Open Access
Chidambaram, Prasanna
Graduate Program:
Petroleum and Natural Gas Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
February 02, 2009
Committee Members:
  • Turgay Ertekin, Dissertation Advisor
  • Turgay Ertekin, Committee Chair
  • Michael Adebola Adewumi, Committee Member
  • Robert W Watson, Committee Member
  • Sridhar Anandakrishnan, Committee Member
  • Artificial neural network
  • History matching
History matching is one of the more critical steps in the reservoir performance predictions. It is during this step the reservoir parameters used in the model are adjusted until the reservoir model mimics actual reservoir behavior. A lot of times, there is no reliable way to measure some of the reservoir parameters required to build the model. This leads to gross approximation of these properties over the whole reservoir. Often these parameters are adjusted and readjusted until a good history match is obtained. Once a good history match is obtained, greater confidence can be placed on predictions made by the model. The most common method of history matching is to make numerous simulation runs with each run using a different set of model parameters. The model parameters are varied in small steps in a trial-and-error fashion between each run until the observed production data matches with simulation model production data. This process is computationally intensive and time consuming. The number of simulation runs required to obtain a good history match also depends on the initial estimates of the parameters. In this study, artificial neural networks are used to build a neuro-simulation tool for predicting properties like porosity, permeability, net pay thickness and two-phase relative permeability curves. Network uses cumulative production and pressure data as input to predict the history match parameters. Predictions made by the tool developed will give a good history match or the least serve as a good starting point to perform history match. The main advantage of using this tool is that the tool will not require an initial guess value for the parameters. This will remove the guess work involved in estimating some of the unknown parameters. The proposed artificial neural network will also reduce the actual number of simulation runs required to obtain a good history match when good estimate of model parameters are not available. The parameters predicted should provide a good history match or the least serve as good estimate for parameters that can be fine tuned to improve history match. A commercial reservoir simulator was used to generate synthetic data necessary to train and validate the artificial neural network. The neural network developed can be used to predict reservoir properties of black oil reservoirs. Two separate networks were developed. A prediction network that will predict reservoir properties and the other, a network designer, that will provide design parameters required to build the prediction network. The results of this study show that the prediction network as designed by the network designer is capable of predicting reservoir parameters within acceptable margins of error. It will also considerably reduce the number of simulation runs required to achieve a good history match, thus reducing the computational resources and time required for history matching process. The developed tool was implemented to real field data from Perry reservoir located in Brayton Fields, west of Corpus Christi, Texas. Neuro-simulation tool was able to obtain a good history match with field production data. With just 50 simulation runs made to generate the training data for the network, it was able to predict the properties of the reservoir without any need for an initial estimate of the parameters.