ESP Frequency Optimization - A Machine Learning Approach as Field Study

Electrical submersible pumps (ESPs) are a widely used artificial lift method that supports the performance of the well and is used to ensure efficient, stable oil production or to support water production in geothermal wells. The efficiency of an ESP system strongly depends on its operating conditions, particularly the pump frequency has a high impact on the operation. While the pump frequency can be easily adjusted, the more difficult part is to find the optimum frequency. The possibilities to measure and process data from oil fields have steadily increased over the last years. Therefore, the use of data-based approaches to optimize the oil production is becoming increasingly important. This thesis presents the proof of concept for a physics informed reinforcement learning (RL) approach which is used to find the best frequency to optimize the production conditions of the ESP system with the goal to maximize the flow rate while considering system constraints. For this approach, a mathematical proxy model referred to as the environment, is calibrated to mirror the real system behavior and is used as a basis for the reinforcement learning algorithm. The agent of the model learns through a feedback loop, selects frequency actions, receives reward-based feedback and continuously improves decisions to identify the optimal frequency settings. In addition to proving the theoretical concept on real field data, the significance of the data quality and availability was analyzed in this thesis. A very important point is the selection of the input data time resolution, as monthly; daily and hourly data were analyzed. A Python based framework was implemented to handle the environment calibration, RL training and frequency prediction. Results demonstrate that the environment can be successfully calibrated by the model and that it sufficiently imitates the real well conditions. The result of the frequency prediction strongly indicates that the RL model is successful and consistent in finding the best frequency for the operation conditions to improve the overall production efficiency.