Forecasting Smart Well Production via Deep Learning and Data Driven Optimization

As smart well technology is increasingly being adopted in oilfield development projects, the need to optimize controls emerged in order to justify its higher initial investment by considerably increasing net present value. While there are numerous methodologies to achieve this goal, a common fact in all is the need for a great number of computationally expensive reservoir simulations, hindering extensive optimizations. This paper proposes the use of deep learning algorithms in proxy models, in order to accurately replicate the behavior of the simulator by forecasting production based on previous data. Thus, a smaller number of simulations are required for a training dataset, and the proxy can then be used in lieu of the simulator for optimization purposes. Other benefits in the use of the proposed methodology include the gathering of insights on production, as problems might be occurring if measured production noticeably deviates from the forecast. Two case studies were done, and the results indicate that a Long Short-Term Memory Network-based proxy is able to forecast production with a remarkably low error, validating the methodology and supporting its use.