Deep Learning Techniques for Gas Well Production Optimization

Javier Fatou Gomez – TNO
Pejman Shoeibi Omrani – TNO
Stefan Belfroid – TNO

2020 ALRDC/SWPSC Artificial Lift Workshop
Oklahoma City, OK, USA
Feb 17-20

This presentation describes a project to upgrade knowledge/practicability of start-up/shut-in of wells. Numerical and multi-tank models were built to model a dynamic reservoir with intermittent production. The models were then validated experimentally in 2019-2020 and melded with field data to develop fully data-driven optimization.

Two main activities presented in this presentation: Virtual metering/back-allocation and Intermittent production: data-driven production optimization. Why Machine/Deep Learning? Physical models or experimental data may not be feasible/available (complex non-linear dynamics, absence of sensors, costs…). Significant amount of field data available for many processes = opportunity. TNO is working/has worked extensively with Machine Learning for Oil and Gas.

With virtual flow metering it was found that Recurrent Neural Networks could predict liquid flowrates with less than 1% of relative error, although time-dependencies are
important. For back-allocation, Artificial Neural Networks could predict single well flowrates with 94% accuracy. Domain knowledge is key to intermittent production optimization and current time step monitoring – a naive approach might result in significant overfit/non-physical results. A 1.3% cumulative gas production error was achieved for a liquid loading/meta-stable dataset in future time steps forecasting. Work was continuing to develop fully data-driven intermittent production optimization.