Validating Hydraulic Fracturing Properties in Reservoir Simulation Using Artificial Neural Networks

Abstract

Abstract This paper presents a workflow to validate the hydraulic fracture properties in the reservoir simulator after the fracture job has been completed. The workflow incorporates artificial neural networks (ANN) as the engine behind the validation process. The required hydraulic fracture parameters that need to be fed to the reservoir simulator are mainly the fracture half-length, fracture permeability, and fracture width. These parameters constitute the fracture conductivity and are used to match the production and pressure profiles after the fracture job is completed. If the reservoir simulation model has a satisfactory history match, the model response after the fracture job should not drastically deviate from the observed well response. If that deviation occurs, the developed ANN is introduced to help in figuring out a set of fracture conductivity parameters that match the observed data. The process starts by running the reservoir simulator to generate several data sets with variable fracture conductivity parameters to be used for training and testing the ANN. The ANN then outputs the fracture conductivity parameters based on the observed pressure and production profiles. The suggested output of the ANN is eventually validated with the reservoir simulator to check for the results accuracy. This workflow has been theoretically tested on a synthetic reservoir model with heterogeneous reservoir properties. A fracking job of up to five stages is performed on wells in a low permeability oil reservoir. Results show the ANN is able to reproduce the fracture conductivity parameters with up to 84% accuracy when the stimulated well has been in production for one year. This number rises to 96% when additional year of production data is collected. The availability of different training algorithms makes the artificial neural networks a fast and reliable tool to analyze the relationships between various parameters. As the results of this paper suggest, the reservoir simulation model can be updated with the aid of ANN and without compromising the captured physics during the history match.