Evaluation of A Cyclic Water Injection Program Using A Combination of
Artificial Intelligence and Reservoir Physics

An innovative technology called Data Physics*, which combines the predictive capacity of traditional numerical simulators with the speed and flexibility of machine learning, was applied to optimize  a mature waterflood in the Neuquen basin in Argentina.

A reservoir model was built by combining machine learning techniques with the partial differential equations of fluid flow as applied in conventional reservoirs simulatiors.  The approach is probabilistic in nature, thereby supporting the quantification of uncertainty.  Given the complex conditions of operation and control in secondary recovery projects, the dearth of data, which was inherently noisy, it is not feasible to define a unique deterministic solution. The proposed approach provides physically possible and statistically viable distributions (ranges and frequencies) of the variables affecting reservoir behavior.

Due to its predictive capabilities, the reservoir model, created at a layer level using raw uninterpreted production and injection data, is capable of reproducing and predicting field production history.   This enables continuous closed loop optimization and provides a program of injection water redistribution.   The result is optimized production, and reserves development with reduced injection costs.

As part of the assessment of different prescriptions, the operator wanted to evaluate the optimum way to cycle two sets of injectors to reduce water handling and injection costs.  A robust connectivity model was built from the reservoir model which was then used to evaluate and compare the different alternatives to cycle injectors during two years versus an optimum injection redistribution.  The results include production forecasts for the scenarios evaluated which, when executed, would allow the customer to reduce water cost while maintaining production levels unaltered.