Physics-Informed Reinforcement Learning for Motor Frequency Optimization in Electrical Submersible Pumps: Enhancing AI-Led Decision-Making in Production Optimization

Abstract

Abstract This paper presents a novel approach, Physics-Informed Reinforcement Learning (PIRL), designed to tackle optimization challenges within complex systems. Focusing on oil wells equipped with Electrical Submersible Pumps (ESP), the study aims to enhance decision- making processes by integrating physical principles with reinforcement learning techniques. The primary goal is to optimize motor frequencies for ESPs leading to maximizing production rates, reducing brake horsepower consumption, and ensuring operational efficiency within the Best Efficiency Performance (BEP) range. By leveraging PIRL, this research offers a promising avenue for enhancing productivity and efficiency in oil extraction processes, addressing the limitations of conventional practices, thereby enhancing decision making capabilities in production optimization. Traditionally, a constant frequency is applied over extended periods of time, disregarding variations in reservoir and fluid properties as well as operational field conditions. This approach often leads to suboptimal performance and unnecessary electricity consumption. In contrast, the proposed methodology employs reinforcement learning (RL) to develop a machine learning (ML) model, acting as an advisory system for dynamically adjusting operational frequencies. By continuously evaluating feedback from the system and adhering to operational, field, and reservoir constraints, the proposed framework offers an automated response to changes during the production phase of ESP wells, ultimately enhancing production efficiency and reducing energy consumption.