You have access to this full article to experience the outstanding content available to SPE members and JPT subscribers.
The complete paper discusses the successful application of a data-driven approach to analyze production data and identify root causes of slugging in a subsea production system on the Norwegian Continental Shelf. The approach used machine-learning techniques to model and analyze historical production data to identify the drivers behind slug flow. The results were used in combination with simulator studies and engineering experience to create a better understanding of the underlying root cause and to make it easier for field engineers to leverage all available information to reduce slugging and optimize production.
Slugging Challenges in Offshore Fields
Subsea production systems, characterized by deep wells and pipeline-riser setups, are especially prone to slug flow. Severe slugging, which can occur in the riser as a result of a pipeline topology-induced low-point angle at the base of the riser, produces only one slug in the riser at a time, and its length can reach that of the entire riser. Slugging may also arise in wells. In gas-lifted wells, where gas is injected through a one-way valve at the bottom of the well, casing-heading slugging may occur. Slug flow can also be present in wells without gas lift. This type of slugging, called well instability, generally features shorter slugs than does severe slugging, with several slugs simultaneously present in the riser. The mechanisms of this regime are not entirely understood. For example, commercial multiphase flow simulators are often not able to reproduce well instability. “Slugging” in this paper refers to either severe slugging, casing-heading, or well instability.
Slugging can lower average oil production. Production loss can also be caused indirectly because operators commonly combat slugging by choking the riser to increase system backpressure to stabilize flow. Production reductions of up to 50% have been reported as necessary to minimize slugging on offshore platforms.
The slug-flow pattern is associated with large and abrupt variations in pipe pressures. The pressure variations may cause wear and tear on the equipment, degrade the separation process, and, in the worst case, force a system shutdown. Reducing or eliminating slugging is thus critical for both economics and safety.
In mature offshore fields, slugging may follow an increase in water and gas production. Conditions for slug flow can also be created when a production system is extended with tie-ins and new wells. In such scenarios, the operator must analyze the causes of slugging and mitigate its effects. While slug flow can be controlled by choking the riser production, understanding slugging root causes may present opportunities for minimizing production losses, for example by altering the fluid composition of the commingled flow in the riser.
Operators of offshore production systems collect and store real-time production data from hundreds or thousands of sensors to monitor and optimize production. In cases where slugging arises, a data analysis to identify root causes can yield crucial insights for decisions on whether to take preventive measures or actively control slugging. Oil companies are increasingly reliant on the application of statistical and machine-learning methods to extract such valuable insights from production data.
The asset, which is being produced with a subsea production system, experienced severe slugging in a riser, which limited production throughput. The system comprises eight subsea wells, four of which are multibranched and gas-lift-capable. The wells produce a mix of oil, gas, and water. Well flows are commingled at the seabed and lifted to the topside processing facility by a 3-km-long riser pipe.
The wells are well-instrumented, with several pressure and temperature sensors installed at the wellhead and downhole. Each well also has a working multiphase flowmeter, providing gas, oil, and water rates. Gas/oil ratio, water cut, and gas/liquid ratio are derived for all well and riser flow rates. With regard to topside instrumentation, more than 100 sensors directly measure the state of flow in the production system.
After several years of steady-state production, the operator experienced an onset of slugging. While some operating regions showed a low degree of slugging, others appeared to create conditions for slugging. The overall trend was that the degree of slugging increased over time, likely the result of higher gas and water production. The gradual worsening made it difficult to pinpoint the causes, or drivers, of the slugging. To avoid heavy slugging, the operator eventually deemed it necessary to choke the riser and lower total production.
To understand slugging scenarios, a series of investigations was conducted. Different operating settings were explored, and a simulation study was conducted. The simulation study failed to reproduce the slugging. After 1 year of investigations, it was still unclear how to best mitigate slugging, and the operator continued production at reduced capacity by choking total production. Further investigation into the root cause of slugging would at that point require an analysis of the historical production data. The hypothesis was that by analyzing the various operating points, identifying the factors that caused a high degree of slugging would be possible. The primary question was whether slugging was caused by riser conditions or problematic conditions in individual wells.
A method suitable for a data-driven analysis of production data, including more than 100 sensor measurements, was devised and implemented. Key components included an algorithm for processing the large amount of production data and several machine-learning methods capable of ranking features (factors/explanatory variables) based on their predictive importance.
A two-step approach was followed to identify likely causes of slug flow. First, a slug-severity measure was designed and applied to all historical production data. Second, machine-learning techniques with feature-ranking properties were applied to create a model ensemble. These models were used to identify the main predictors for slug behavior, using as much data as possible. However, because the goal was to improve steady-state production, intervals with transient behavior, such as well ramp-ups and choke changes, were removed from the data set before training the models. Intervals with little or no production were also removed. The complete paper describes the two steps in detail.
Results and Discussion
Rankings from the initial screening and from the recursive-feature-elimination step are shown in Figs. 1 and 2, respectively, of the complete paper. The riser liquid rate was the highest-ranked feature according to the random forest ensemble, with a clear positive correlation between slug severity and riser liquid rate. While the random forest technique preferred riser measurements, in line with the expected outcome, lasso regression often prefers well measurements over riser measurements. This is partially explained by flush production (increased flow during the first days after startup). Data points associated with flush production in one or more wells are marked in Fig. 1. Several cases of temporarily increased severity exist after well startup, which explains why some wells were theorized to be causing slugging. However, no evidence exists that these wells were more problematic than others after they had been stabilized. The key takeaway, therefore, is that riser liquid rate was the main driver behind slugging in this case, with water being slightly worse than oil.
- Quantifying slug severity enabled the use of supervised learning to build predictive models for slugging. Training the models with actual production data made it possible to use feature-importance ranking to identify likely drivers of slugging.
- The slug-severity measure was differentiated successfully between low, moderate, and high slugging. However, the measure is noisy, and some tradeoffs had to be made in terms of what could be represented. These tradeoffs can be asset-specific.
- After identifying liquid rate as the main driver, a well branch with significant water production was identified and closed. The operator confirmed that by closing the branch, the slugging was reduced. As a result, total oil and gas production could be returned to the level seen before the onset of slugging and subsequent choking of riser production.
- Applying machine learning is by itself insufficient to understand the root causes of slugging in the presented case. Only by combining machine learning with engineering knowledge can this goal be achieved. However, this data-driven approach accelerates the analysis of large and high-dimensional production data greatly. As demonstrated in this case, it can provide valuable insights for important decisions regarding slug mitigation and avoidance.
- The method described in the complete paper is not specific to the studied case. The same approach can be used on other fields experiencing slugging. One of the great advantages of data-driven approaches is that they are easily automated and used in real-time settings where decisions must be made rapidly.
- Actions were implemented to reduce water production, and this led to reduced slugging.