Abstract
Given the depletion of hydrocarbon reserves,the oil industry is developing enhanced recovery processes,including waterflooding,to increase the quantity of hydrocarbon to be extracted from each reservoir.However,one of the main challenges in assessing these processes is predicting water and oil production associated with the method,given the uncertainty,lack of reservoir information,and access to commercial software.This work proposes a methodology based on multivariate Long-Short Term Memory(LSTM)Neural Networks to predict oil and water production time-series in an oil field exploited through waterflooding.First,we used the oil and water production's time series of the producer wells in an inverted five-point injection pattern developed in a numerical reservoir simulation software based on a Colombian oil field exploited with waterflooding.In this scenario,we evaluated three multivariate LSTMs to determine the input features mat should be included.Thus,through a define-by-run process,we assessed the LSTM structures most suitable.Finally,we found that the information obtained by operational variables,such as bottom-hole pressure or water injection rate,allows significant approximations to predict oil or water production.
NSTL