Management and Optimization of a Large Waterflood Operation Applying a Physics Embedded Machine Learning Workflow

The paper describes the process to model a large waterflood field located in Montana using a workflow that creates a predictive model from production and injection historical data to infer petrophysical parameters in a reduced period of time while honoring the reservoir physics at all times; data from 2010 was used to create a layered model which was later used for proactively managing the field as well as for optimization purposes. Different workflows have been carried out using the model for surveillance and forecasting, as well as generating the pareto front and selecting a target scenario to increase production by redistribution and implementing its prescriptions in the field.

The workflow aims at significantly reducing the time spent in creating a model by blending traditional reservoir physics with modern data science techniques (AI and Machine Learning) to create a continuous model able to replicate the field behavior. Because the model runs very fast (minutes) it can be used for day to day management and optimization of the field. The former is achieved by proactively using the base case model to create scenarios of various operational decisions; the latter leverages evolutionary algorithms to run thousands of variations of the base case with the objective of finding the optimal operational injection strategy (pareto front) which was later implemented in the field.

These functionalities were put to work in the implementation of a closed-loop optimization workflow in which the input data and the model are updated frequently, maintaining an evergreen model of the field which can be used to run “what-if” scenarios in minutes and generate the pareto as mentioned above. Some of the what-if scenarios evaluated the impact of operational decisions like injection changes, wells shut-in, inactive well reactivation and a thorough injection sensitivity analysis to quantify the impact of each injector and rank them by importance. The pareto was generated running the model more than 15000 times and three optimum scenarios were assessed for implementation. The final implementation is in place and the initial results encouraging, validating the predictive ability of the model.