This paper introduces a new sensor analytics application for flow profiling that is based on physics-informed machine learning (ML) techniques. The application was utilized in the interpretation of point and distributed sensor data acquired from a carbonate reservoir in the Middle East and the results enhanced the confidence in the utilization of fiber-based flow monitoring solutions. The application was tested on two data sets: (1) a distributed fiber optic sensing (DFOS) data set acquired during acid stimulation and water injection periods from an injector and (2) a point acoustic sensor data set acquired from an oil producer. In both cases, the output of the application was qualified using independent measurements.

A combination of ML and first-principles models has been used to develop the real-time sensor analytics application. Over 1500 data points were acquired in laboratory conditions where flow conditions were simulated. Various ML models were trained on the labelled data from these experiments to provide flow diagnostics for the subject well completion type and those with highest accuracy were selected for development. Independently interpreted production and petrophysical logs were used in the qualification process to validate the interpretation results obtained with the developed sensor analytics technology.

The results from both cases agreed with the results from the qualifying measurements. In the first case, the production log and the DFOS measurements were taken under two different injection conditions. Hence, some discrepancies were observed between the two which were explained by the heterogeneity within the reservoir section. In the second case, data from the point acoustic sensor and the production logging (PL) tool were acquired under the same conditions. The flow profiles predicted by the ML model applied on the point acoustic sensor data agreed with the production log.

You can access this article if you purchase or spend a download.