Design of a Neural Network Model for Predicting Well Performance after Water Shutoff Treatments Using Polymer Gels

Abstract Current selection of candidate wells for water shutoff by polymer gel technology is based on empirical criteria which make uncertain its result. Although sophisticated models to resemble the injection of polymer gels in a porous media are available, they require extensive knowledge of the well and reservoir, plus specific yet to be characterized gel kinetics parameters. In this work a model is proposed to predict a well future production response based on well and reservoir information that is commonly available. Neural network models were designed, trained and validated for predicting wells production performance after a polymer gel treatment for water shutoff. A total sample of 31 historical applications of gel treatments were used for training and validating the proposed networks. Historical applications gathered include different well and reservoir types, which make the model prediction validity wide broad. Neural networks with two different tasks were developed during this study, one of them with a bi-valuated response (output), and qualified as classification networks and the others with a wider range of potential output, identified as regression networks. The model aims to predict a well future oil production and the percentage of water associated to that production after the execution of a gel treatment. Training of the neural network was carried out using 90% of the wells set gathered and 10% of the wells were used for validation. An average relative error of 20% was obtained when comparing the actual production performance of wells and the neural network prediction for oil rate and water content. The proposed neural networks allow the selection of future candidate wells for gel treatments, based on its potential success. Additionally, the proposed network allows improving future gel treatments by evaluating the effect of the volume of gel to be injected on the treatment's result.