No abstract

Traditional

The maintenance of electrical submersible pumps (ESPs) is a highly capital-, resource-, and manpower-intensive exercise that is traditionally performed by reactive process monitoring of multivariate sensor data. In the reactive paradigm, it is difficult to proactively distinguish between sensor fluctuations, trip events, and failures in real time. This paper presents a real-time alarming system for predictive ESP failure identification by constructing dynamic operating envelopes on real-time sensor indicators using machine learning (ML). This ML model identifies the complex relationships between pressure head, pump losses, and supplied electrical energy. Operating envelopes are dynamically updated and validated by continuously incoming sensor data to provide indicators of ESP trip events. Recommendations for the amount of data required to provide reliable predictions and an alarming system to classify an ESP trip as a failure are also incorporated in the model. The algorithm can identify longer-term trends and deeper functional relationships from historical data, as compared to the traditional engineering approaches used for ESP diagnoses. The new workflow with the predictive model can provide signals two weeks before the actual failure event, as indicated by the traditional workflow. Issues related to sensor data quality, including missing, misaligned, and erroneous data that may result in false positive notifications can be easily improved by incorporating more domain knowledge by subject matter experts and field engineers. The proactive identification of failure events using this real-time alarming system can improve production efficiencies by avoiding deferment losses. It also contributes indirectly by improving, reducing, and automating the time spent in analyzing failure events, such as dismantle, inspection, and failure analysis (DIFA). The alarming algorithm can be gradually incorporated into traditional systems to provide continuous improvements and to add value without incurring the high costs of initial deployment and change management associated with it. This paper presents predictive ML models deployed to analyze real-time sensor data to proactively predict failures in ESPs results. It also provides recommendations related to the use of these workflows to improve operating practices and production efficiencies and to reduce deferment. Similar approaches can be extended to the monitoring of other equipment in real time.

The current cycle for reservoir management requires several months to years to update static and dynamic models as additional data from the field [logs, production, pressures, core, four-dimensional (4D), etc.] are obtained. These delays in updating the models result in increased risk and contribute to a significant loss of economic value. The ultimate goal for next-generation reservoir management is to reduce the cycle from several months to a few days. The current challenges for developing a proactive/real-time reservoir management solution include but are not limited to the time and manual intervention involved in conditioning and interpreting the logging-while-drilling (LWD) and well log data acquired during and after drilling a well; updating three-dimensional (3D) petrophysical/static models; and the computational cost and time involved in generating reservoir models from static and production data (history matching). However, the current widespread use of machine-learning and cloud-computing capabilities leads to faster and more accurate models, enabling real-time or near-real-time decision making. Using machine learning, one of these challenges—updating the 3D static models—was successfully addressed, namely, updating porosity prediction in a 3D model after new information comes to light, such as logging from a newly drilled well. The conventional geostatistical approach does not always honor geological variations in the subsurface formations, because only one or two seismic attributes can be used for co-simulation, and only with first-order interactions. Additionally, and most important, generating hundreds of realizations on a 3D grid is computationally intense and time consuming. Typically, several weeks are necessary to generate these static models before feeding them into the reservoir model. The proposed solution is a machine-learning-based approach that integrates 3D spatial availability of seismic data with petrophysical properties. One important goal of reservoir management is to understand reservoir uncertainties before they adversely affect field development. This machine learning solution proved to be computationally less costly, more accurate, and much faster than the conventional geostatistical approach.