Artificial intelligence for dynamic and intelligent methane inventory

Atmospheric methane contributes to approximately 20-30% of the current global radiative forcing by greenhouse gases. Despite the potential for a 39% reduction in emissions from the oil and gas sector at no net cost, the lack of dependable emission data hinders governments from implementing timely and impactful mitigation actions aligned with the Global Methane Pledge. Existing regulations rely on national methane emission inventories, significantly underestimating methane sources across various emission sectors as revealed by recent studies. The primary cause of this discrepancy is the exclusion of super-emitters in these inventories. Super-emitters, characterized by high emission rates, collectively account for an average of 40% of total methane emissions. To implement effective regulations for reducing methane emissions, a novel, reliable, and accurate inventory methodology is needed. We propose here a framework for an innovative dynamic and intelligent inventory based on artificial intelligence tools.  The dynamic component involves the collection and automatic association, over time, of methane plume detections from satellite source points with the oil and gas infrastructures at their origin. The intelligent part of the inventory enables automatic statistical and forecasting analyses contributing to the definition of multi-level emission profiles in near real-time, spanning country, region, basin, operator, site, and infrastructure levels. The proposed framework is divided into two main parts, the first part focusing on instantiated detection of potentially methane-emitting infrastructures, without recourse to fixed inventories of oil and gas (O&G) infrastructures. As the landscape of O&G infrastructures is constantly evolving, the use of an emission inventory produced at time t can quickly become inaccurate. The principle of snapshot instantiation is essential for building up an up-to-date inventory of infrastructures especially in the context of quasi-continuous monitoring. This first part is based on the use of object detection algorithms to automatically detect and recognize O&G infrastrucutres for each methane plume detection with an accuracy of over 94%. The second part of the framework consists in matching the infrastructure closest to that of the detected plume, using the K-nearest-neighbor algorithm. Carried out successively in time, this method allows to build up a time series of the rate and frequency of methane emissions by O&G infrastructures which form the basis for methane emissions spatio-temporal analysis and forecasting. To show how this framework can be used, we present a study case that consists in estimating a methane emissions inventory for compressors, tanks and wells in the Permian Basin (USA).