Implementation of Data Driven Machine Learning Methods and Physics Driven Concepts for Real-Time Well Performance Estimation in Kashagan Field

Abstract

Abstract This paper presents the implementation and value of Data Driven Machine Learning methods and Physics driven Concepts for real-time well performance estimation in Kashagan Field. The models are expected to be used to detect fluid properties changes, restrictions in the well tubing and instrumentation malfunctioning. The paper also describes how the data driven and physics driven models are built, tuned and maintained. The comparison of the methods will also be attempted in terms of redundancy. The data driven machine learning method (DDML) used in the paper is artificial neural networks model. The model consists of several inputs, layers of artificial neurons and output parameter(s). The built DDML model is trained by feeding labeled input test data and trained. Then the trained DDML model is validated and tested against separate labeled data which the model did not "see" before. Physics driven model (PDM) is built on Vertical Lift Performance (VLP) equation. The PDM is also "trained" with input data by tuning the model in order to fit to test data. The tuning parameter is friction coefficient. The trained DDML and PDM models are then tested on same set of data in order to compare the performance. Both methods showed similar degree of accuracy. The reference data comes from MPFM, Test separator and downhole gauges. Downhole pressure gauge signal is used by both of the methods. However, if the downhole gauge is broken, the PDM will not function for estimation of the rate. On the other hand, DDML model can be re-trained with available signals (tubing head pressure, temperature, choke size, flowline pressure and temperature, etc.) by excluding the downhole gauge signal from the training data. It is clear that the DDML model trained without downhole pressure signal is not as accurate fully trained DDML model. However, it is still possible to have reasonable estimate of the parameter with even partially trained model. Both models can be used to estimate downhole pressure in cases where the gauges are broken. The proposed methods can be used by practicing engineer to estimate well performance parameters. The methods are mostly suitable for so called smart fields where the data is readily available and makes it possible to extract useful insights from the archived data. With hands on experience with such simple implementation of data driven machine learning techniques an engineer can try to implement ML methods for other applications, e.g. well interference, history match, anomaly detection, etc.