Hybrid Physics/ML Framework for Virtual Flowmetering Optimizes Production

_

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 233457, “A Hybrid Physics/ML Framework for Virtual Flowmetering in ESP-Lifted Oil Fields: From Troubleshooting to Robust Deployment,” by Khaled A. Raslan, SPE, Badr Petroleum; Hossameldeen Elnaggar, University of Wyoming; and Mostafa A. Sobhey, Khalda Petroleum, et al. The paper has not been peer-reviewed.

_

This study presents a hybrid physics/machine-learning (ML) framework for virtual flowmetering in electrical-submersible-pump (ESP)-lifted oil fields, with a focus on practical deployment and real-world applicability. Unlike conventional approaches that evaluate performance on randomly split data, the proposed workflow employs a well-based data-partitioning strategy to ensure that model performance reflects true generalization to unseen wells. In addition, the learning task is reformulated in a normalized space relative to pump best efficiency point (BEP), enabling consistent scaling across different pump sizes and operating conditions.

Methodology

Data Set and Study Scope.

The data set employed for the purpose of the current study contains approximately 35,000 samples acquired from approximately 50 ESP-lifted wells. In each sample, a combination of surface measurements, pump conditions, and completion attributes is accompanied by an observed gross fluid-production rate.

Data Cleaning and Standardization.

To achieve uniformity across all wells, categorical features such as formation type, pump type, and operating mode were made uniform by trimming excess white space and making text uniform. Also, the date column was converted to a consistent date/time format.

A normalization procedure was conducted to convert perforation-interval values into a consistent format by replacing incorrect dashes, adding spaces to separate intervals, and eliminating characters other than numerals. Intervals that were physically impossible were discarded during this procedure.

Completion-Feature Extraction.

Following normalization, perforation intervals were converted into physically meaningful completion features.

Physics-Informed Feature Engineering.

To achieve better robustness and interpretability in the model, some additional features were engineered according to ESP operation principles. The pressure difference across the pump was computed by subtracting the intake pressure from the discharge pressure. Frequency was expressed as the speed-equivalent version of the variable to accommodate the difference between the two- and four-pole motor types.

Head coefficient was represented using the ratio of the pressure difference to the number of stages and the squared equivalent frequency value because a direct dependency exists between the head generation and rotation speed. Other engineered variables include torque coefficient, power normalized to the number of stages, and the pressure drop induced by the choke.

Well-Based Data-Splitting Strategy.

A data-partitioning method was used to make sure that the assessment was performed with zero leakage and was reflective of the real situation. In contrast to the random allocation of data points, the wells were assigned to training, validation, or test data sets. Approximately 15% of the wells were selected for testing, without any contact during training or hyperparameter optimization. The other 85% was divided among training and validation subsets, resulting in roughly 70% for training and 15% each for validation and testing.