Bayesian neural networks for virtual flow metering: An empirical study

Recent works have presented promising results from the application of machine learning (ML) to modeling flow rates in oil and gas wells. Encouraging advantageous properties ML models, such as computationally cheap evaluation ease calibration new data, sparked optimism for development data-driven virtual meters (VFMs). Data-driven VFMs are developed small data regime, where it is important question uncertainty robustness models. The may help build trust which a prerequisite industrial applications. contribution this paper introduction probabilistic VFM based on Bayesian neural networks. Uncertainty model measurements described, shows how perform approximate inference using variational inference. method studied by large heterogeneous dataset, consisting 60 wells across five different assets. predictive performance analyzed historical future test an average error 4%–6% 8%–13% achieved 50% best performing respectively. Variational appears provide more robust predictions than reference approach data. Prediction explored detail discussed light four challenges. findings motivate alternative strategies improve VFMs.