Neural networks to locate and quantify fugitive natural gas leaks for a MIR detection system

Abstract

Abstract Fugitive natural gas leaks from abnormal operations or failed containment at oil and gas production fields and gas supply chains are a significant source of atmospheric methane (CH4), a potent greenhouse gas. This realization of significant CH4 leaks has stimulated efforts at mitigation using cost effective detection methods. US DOE's ARPA-E MONITOR program has supported transformational research to locate and quantify fugitive methane leaks at natural gas facilities in order to achieve a 90% reduction in CH4 emissions. This paper describes the development, application, and evaluation of a MONITOR program sponsored source attribution technology solution, wherein an artificial neural network (ANN) infers as outputs, a leak location and release rate given inputs of measured sonic anemometer wind velocities and methane sensor time series. We discuss development of our ANN model, its training, computational implementation and use to design effective sampling strategies over a range of source and sensor locations, and meteorological conditions. We report on the deployment and performance of our ALFaLDS in blind tests at Colorado State University's METEC well pad facility in Ft. Collins. Test results demonstrate that our ALFaLDS locates the engineered methane leaks to better than 8% of the domain length scale and has the ability to quantify fluxes, more accurately at high flux rates, less so at low rates. Our ANN overestimates leak source rates by an average scale factor of about 1.83 for small pads and by an average of 1.77 for larger domains. The scale factor can be used operationally but needs more research to assess its generality. This capability of leak location with high skill, speed and accuracy at moderate cost promises new automatic affordable sampling of fugitive gas leaks at well pads and oil and gas fields.