A Physics-Informed Neural Networks (PINN) oriented approach to flow metering in oil wells: an ESP lifted oil well system as a case study

Abstract

A virtual flow meter (VFM) is an interesting and growing method to replace or support expensive and high-maintenance dependent physical flow rate sensors in oil and gas production systems. This paper presents a new use of Physics-Informed Neural Networks (PINN) as a hybrid virtual sensor to be used in an oil well system. This strategy combines prior knowledge about the system dynamics in the form of a phenomenological model in order to drive a long short-term memory (LSTM)-type recurrent neural network (RNN) training. The resulting model is able to predict the average flow rate in the production column some time steps ahead using measurements of states and exogenous inputs available in an oil well information system. The proposed scheme is also flexible as it can estimate fluid properties via their model parameters as an additional feature. The results obtained in this study through simulation used a benchmark electrical submersible pump (ESP) model to demonstrate the proposed PINN-based VFM potential as an accurate alternative to infer flow rate in oil wells.

Keywords

• soft sensor
• Physics-Informed Neural Networks
• electrical submersible pump
• virtual flow meter
• recurrent neural network