Oil wells in Western Siberia usually placed on artificial drilling pads, forming well clusters up to 30 wells. The flow rate of each well in the cluster measured by an automatic measuring unit one by one. Often flow rate measurement requires several hours and flow rate of a single well can be measured once a week or less. This led to situation then events affecting well rate can be invisible between measurements. Identifying such events can be extremely useful in many cases, for example for wells with unstable behavior or transient regimes. The same challenges are also faced at distant green fields during their development, there the flow rates can be measured once a month with a mobile unit. The objective of this paper is to develop a virtual flowmeter model based on indirect high-frequency data of well operation and ESP.

In Gubkin University, at the Petroleum Reservoir and Production Engineering Department, bench tests of ESP5-50 (118 radial stages) on gas-liquid mixture in a wide range of volumetric gas content (βin = 0-60%), intake pressure (Pin = 0.6-2.1 MPa) and pump shaft speed (n= 2400-3600 rpm) were performed. Three vibration sensors were installed on the unit: on the ESP, at the ESP discharge, on the pipeline, which simulates the wellhead production tree. During the bench tests were recorded series of pressures at the intake, discharge and along the pump length, series of current and power consumption, as well as vibrations with frequency several times per second.

Based on the bench test results, we investigated the possibility of indirect determination of well operation parameters during artificial lift modelling by machine learning. As a result, the approaches to modelling taking into account various sets of parameters (features) have been studied:

  1. based on hydraulic parameters – ESP intake and outlet pressure;

  2. based on hydraulic and electric parameters – current and power consumption;

  3. based on hydraulic, electric and vibrating parameters.

The analysis of data series allowed to define the boundaries of stable ESP operation, namely the transition to surging and pump starvation.

The novelty of the work is:

  • machine learning modeling of the gas-liquid mixture pumping process by electric submersible pump;

  • solving both direct and inverse issues: as virtual liquid flowmeter as, virtual gas content flowmeter at the pump intake.

You can access this article if you purchase or spend a download.