Improving ESP Production for Unconventional Wells Through Real-Time Machine Learning Based Changepoint Detection

Abstract The operation of unconventional wells utilizing electrical submersible pumps (ESPs) is filled with challenges such as broken shaft, intake blockage, blockage of pump stages, increase of free gas at intake, etc. Additionally, any deviation from optimal operation conditions means loss of production for operators. Detecting these events and conditions early is imperative for minimizing downtime and optimizing production. Most operators rely on manual and/or rule-based processes for the detection of these events, leading to delayed event detection and associated productivity losses. In this work, a novel approach is proposed for earlier detection of these events, which are referred to as changepoints. Most of these changepoints can be identified as repeated patterns on the time series data collected during the well operation. Hence, a machine learning (ML) model trained on well historical data should be able to detect these changepoints. In this work, historical data from over 200 Permian unconventional ESP-operated wells are used for training and evaluating the ML models. This data includes high- frequency pump measurements, intake pressure, electrical current, and temperature as well as production rates. Most of the changepoints show very similar changes in well data when comparing one well to others in the asset. Therefore, the objective is to develop a global model applicable to all the wells. After labeling the known changepoints, different machine learning models are developed and evaluated. The sliding-window technique was leveraged for creating the training data. Various window sizes were tested. Statistical attributes of the data within each window were used as features for ML model training. These features include mean, standard deviation, and time series similarity metrics such as dynamic time warping, cross-correlation in the time/frequency domain, and power similarity. Out of the different tested algorithms, the Extreme Gradient Boosting algorithm proved to be the most successful in detecting the changepoints. The trained model proved to be able to successfully identify close to 90% of the changepoint events in less than a few hours (compared to a few days or weeks of detection time using the manual monitoring processes) from the time of occurrence. For some events, the models were able to identify earlier data patterns that were leading to the main problem events hours before they happened. Although a similar method is applied in other industries, this work is the first application of ML for ESP wells monitoring and changepoint detection. This work proves that ML models can lead to early detection of well problems, which helps operators take corrective actions before they cause disruptions and downtime, reducing repair costs and improving efficiency.