_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 219528, “Autonomous ESP Optimization Using Machine Learning Demonstrated in the Permian Basin, ” by Ryan D. Erickson, SPE, Reynaldo Ramos, SPE, and James B. Meek, SPE, Vital Energy, et al. The paper has not been peer reviewed. _ This paper highlights the demonstration, refinement, and implementation of a machine-learning (ML) algorithm to optimize multiple electrical-submersible-pump (ESP) wells in the Permian Basin. The complete paper presents two case studies for autonomous ESP optimization driven by this ML model. The paper discusses key learnings from each study to assist operators in their digital journey with considerations for effective field implementation. Digital Maturity Framework The framework is a two-variable matrix clarifying important milestones for both maturity of capability and solution implementation (Fig. 1). For the scope of implementing an ML inference model (MLIM) to source set-point recommendations to optimize a well producing with an ESP, the capability and implementation maturity paths are defined next. Maturity of Solution Capability - Measure—Predefined sensor and telemetry data are collected in a timely, consistent, and holistic manner. - Optimize—Key artificial lift tuning parameters, including ESP frequency and well flowing tubing pressure (FTP) are sourced from the MLIM. - Automate—MLIM set-point recommendations autonomously write to field control devices. Maturity of Solution Implementation - Develop—Configure field-data capture, create MLIM that autonomously generates set-point recommendations of ESP frequency and well FTP with a user interface (UI). - Demonstrate—Use ML capabilities on a subset of wells producing in the Midland Basin. - Refine—Improve the MLIM with reinforcement and feedback from domain experts and enhance UI. - Deploy—Distribute the technology at scale across the field and transition into core business. MLIM Overview. After building and testing a mature data-acquisition environment, the next step in autonomous ESP optimization was to develop an MLIM to deliver high-quality ESP set-point recommendations. A previous publication focused on developing an ML model for ESP optimization using an extensive data set from 193 ESP-operated wells. The data-preparation process involved cleaning, handling missing values, and normalizing features to ensure high-quality inputs for the model. Ultimately, an artificial neural network (ANN) was selected as an algorithm because of its flexibility, control over the training procedure, and superior accuracy. The ANN model was designed to forecast daily production rates based on current telemetry, ESP operating mode, flow rates, and other well-specific values. It consisted of two submodels—the telemetry model, which predicted daily telemetry values, and the production model, which forecast daily production rates. The training process involved splitting the data set into training, validation, and test parts, and conducting training in two stages to achieve optimal results. The model demonstrated exceptional performance, with an average R2 score of 0. 98. Furthermore, the observations from the 193 wells were reinforced by expert petrotechnical input. This high accuracy underscores the model’s potential for real-time ESP optimization, enhancing production outcomes and overall well profitability. A UI was developed to house the MLIM configuration screens, historical telemetry data for ad-hoc analytics, and visualizations to make the ANN-sourced opportunity clear and easy to understand.
Chris Carpenter (Wed,) studied this question.