On Machine Learning Approaches for Wet Gas Flow Metering with Actual Data: From Handling Available Data to Developing Models

In the context of wet gas production, wet gas flow meters (WGFM) are solutions that can be placed in-line and ally measurements from different sensors with complex data processing to assess individual flowrates in a multiphase flow composed mainly of gas, oil, and water. Machine Learning strategies emerge as powerful tools for modeling the relationships among variables, but previous studies have focused on data obtained from controlled lab-scale measurements. In this context, the present paper presented for the first time a detailed methodology and evaluation of the development of data-driven models for the prediction of individual flowrates of gas, oil and water in a wet gas stream considering data from the operation of a WGFM device in an actual production site operated by an energy company in Brazil. Particularly, the themes discussed cover data pre-treatment, instance selection and model training, with PCA and Neural Networks being explored. The main strategy in which developments relied was considering on/off production information to split data into quasi-steady state and transient production regimes; then, only transient data was considered for model training, which offered a rich and concise source of process information that allowed developed models to learn to represent not only transient, but also the quasisteady state data. Particularly, the developed models presented exceptional performance on the quasi-steady state predictions both for the temporal window for which the transient data used for training was obtained and subsequently, indicating that the developed models could be applied to in-line applications. Typical prediction errors were generally below 6% of actual measured values. It is believed that the presented developments contribute to improving WGFM technologies, as virtual sensors can address hard-sensor failures, replace gross-error measurements and also be used to filter random errors. Such developments may enhance the technical and economic feasibility of WGFM at plant sites