Machine Learning Assisted Reservoir Management for Injection Projects

Reservoir management is critical for optimal hydrocarbon reservoir performance. A key component of reservoir management is decision making utilizing numerical reservoir models. One downside of these models is the large computational footprint for development and deployment. Recent developments in machine learning provide technologies that can augment reservoir management workflows. In this work, machine learning algorithms are used to optimize selection of injector well location, perform fluid production prediction and optimize well controls. Selection of optimized water and gas injection well locations is a key challenge in secondary and tertiary recovery projects. Traditional approaches using numerical simulation require a high number of simulation runs with optimization algorithms. This is further complicated if there is geological uncertainty (e.g., multiple geological models are required to capture the uncertainty range). An alternative approach is proposed, using machine learning algorithms trained on a geological ensemble. Well level aggregations are proposed to efficiently reduce the number of injector well location evaluations. Proxies were trained with several target variables including oil recovery and CO2 storage. These proxies provide rapid evaluations of injector locations within seconds at accuracies between 0.94 - 0.98 (R2 score comparing machine learning proxy predictions and numerical simulation results). Blind tests with different geological realizations yielded results comparable to numerical simulation. Posterior sampling was used to determine optimal injector locations across a large geomodel ensemble. To predict time varying production without numerical simulation, deep learning models including a Convolutional Neural Network (CNN) were trained with well level data. These algorithms successfully predicted production from water and miscible flooded patterns. Models trained on several wells could predict production from a neighbor well within the same pattern area. Finally, a Graph Neural Network (GNN) proxy was used within a Closed Loop Reservoir Management (CLRM) workflow to optimize oil production of a waterflood. Kriging was used to initialize the adjacency matrix, resulting in improved network performance. The GNN proxy was able to perform monthly optimization over a period of one year within eight minutes, compared to 6 hours using traditional simulation.