Neeraj Zambare and Kapil Mukati, Kongsberg Digital, Norway, explain how a combination of multiphase flow simulation and machine learning is enabling enhanced virtual flow metering and anomaly detection.
Well flow metering is one of the most important measurements for offshore oil and gas facilities. It not only provides a means of flow allocation, but also helps in predicting the overall reservoir depletion and reservoir life. The most dependable way of metering a well under any circumstance is still by routing a single well to a test separator and measuring the phase flow rates topsides. However, this process is costly as it involves deferring production from other wells. For older fields, it is sometimes not possible to flow a single well to test separator because of flow assurance issues – for example, too low flow rate, too high water cut, slugging etc.
Multiphase flow meters (MPFM), typically installed subsea on well heads, have been around for more than 25 years. These meters can measure multiphase flow rates on a continuous basis, thus reducing the need for costly test separator measurement. Though the MPFM technology has advanced over the years, it has its own problems. Firstly, it is not cheap. Secondly, it requires calibration at a certain frequency based on well test data; thirdly, its useful life is limited; and lastly, subsea replacement installation costs are so high, most operators choose not to.
Virtual flow metering
Physics-model based transient multiphase flow simulators have been around for more than 25 years as well. Some of these models boast high accuracy in estimating phase flow rates in a pipeline (e.g. LedaFlow from KONGSBERG). This gave rise to Virtual Flow Metering (VFM), which is a software-based solution deployed on a Microsoft Windows based computer. Like MPFMs, VFM needs to be calibrated from time to time, however it has clear advantages over MPFM by being less costly and easier to maintain. Model based VFM depends on good quality measurements from the field such as bottom hole and well head pressures and temperatures for accuracy. Incorrect measurements or failure of key sensors can adversely impact accuracy. This can be improved to some extent by having fall back strategies for instrumentation failures.
Figure 1. Hybrid approach using physics-based model and ML.
For any process, it is very important to detect anomalies as soon as possible, since in most cases, these earlier observed anomalies lead to either undesirable shutdowns or worse – disasters. Again, there are physics-model based anomaly detection solutions available today. These solutions are based on the common principle of comparing measured data against data predicted by models built using available vendor-provided design data. A statistical deviation in any of the measurements (prioritised based on which measurement is affected the most by the monitored anomaly) from the reference model data implies that there is an anomaly – meaning the equipment or system is not performing according to specifications.
Machine Learning (ML) is a type of Artificial Intelligence (AI) where computers learn from data continuously and get progressively better at performing some task in an autonomous fashion. There has been a surge recently in pursuing AI solutions in the oil and gas industry, with applications ranging from hydrocarbon exploration to reducing operating costs, and to improving data management. ML-based VFM has the potential to reduce asset development and operating costs by replacing expensive MPFMs.
The principle behind ML-based VFM is to use historical well sensor data (pressures, temperatures, flows and choke positions) to learn how a well has been producing in the past and estimate phase flow rates as the current information is fed in real time to the ‘trained’ ML model. However, just like any other ML technique, accuracy depends on both quantity and quality of available data. Sometimes quantity of data is available, especially if the well has been producing for several years, but quality is lacking i.e. information over the full range of choke opening or wider range of water cut. More frequently, however, it is difficult to get a sufficiently large data set covering the whole operational range.
Machine learning techniques can be effective in detecting anomalies, at least when compared to the data it was trained with. Even though the black-box based approach of ML models is not very well accepted in the oil and gas industry, it has significant potential to be more industrially acceptable for real time applications compared to physics-based models since they are computationally much faster and more stable. The main reason for computational simplicity is that more complex activity of learning is achieved up front and ‘trained’ ML models are much easier to compute in an online environment.
Like physics-model based VFM, instrument failure can be detrimental to ML solutions since they depend on real time sensor data to determine process state. Measured data also can rarely be used ‘raw’ because ML requires information relevant to the task it is being trained to perform. It is not unusual to spend more than 50% of time in an ML project on data massaging i.e. cleaning, preparing and organising data that can be used by ML algorithms.
Figure 2. Example of water breakthrough detection and water rate estimation using Hybrid ML VFM.
Hybrid solutions – where physics-based models are used in tandem with ML techniques – provide clear benefits over pure model-based or pure ML-based solutions. In a hybrid solution, machine learning is done using synthetic data generated from physics-based models ‘tuned’ to the actual real-world conditions. Any gap in quantity or quality of training data can be easily filled with physics-based models since they can simulate a wide variety of conditions and scenarios, even if they have not happened in the real-world yet, e.g., water break through, GOR change or choke blockage. ML VFM trained in this manner can provide very accurate flow estimates not just for existing process conditions, but for any expected future changes as well. Using physics-based models for training ML models makes it possible not just to detect anomalies, but to determine their nature as well.
In addition, synthetic datasets created using physics-based models require minimal data cleansing effort since they are generated with a specific ML task in mind. Consequently, Hybrid ML solutions can be quickly deployed and are cost-efficient.
Figure 1 demonstrates a state-of-the-art proposed Hybrid ML approach. The calibration and learning steps can be automated if a traditional physics-based model real time simulator is deployed together with a ML solution. Any detected anomaly or performance deterioration can then trigger calibration and ML re-learning steps, resulting in a truly adaptive and user intervention-free deployment, the added benefits being redundancy and confirmation of model outputs. In addition, the parallel physics-based real time simulators can provide engineering insight into the problem that the black-box approach of ML lacks.
Industrial example – water break-through detection
In general, the presence of water in oil and gas production is undesirable because it creates high risk flow assurance problems in addition to lowering overall production. Once a well starts producing water, usually there are a series of planned actions that must be undertaken to avoid shutdowns or potential catastrophes related to hydrate formation. However, it does not matter how well this has been planned for if no one knows when a well starts producing water. This phenomenon, called ‘water break-through’, is an even bigger problem in multi-well subsea developments as it impacts flow allocation as well. Even MPFMs are not designed to detect this early onset of water production because it is typically out of its measurable range. In such multi-well subsea developments, it is difficult to determine which well has started producing water even after it arrives topsides, since what we get topsides is commingled flow from all the wells.
A hybrid VFM can be used to detect water break-through because a training dataset can be generated by simulating the physics-based model over the full range of water cuts. ML computation then ‘knows’ what to expect when water break-through occurs in a well and can detect not only the event but provide water flow estimates as well, in addition to oil and gas rates.
This will be impossible with a sensor-data-only ML solution because there is no real-world data to learn from before an actual water break-through event.
Figure 3. Physics-based models used in tandem with machine learning techniques can deliver accurate flow estimates.
Read the article online at: https://www.oilfieldtechnology.com/digital-oilfield/30042019/a-hybrid-approach/