Prediction of Frictional Pressure Drop in Flowing Deviated Gas Condensate Wells through Utilization of Artificial Neural Networks

Open Access
Seren, Doruk
Graduate Program:
Petroleum and Natural Gas Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
April 09, 2008
Committee Members:
  • Turgay Ertekin, Thesis Advisor
  • Production Engineering
  • Reservoir Engineering
  • Artificial Neural Networks
  • ANN
This thesis presents an inverse solution to the prediction of pressure drop observed in flowing deviated gas condensate wells through the use of an Artificial Neural Network (ANN) which is trained entirely with data from a deepwater, high-pressure rich gas field in the Gulf of Mexico. The field data is obtained from downhole pressure gauges (DHPGs) which are installed at a significant distance above perforations. The ANN is trained first to predict pressure drop between the DHPG and wellhead, and then adapted to predict true flowing sandface pressure (PSF). In order to validate the ANN predictions of true PSF, history matching studies are conducted on the field. The ANN designed in this study is a feedforward-backpropagation network, and requires supervised learning. During training, wellhead pressures (PWH), fluid rates, static well parameters, and node-to-node lengths, and functional links are used as inputs, and output is DHPG reading. To adapt the same network to ‘extrapolation’- that is, prediction of pressures outside the training range, with greater well deviation in the section between gauges and perforations, adapted normalization parameters had to be applied to the training data. To validate ANN predictions of PSF, three history matching studies were conducted. First, bottomhole pressure (PBH) behavior is matched to uncorrected DHPG data, then lift-table corrected PWH behavior is matched to PWH history, and lastly, PBH behavior is matched to ANN-corrected ‘true’ PSF history. The reservoir property distributions that produce satisfactory history matches for each stage are compared against the properties taken from whole cores, and the match to ANN-corrected PSF history proves to be the most reasonable match. While the ANN proved highly successful in predicting pressure drop between wellhead and DHPG, the ‘true’ PSF predictions show some irregularities and are not reliable under all conditions, due to the extrapolation of the prediction outside the range of the original training data. However, use of the ANN predictions along with good engineering judgment in a history matching study demonstrated a more reasonable reservoir property distribution than alternatives. Furthermore, the training process identified strong dependence of pressure drop on the average deviation parameter, as observed from weights. Additionally, the success of the training process demonstrates that the ANN is robust to the nonlinearities that cause difficulty in application of classic semi-empirical lift correlations in highly deviated gas condensate wells such as unsteady liquid holdup, slugging, and transitioning flow regimes. The contributions of this work are diverse. In addition to the design and validation of an artificial expert system to model flowing frictional pressure drop in wells that are unreliably modeled by other means, weaknesses of the other methods are identified in the process. These include the deviation parameter applied to the ANN training data (absent from existing correlations) and the significance of phase change within the production tubing identified by internal inconsistencies among semi-empirical correlations in liquid holdup and flow regime calculations. The importance of the ANN’s predictions are validated by their application to a history matching study, and the ANN proves useful in an extrapolation mode, unlike the more common interpolation mode often used in similar applications.