Predicting power plant emissions using public data and machine learning

Abstract

Accurately predicting emissions from electric generating units using only publicly available information is an important but challenging task. It provides a critical link in evaluating the environmental impact of energy transitions in the power sector, makes it possible to engage stakeholders in electricity product cost modeling and electricity markets without accessing proprietary data, and serves as an auditing tool to detect anomalies in self-reported emissions data. However, the absence of proprietary data also limits the prediction accuracy. In this paper, we adopted two novel and effective strategies to overcome this challenge. First, we utilized not only the emission monitoring data (such as the Continuous Emission Monitoring System (CEMS) data) as previous studies did but also a variety of auxiliary datasets in the public domain such as the EPA Field Audit Checklist Tool (FACT). Second, we employed machine learning techniques (Extreme gradient boosting (XGBoost) and neural networks (NN)) to take advantage of the large amount of public data available. We evaluated the effectiveness of our strategies by predicting NO_x, SO₂, and CO₂ emission rates for all thermal electric generating units in New York State (NYS). Two models were developed: a full model to take a full inventory of public information and a reduced model for use in data-limited scenarios based on unit-level features that could be derived from a simplified power systems economic dispatch model. The models performed well for NO_x emission rates overall compared to the previous results, achieving R² values over 0.9 for both the full and reduced models. XGBoost and NN were shown to outperform the Linear Regression (LR) model consistently and significantly, which was employed previously to estimate unit-level emissions, especially in reduced models with a limited number of features available. The predictions of SO₂ and CO₂ emission rates showed strong overall predictive performance as well. We recommend stricter enforcement of the data reporting procedure, providing emission control operational information, and obtaining related data from multiple sources in the public domain as key steps to further improve the emission predictions.