A convolutional neural network-based proxy model for field production prediction and history matching

Production prediction has been playing an essential role in reservoir development and management. Reservoir history matching, typically involving time-consuming numerical simulations for model calibration and parameter estimation, is most applied to catch several reliable reservoir models for reservoir production prediction. This research introduces the proxy model, mainly based on the convolutional neural network (CNN), to accelerate the forward computation process and improve the history-matching efficiency. Firstly, numerous embedded discrete fracture models designed via Latin Hypercube Sampling and constructed by MATLAB Reservoir Simulation Toolbox are conducted to generate training and testing datasets with high complexity and diversity. Then, we introduced the concept of a tier system consisting of multiple cubes that record reservoir and fracture information in the form of channels in spatial space. Six channels were utilized to record fracture information, including fracture shape, time, matrix porosity, matrix permeability, fracture porosity, and fracture permeability. The cube data of the tier system served as input to our proxy model, whereas oil and water production rates served as output. Finally, we trained and tested the CNN-based proxy model to predict oil and water production accurately. We performed assisted history matching by the genetic algorithm using this artificial intelligence-based proxy model. The results indicate that the CNN-based proxy model can completely match the numerical simulation with an accuracy of 99% on synthetic production data, and the genetic algorithm-based-proxy model can accelerate the history-matching progress more efficiently than that using the traditional method.