Smart Well Data Generation via Boundary-Seeking Deep Convolutional Generative Adversarial Networks

The current trend in the Oil & Gas industry is the use of more complex and detailed reservoir models, seeking better refinement and uncertainty reduction. Alas, this comes with a great increase in computational time, encumbering the optimization process. With the growing adoption rate for smart wells in oil field development projects, these optimizations are indispensable as to justify the investment on the technology and maximize financial return, by finding the optimal valve control schedule. The present paper seeks to establish a new methodology for creation of smart well data by means of a deep generative model, capable of modeling complex data structures. This generation of data is advantageous to the industry as it can then be used for various other applications. Other benefits besides the reduction of optimization time include the use in data augmentation, where the network is used to diversify existing data as to improve lacking datasets, and data privacy, as the generated data, while next to real, can be shared without the original, protected model. A case study was done in an industry-recognized bench-mark model, and the results completely support the use of the proposed methodology, as it was able to achieve all expected objectives.