Energy & Environment Lab Predicting Methane Leaks
Methane, a potent greenhouse gas with 25 times the 100-year global warming potential of carbon dioxide, continues to imperil the global climate challenge. This methane contributes aggressively to climate change, while carcinogenic gases that escape alongside it elevate the risk of adverse health outcomes. The oil and gas (O&G) sector is the second largest contributor to U.S. methane emissions: an estimated 2% (13 billion kg) of U.S. natural gas production leaks into the atmosphere each year. If leakage is indeed that high, then the climate impacts of natural gas consumption may be higher than those from coal consumption, with important implications for the transition to a net zero economy. To set good policy, regulators and firms must know how large emissions truly are, where they occur, and what forces can reduce them to economically efficient levels. Yet regulators and many firms are unable to consistently monitor emissions, instead relying on infrequent on-site facility inspections to detect leaks.
The epicenter for work on this challenge is the state of Colorado, the first state to pass regulations aimed at reducing methane emissions from its sizeable and growing shale oil and gas industry. We partnered with the Colorado Department of Public Health and Environment (CDPHE) to identify tools that can successfully address the gap in methane emissions monitoring and enforcement. We developed a machine learning model that leverages historical inspection data to predict the location of methane leaks, which CDPHE is now using as part of its annual inspection targeting strategy.
We have also launched three randomized control trials to better understand compliance behavior. Working closely with CDPHE, we first varied information provided to firms about 1) leak and inspection risk, as predicted by our machine learning model, and 2) observed leaks, then measuring the impact on facility methane emissions. Preliminary results from our evaluations suggest that notifying operators of certain facilities being at higher risk of leaks and inspections reduces the likelihood of emissions from those sources. Furthermore, the first year the machine learning model was used by CDPHE to target inspections their hit rates rose drastically, and the model was shown to be effective at predicting emission risk. If scaled, our approach could fundamentally change emissions monitoring and enforcement, and serve as a proof-of-concept on how to decrease methane emissions from the oil and gas sector.