located in Portugal. Based on a multivariate time series analysis containing data that ranges from July 2017 to April 2020, several conclusions are identified. Predicted values of sulfur dioxide emissions of the five models exhibiting the lowest forecast error are strongly correlated, particularly in the interval 0.35 +/- 0.10 mu g/m(3). The application of multi-step ahead forecasting analysis and nonlinear ensemble algorithms reinforces the main result from the one-step ahead forecasting exercise, where it is demonstrated that machine learning models have a better generalization power compared to classical approaches. Additionally, an identification strategy is proposed to assess the efficacy of a firm-specific measure adopted in 2017 (i.e., qualitative improvement of the desulfurization process to reduce the level of sulfur dioxide emissions). Super learning algorithms confirm that sulfur dioxide emissions in 2017 were approximately 19% greater relative to the period 2018-2020, which allows to conclude that the effort promoted by the firm was effective. From a regulatory point of view, this study confirms that Portugal is likely to satisfy 2030 targets imposed by supranational regulation for sulfur dioxide emissions and provides useful recommendations to ensure the persistence of best air quality sustainability practices.

Automated summary

The study suggests using Bayesian, econometrics, and machine learning models to predict sulfur dioxide emissions near a thermal power plant in Portugal, concluding that machine learning models have better generalization power.