Restricted (Penn State Only)
Putcha, Venkataramana B
Graduate Program:
Energy and Mineral Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
August 02, 2017
Committee Members:
  • Turgay Ertekin, Dissertation Advisor
  • Turgay Ertekin , Committee Chair
  • Eugene C Morgan, Committee Member
  • Kamesh Madduri, Committee Member
  • Sridhar Anandakrishnan, Outside Member
  • numerical simulation
  • machine learning
  • gas lift
  • neural networks
  • petroleum
  • production
  • optimization
As the reservoir pressure declines with time, many of the wells do not have adequate bottom-hole pressure to carry the fluids to the surface. Under such circumstances, artificial lift mechanisms must be employed. Amongst various artificial lift mechanisms, a significant proportion of wells utilize the gas-lift mechanism, which is an extension of the natural flow. In gas-lift implementation, high pressure gas is injected into the wellbore through a valve, where injected gas supports production by altering the composition and reducing the density, and increasing the velocity of the produced fluids. In order to design a gas-lift system, a study of the inflow performance of the fluid from the reservoir into the wellbore, combined with the outflow performance of the fluids from the bottom of the wellbore to the surface is necessary. For this purpose, existing technologies for optimization of gas-lift systems predominantly use empirical correlations in order to reduce the computational overhead. These systems use a single-equation based inflow performance relations and black-oil outflow performance correlations that have restricted applicability in systems where the fluid composition varies spatially and temporally. The contemporary protocols consider the oil flow rate, water cut and formation gas-liquid ratio and well productivity index at a given instant of time to calculate the optimal quantity of gas lift injection. Due to this methodology, the effects of pressure decline and subsequent variations in well performance are not adequately captured. This results in a solution which determines the maximum liquid flow rate expected for a given gas lift injection rate only for the instantaneous period at which the study has been performed. This optimal gas lift injection rate may or may not provide the maximum total output of oil over the producing life of the well. As a first step, a compositional coupled numerical reservoir and wellbore hydraulics models has been developed as a part of this work. These hard-computing tools simulate the variations in composition, pressure and production profiles of a gas lift well and its associated reservoir from inception to abandonment. One more advantage of this method is that it can predict the future performance of a well with or without the details of well production history. This capability can be useful when gas lift is introduced in a well immediately after its completion post a drilling or a work-over job. Soft computing tools have gained popularity in the petroleum industry due to their speed, simplicity, wide range of applicability, capacity to identify patterns and ability to provide inverse solutions. The fully numerical coupled reservoir-wellbore simulator developed is computationally expensive. In order to develop a faster system, firstly, an ANN based wellbore hydraulics tool is developed and coupled with the numerical reservoir simulator. The data utilized for training the ANN tool was generated using the numerical wellbore hydraulics tool. Both the numerical and ANN wellbore hydraulics models were validated against cases from the field and another compositional numerical model from the literature. The average relative deviation with respect to field data was observed to be 2.2% and 2.4% respectively for the ANN and numerical wellbore hydraulics model, respectively. When compared against another compositional numerical model, the average relative deviation for the ANN based model was observed to be between 3.3% and 7.1%, while it was between 2.3% and 8.1% for the numerical model developed in this work. While the ANN based wellbore hydraulics model maintained the accuracy of the numerical model, it outperformed its counterpart the numerical model, by four orders of magnitude in terms of speed-up. The ANN based wellbore model was also coupled with the numerical reservoir simulator. This resultant model which involves a coupled numerical-ANN system is faster than the fully numerical coupled system by about 160 times. This coupled tool was used to generate a gas lift database of cumulative oil production of a well with various reservoir and wellbore operating conditions under a range of operating gas lift injection depths and flow rates. This database was used to develop an ANN based gas lift model that is capable of generating performance curves plotting total oil produced during the producing life of a well as a function of gas lift injection rate. Blind testing of the ANN gas lift model showed an average absolute error of 16.6 % with respect to the predictions of the coupled numerical-ANN reservoir wellbore model. This fully ANN based gas lift model provided a speed-up by four orders of magnitude with respect to the coupled numerical-ANN based model. Hence, a fast, robust and versatile model has been developed for maximizing total primary oil recovery using gas lift optimization through integration of numerical and neuro-simulation.