Machine learning for surveillance of fluid leakage from reservoir using only injection rates and bottomhole pressures

Carbon-neutral economies would require preventing the release of industrial-scale CO2 into the atmosphere by injecting into geologic formations. Large-scale injection of CO2 into deep reservoirs carries a potential for its undesired leakage into above zones, which can act as an obstacle to its large-scale implementation. Current methods for surveillance of CO2 leaks are costly and not very robust, especially the methods that simulate expected pressure behavior based on an assumed reservoir model. This study proposes a machine learning method for surveillance of fluid leakage using deconvolution response function (a non-linear function of time varying bottomhole pressure and injection rates) from injection and monitoring wells as a measure of leakage that is simulated via multivariate linear regression of all the wells present in the reservoir. Leakage is detected by comparing “expected” (baseline without leaks) deconvolution response of all monitoring wells with their “observed” deconvolution response. Three key advantages of the proposed method are that it i) uses only injection rates and bottomhole pressure data (with no reservoir or geological model), ii) is independent of physical process parameterization uncertainties, and iii) applicable to both conventional and unconventional (e.g. fractured tight formations) reservoirs with any fluid (e.g. compressible, incompressible). The proposed method is first trained to learn well history with no leakage, followed by its validation after which it can be used to detect leakage by tracking a meaningful deviation error (at least twenty times the error of no leakage base scenario over same time period) between expected well response and observed well response at all monitoring wells. The well history required for the proposed method comes directly from measurements made at wells in a real field, but in absence of field data the proposed method is illustrated through well history simulated by reservoir simulations; no such numerical simulations are required for application of this method in a real world scenario with well measurements.