A New Approach to Creating a Digital Twin of Well for Production Monitoring in Western Siberia Fields

Abstract

The purpose of this work is to create a stable and scalable digital twin of oil well for virtual flow metering task capable of predicting the flow rate of wells equipped with ESPs for many fields in Western Siberia at various stages and conditions of development. The basis of the suggested approach is to use the best aspects of classical petroleum engineering methods and machine learning algorithms. There is a digital twin of the well, which includes a hydraulic and electrical parts. The adaptation of the well model is carried out through the calculation of the degradation coefficients of the ESP and then a regression task is set to predict these correction coefficients using a gradient boosting model on decision trees. The final element is the prediction of the flow rate according to the well physical model based on correction coefficients predicted by machine learning methods. The described approach proved its effectiveness after testing at several fields in Western Siberia for various operating conditions. The algorithm was especially useful for estimating the flow rate of a well in severe cases: unstable well operation, joint measurements, a new well after drilling, a broken flow rate measuring system. A comparison was also made with the classical approach of forecasting the flow rate for one well - the area of applicability. An assessment of the sufficiency of data for the construction of the model, the degree of degradation of the approach in the absence of data, the ability to scale and increase computational costs has been tested. Metrics have been obtained and the method of assessing the quality of forecasting in relation to the problem of virtual flow metering has been improved. Practical recommendations on the implementation of this approach are given. The novelty of this work lies in the method of combining physical and statistical calculations - the sequence of calculations according to the scheme: white-box - black-box - white-box. As well as working with telemetry time series along with machine learning algorithms.