History matching a large reservoir model with high well count and long production history might be a challenging and time-consuming task, especially if it's done using only global multipliers as advised by common practice. Using local modification is unavoidable. This paper describes the application of an automated workflow creating local modifications on well-by-well regions. This application is done on top of the global multipliers and all the other uncertainty parameters. It's done to generate an ensemble of simulation runs leveraging machine learning proxy models to history match a large onshore carbonate reservoir.

We applied a machine learning (ML)-based optimization process to perform history matching with a large number of uncertainties. A ML proxy-model was built from an ensemble of reservoir simulation models capturing varying uncertainty parameters. The ML proxy-model captured field and well level simulation results assessing the quality of the history match. An optimization framework was defined in which to embed the ML model, minimizing the mismatch between simulation results and historical production/injection. The objective function as defined assesses history match quality by comparing the historical production/injection data with simulation results predicted by the ML model.

The workflow was successfully applied to a large onshore carbonate reservoir with approximately more than 1000 wells and 20 uncertainty parameters. An ensemble of simulations was generated to capture the variation of the uncertainty parameters. This ensemble was used to train a high-fidelity reservoir proxy-model that predicts field level and well level simulation results. The optimization results provided a set of parameters by minimizing the mismatch between the results predicted by the ML model and the historical production/injection data. As a result, a high-quality history match was obtained on the key parameters considered. The results were then validated by running the reservoir simulation model with the optimum uncertainty parameters, which delivered results very close to the ML model prediction.

The optimization process with a ML model is orders of magnitude faster than attempting to perform the history matching using the reservoir simulation model. The ML-based optimization for the history matching framework developed demonstrated good results on a complex reservoir model with a high number of uncertainties and high well count, paving the way toward significant field development planning acceleration.

You can access this article if you purchase or spend a download.