Abstract
The efficiency of perforation is an important aspect in gas well since it affects near wellbore pressure drop related to turbulent flow. The perforation efficiency is correlated with non-Darcy skin that is able to be distinguished by pressure transient analysis of isochronal test (Swift et al., 1962), or evaluated from multi-rate flow test data plot coefficients (Jones et al., 1976), or type curve of single build up test following constant-rate production (Spivey et al., 2004). A simple single rate pressure transient analysis which is supported by parameters derived from historical multi rate test data was also proven to differentiate skin damage and non-Darcy skin (Aminian et al., 2007). Unfortunately there are trade-offs between accurateness and analysis time in these aforementioned methods.
Quick analysis of perforation efficiency is often needed during well completion and workover activities, to decide whether re-perforation job is required or not. To overcome the challenges of limited time for data acquisition and evaluation, an empirical relation between actual perforation length, skin damage, and laminar-turbulence flow coefficients that are obtained from short-time multi rate test is important to predict the perforation efficiency.
The empirical relation will be developed using machine learning. A simple gas reservoir model is built and then run with variations of reservoir permeability, perforation interval length, near wellbore permeability, and vertical anisotropy to generate large numbers of hypothetical multi rate test data. The data set of laminar coefficient, turbulence coefficient, absolute open flow, skin damage, and perforation length will then be trained and tested to create empirical relation using supervised regression method which will afterwards be applied to several actual field cases.
This study will elaborate the development of empirical relation of perforation efficiency with the distinct parameters obtained from simple short-time multi rate test data, what other factors will influence the empirical relation, as well as become the possible condition limit of the field application of the developed empirical relation.
Circuit convergence study using machine learning compact models
