Applying Machine Learning Techniques to Interpret Flow Rate, Pressure and Temperature Data From Permanent Downhole Gauges

Abstract

AbstractPermanent downhole gauges (PDGs) provide a continuous record of pressure, temperature, and sometimes flow rate during well production. The continuous record provides us rich information about the reservoir and makes PDG data a valuable source for reservoir analysis. It has been shown in previous work that the convolution kernel based data mining approach is a promising tool to interpret flow rate and pressure data from PDGs. The convolution kernel method denoises and deconvolves the pressure signal successfully without explicit breakpoint detection. However, the bottlenecks of computational efficiency and incomplete recovery of reservoir behaviors limit the application of the method to interpret real PDG data.In this paper, three different machine learning techniques were applied to flow rate - pressure interpretation. We formulated the problem into a linear regression on parameters that connect the nonlinear flow rate features with pressure targets. The linear process leads to a closed form solution, which speeds up the computation dramatically. Linear regression was shown to have the same learning quality as the convolution kernel method, and outperforms it with much less computational effort.Kernel ridge regression was applied to address the issue of incomplete recovery of reservoir behaviors. Kernel ridge regression utilizes the expanded features given by the kernel function to capture the more detailed reservoir behaviors, while controlling the prediction error using ridge regression. It was shown that kernel ridge regression recovers the full reservoir behaviors successfully, e.g. wellbore storage effect, skin effect, infinite-acting radial flow and boundary effect.Some potential uses of temperature data from PDGs are also discussed in this paper. Machine learning was shown to be able to model temperature and pressure data recorded by PDGs, even if the actual physical model is complex. This originates from the fact that by using features as an approximation, machine learning does not require perfect knowledge of the physical model. The modeling of pressure using temperature data was extended to two promising applications: pressure history reconstruction using temperature data, and the cointerpretation of temperature and pressure data when flow rate data are not available.