Creating a Multiphase Production Model Tailored to Deviated Oil-Producing Wells for Integration as Input into a Machine Learning Model for ESP Survival Analysis

Abstract

Abstract A comprehensive model estimates wellbore production variables, including flow velocities, phase fractions/holdups, fluid properties, pressure/temperature profiles, and Electrical Submersible Pump (ESP) performance metrics. Coupled with instrument data, these inputs support productivity surveillance and ESP status prediction driven by machine learning (ML). The wellbore production model's effectiveness is evaluated by comparing it with industry-sourced field data encompassing muti-probe production logging tool (PLT) measurements from multiphase producing wells. This work is based on Nodal Analysis, which involves dividing the wellbore into numerous sections and applying the conservation of mass, momentum, and energy principles to model pressure, temperature, and other essential profiles. Three multiphase flow methods are employed: homogeneous flow, separated flow with slip (Hagedorn-Brown method), and separated flow with slip and flow pattern (Mukherjee-Brill method). Affinity Laws are used to describe the ESP performance. A machine learning model is trained using manually labelled historical data subsets comprising model results and actual field measurements. Its purpose is to recognize ESP operational statuses such as pump off, normal operation, electrical wear, and mechanical wear. Supervised feature selection methods are utilized to identify the most relevant parameters. The along the wellbore measurements from PLT e.g. the phase holdup (also referred to as in-situ volume fraction) is in agreement with the modelling results. For a wide range of liquid rates and gas-liquid ratios, flow rates can be determined with an average deviation of less than 10%. Machine learning feature selection methods, such as sequential backward elimination, reveal that production modelling results are crucial for identifying ESP statuses, including mass rate, viscosity, and pump parameters like efficiency. This study demonstrates that hydrodynamic modelling results provide additional information for ML training that electro-mechanical raw data may lack. Thanks to the integration of hydrodynamic modelling and raw data supplied to the ML algorithm, it can classify operational statuses with 99% accuracy and predict ESP failure months in advance. When the model is connected to standard wellbore instrumentation, it enables near real-time production monitoring and provides essential hydrodynamic input to ML-based algorithms for continuous monitoring ESP equipment. It can be used as a virtual flowmeter (VFM) or a validation tool for multiphase flowmeters (MPFM), enhancing allocation split accuracy and enabling operators to concentrate on true contributors. The methodology can be integrated into a digital oilfield (DOF) system, employed as a digital twin, or, as demonstrated in this study, integrated into asset modelling with ESP survival analysis and failure prediction using ML.