Work smarter, not harder

But the task of satisfying this growing demand is becoming ever more difficult with each passing year. There is hardly any ‘easy’ oil left and the future is set to be one of increasingly complex horizons with a pronounced heterogeneity of productive formations at even greater depths.

In view of the high demand for ‘black gold’ and the difficulty of working the reserves, oil producers are actively seeking new technologies that will enable more efficient development of existing fields to satisfy demand and hold down production costs. One method of enhancing oil efficiency is to manage the well’s flow rate.

Managing oil production rates: modern capabilities

Electric submersible pumps (ESPs) capable of working in difficult conditions at depths of over 4 km have become widely used in the oil production sector. They are deployed in Russia, the US, China, Canada and Southeast Asia, where there are many brownfields with significant residual reserves.

ESPs can work in a wide range of settings (± 30% of nominal capacity), and if producers also use control stations with frequency regulation, they can change the shaft speed at any moment in time and thereby influence the rate of fluid production from the formation. Well flow rates cannot be changed in this way with any other existing production method.

Figure 1. Illustration of the algorithm that is used to recalculate the measured fluid flow rate at the data preparation stage. The illustration allows comparison of the difference between the measured fluid flow rate and the actual fluid flow rate (on the left side). If the measured flow rate was equal to the actual one, the ESP operation cycles would look like the cycles on the right side of the illustration. The yellow line shows the period of fluid flow rate measurement. The red line shows operation cycles of ESP during 24 hours. 

The principle of altering the flow rate is simple: when the current frequency increases, the pump’s working assemblies start to rotate faster and the volume of liquid produced from the well increases, and vice versa.

By regulating the shaft speed, it is possible to ensure optimal production of formation fluid.

Over time, however, the reservoir energy changes, as does the volume of fluid that can be produced from the well. In addition, there are more short-term factors that determine the potential production rates, such as the rate of water injected to maintain formation pressure. For this reason, it is not as easy as it might seem to determine the optimal fluid production rate.

Decisions to intervene in the operating mode of an ESP are currently taken by engineering and process staff in the production department. Their decisions are aimed at ensuring that pumps are operating in optimal mode through adjustments to the rotation speed, the duration of the ESP operating cycle, and the ratio of production time to fluid accumulation within the cycle.

The optimisation techniques used by different companies are often similar as they are underpinned by the same basic algorithms and patterns. Yet the decisions taken are not always effective, as they do not take into account the impact of the change in ESP operating mode on all subsequent production phases.

It is possible however to maximise a well’s potential through deployment of advanced equipment and integrated evaluation of the potential of the entire production chain using machine learning technology.

Experts in artificial intelligence (AI) at Zyfra have recently finished testing an AI solution based on machine learning for optimal control of oil production mode at a field in West Siberia, Russia.

Solution

The solution is based on algorithms which take into account the following well operating parameters and their correlations:

• Speed of rotation of the electric motor shaft.
• Pump start – stop (1 – on, 0 – off).
• Water cut (%).
• Density (kg/t).
• Temperature of motor (°C).
• Active power (kW).
• Well-head pressure (kgf/cm²).
• Nominal power of motor (kW).
• Motor current (Amperes).
• Input voltage (V).
• Reservoir pressure (kgf/cm²⁾.
• Formation pressure (kgf/cm²).
• Nominal daily rate (nominal pump capacity) (m³/d).
• ESP setting depth (m).
• Extension to formation perforation interval (m).
• Dynamic level (m).
• Annular pressure (kgf/cm²).
• Intake pressure (kgf/cm²).
• Measured fluid rate (m³).

The solution consists of five main modules:

• Data preparation module.
• Data assurance module.
• ‘Well-formation’ system simulation module.
• Recommended mode computation module.
• Economic impact computation module.

The process is initiated when data is filtered into the data preparation module to remove inaccurate measurements, sensor surges etc. Auxiliary aggregate parameters are also devised in this stage. One of the main parameters is true produced flow rate, which is particularly important for wells operating in short-term production mode.

Figure 2. Initial black box modelling approach of the 'formation well pump' system. Controls - a range of rotation speeds, ratios of production times to fluid accumulation within the cycle, observations - key measured parameters. 

The data assurance module then checks that the data is sufficient for subsequent construction of algorithms and machine learning models.

One of its other functions is checking the wells to determine the accuracy and resiliency of the data for devising recommendations. Recommendations can be produced only for wells whose operation is described in some detail and whose operating mode can be regarded as stable. The criterion for data sufficiency is the availability of all critically important parameters (CIP) at a given frequency: motor temperature, current, intake pressure, cable insultation resistance etc. The stability of the data is assessed using basic time sequence prediction techniques. If the variance in assessments by these methods using historical values does not exceed specified limits, the well is deemed to be operating in stable mode.

Figure 3. General illustration of the final modelling approach. The agent based on the predicted observations checks the feasibility of adjusting in accordance with physical constraints and initiates prediction recalculation.

The third module simulates the ‘well-formation’ system. The element of greatest interest within this module builds a predictive model for fluid flow rate and the well CIPs following any alteration to the well’s control mode. The predictive model consists of two parts: an individual forecast for each well and an overall forecast for all wells in the field. Several machine learning models are used, including random forest, linear regression and others. The resulting forecast for each of the prediction parameters for each well is calculated as a linear weighted sum of the individual and overall forecasts with a derived confidence factor. The confidence factor is calculated on the basis of data sufficiency for a description of the well’s operation.

In addition to the measured values of physical parameters, increments of those values are used as input data to build the prediction. This removes the constant measurement error or differences in the measurement of physical parameters and enhances the accuracy of the general predictive model for all wells.

Figure 4. Details of the final modelling approach. The 'well-formation' system is modelled based on statistics, pump model is physical. 

The models learn from historical data recorded by the telemechanics systems over several years.

Next, by comparing the flow rate predictions and all CIPs, an ESP operating mode is selected to maximise the volume of fluids produced with minimal risk.

The system generates two types of recommendation:

• Increase in motor operating frequency (or in ratio of operating and idling time in the case of short-term operation) to boost the production rate.
• Reduction in frequency to reduce the risk of well shutdown (drying out, overheating of motor etc.).

Recommendations on the optimal well operating mode are issued every day. The current requirement is for these recommendations to be evaluated by an engineer, who then gives the order to adjust the pump’s operating mode on that basis. Automatic control of the pump based on the recommendations is also possible though.

Figure 5. Principle of 'well-formation' model update. 

The solution also performs a calculation of the target economic indicators for each change of operating mode and an analysis of plan vs actual. This helps to properly evaluate the accuracy of the predictive models and provides feedback for updating them.

It could be said that the machine has enabled people to see all sorts of interconnections in the production process and, as a result, to maximise the potential of the wells by controlling the oil production mode on the basis of recorded data.

Aspects of the approach and modelling

The solution functions through a combination of physical and predictive-statistical models.

The pump model in the solution consists of a physical model, into which all the currently known physical patterns of operation of the ‘formation-well-pump’ system are loaded. The statistical part is designed to determine the short- and long-term reactions of the ‘well-formation’ module. From the physics of the process, restrictions are set on maximum frequency, load, current etc. In other words, if the mathematically derived recommendations conflict with the physics of the process in terms of upper limits on parameters (e.g. the system calculates that the maximum flow rate will be achieved at a frequency that is impermissible for the well in question), the system simply rejects them and issues the most effective implementable recommendations.

Test results

This solution has been tested for three months in 500 wells at the oilfield. The results have shown that it is possible to boost oil production by an average of 1.5% by adjusting the well operating parameters. For example, by applying the optimal mode recommended by the system, a well that produces 50 m³ of oil in normal operating mode can produce 51 m³ of fluid.

Conclusion

The solution can be seen as part of the ‘smart oilfield’ concept, which is predicted in some quarters to be the future of the industry. Gartner has calculated that the new concept could help oil companies cut costs by 5% and increase oil production by 2%.² Similarly, Cambridge Energy Research Associates claims that the ‘smart oilfield’ could lower lifting costs by 1 – 6% while reducing oil well downtime by 1 – 4% and labour intensity by up to 25%.³

It is therefore anticipated that machine learning-based technology for controlling oil well operating mode will be much sought after in the oil industry, especially in countries with a large number of brownfields.

References

1. EXARHEAS, A., ‘IEA Cuts Oil Demand Growth Forecast’, https://www.rigzone.com/news/iea_cuts_oil_demand_growth_forecast-15-may-2019-158835-article/ (May 2019).
2. MCAVEY, R., and CUSHING, S., ‘Oil and Gas Industry Innovation Through Digital Technology Primer 2019’, https://www.gartner.com/en/documents/3900966/oil-and-gas-industry-innovation-through-digital-technolo (February 2019).
3. WOOD, T., ‘The Connected Oilfield’, https://www.cisco.com/c/dam/en_us/about/ac79/docs/wp/Connected_Oilfield_0629b.pdf (May 2007).

Read the article online at: https://www.oilfieldtechnology.com/digital-oilfield/19082019/work-smarter-not-harder/