Production Optimization Using Integrated Modelling and ESP Survival Analysis Based on Historical Data and Machine Learning

Abstract

Abstract Most oil wells are converted to artificial lift, such as an electrical submersible pump (ESP), at some point in their lifecycle due to reservoir pressure depletion. This paper presents an integrated approach to model the entire wellbore-ESP system, allowing the computation of production rates and the assessment of equipment health for failure prediction. Historical data logs from two Middle East fields, operating about 100 instrumented wells, are used to validate the work and develop a failure detection process using machine learning (ML). The production model utilizes Nodal Analysis, encompassing multiphase flow and accounting for slip and flow pattern, to determine the in-situ density used for computing the pressure traverse. The temperature traverse is computed from transient heat transfer between the wellbore and formation. Affinity laws are used to describe the ESP performance. Feature selection methods, including the Pearson correlation, Sequential Forward Selection, and Backward Elimination, are employed to determine the most important features. Feedforward neural networks with fully connected (dense) layers are trained on manually labeled subsets of measured and calculated parameters to detect operational statuses, including anomalies. The identified statuses entail pump off, normal operation, electrical wear (including harmonics stemming from poor power supply), and mechanical wear. An analysis is performed for about 20 failure cases from the historical data by reviewing Pull Out Of Hole (POOH) reports, Dismantle, Inspection, and Failure Analysis (DIFA) reports, Teardown Analysis reports, as well as historical high-density field measurements. The observed failure modes include shaft breakage and electrical failure at the ESP cable, the cable penetrator, the Motor Lead Extension (MLE), and any splices located in the system. The root causes of the failures are motor overheating, hydraulically induced mechanical load peaks, degradation of insulation of electrical conductors, and voltage harmonic distortions stemming from poor power supply quality. Using the presented methodology, it is possible to detect wear at the onset, allowing for the prediction of failures like shaft breakage months in advance. The integrated production model is validated by comparison with field data. Abnormal ESP behaviour is predicted correctly in the early stage of development with more than 99% accuracy. Data science tools enable the detection of equipment degradation months in advance, facilitating well-workover preparation and minimizing unnecessary downtime and production loss. This capability leads to significant cost savings. Besides failure prediction, the analytical production model calculates parameters along the wellbore that can greatly assist in identifying production problems, such as flow pattern and in-situ along the wellbore parameters, including the black oil model properties. Integrated modeling fills information gaps between rate measurement and serves as a verification tool of physical metering. Combining continuous production surveillance with predictive maintenance leads to reduced production deferment, improved allocation, and better well and reservoir management (WRM). The pump failure prediction methodology can easily be extended to other operational conditions, such as motor or wellbore inflow related issues. Furthermore, this approach can seamlessly integrate with digital oilfield (DOF) or digital twin processes.