SPE SET Number FILLIN “What is your paper number?” 1 1XXXXX
FILLIN “List all authors and companies” * CHARFORMAT Author, Montanuiversitaet Leoben
This paper was prepared for presentation of the Literature Review Project held at the University of Leoben, Leoben, Austria 26. June 2018.
The production and operations in oil and gas fields are becoming more complex with more demanding restrictions for safety, energy efficiency and environment. And although artificial lift methods are well known and widely used, of todays oil wells, over 90 % are using artificial lifting systems, there are still ….
ESP pumps are responsible for more than 50% of the worldwide oil production, even though their number is much lower than sucker rod and other types of pumps. The reason behind this is their high production capacity. They can also be modified to accommodate a range of production capacities, while achieving any possible drawdown pressure. They also perform impressively in deviated wells and high depths due to their less moving parts.
ESP compared to the other pumping units is very sophisticated and require high maintenance and skilled proffesionals to deal with design and installation to operating troubleshooting. Causes of failure of an ESP pump begin as early as in the designing stage through installation and up to improper operating conditions.
During the operation stage in the lifetime of an ESP, Monitoring and diagnosis are the critical perquisites for an optimal production and minimizing required maintenance. The operator needs to understand and be able to evaluate the given data output and analyse any abnormalities in the pump. There are 3 main parameters that are used to analyse and diagnose the performance of the pump; production rate from surface readings, Motor current from ammeter chart and pump intake pressure (PIP).
Based on these 3 parameters, different ESP diagnosis and monitoring levels has developed in the past years. In the beginning, operators used only ammeter card readings to diagnose the ESP motor performance, then acoustic surveys and downhole measurements were introduced. Table 1 shows the different levels of ESP monitoring used in the industry.
Even though all of these real-time operating parameters are available, analyzing and understanding the information is still insufficient and highly dependable on the skill and knowledge of the operator. Even with high skilled operators some readings and patterns are still ambiguous and cannot be interpreted in a real time manner. Moreover, Ammeter card dignoses does not provide information about VLP, efficiency and pipe leakage. Therefore, several new approaches are developed using neural networks and artificial intelligence to handle the enourmous amount of real time data produced and recognize different problems and diagnose the root causes.
In this paper, we are going to discuss several new approaches for automated ESP diagnosis. These approaches and methods use neural networks and artificial intelligence for different methods of ESP monitoring and diagnosis.
Figure SEQ Figure * ARABIC 1: Flowing System of Baker Hughes Artificial Lift Research and Testing Facility. Reference
Artificial Neural Network
An Artificial Neural Network is a mathematical computational model which is based on the human neural system. A computational unit in ANN is an artificial neuron which is emulation of the human neuron, which is interconnected to other artificial neurons and process information using a connectionist approach to computation Artificial Neural Network is a mathematical computational model which is based on the human neural system. A computational unit is an artificial neuron which is emulation of the human neuron, which is interconnected to other artificial neurons and process information using a connectionist approach to computation. The following figure shows a schematic design of an artificial neuron.
Figure SEQ Figure * ARABIC 1: Flowing System of Baker Hughes Artificial Lift Research and Testing Facility. Reference
Since Neural networks are capable of recognizing trends or patterns in large volumes of data it can be best suited for solving the problem of Electric Submersible pump surveillance. Traditional approach for ESP surveillance involves real-time monitoring of ESP data both surface and downhole. These systems either make predictions using complex engineering calculations or makes the real-time data available to experienced Petroleum engineers and experts who can make decisions of the health of ESP system.
Since a neural network can be trained to effectively recognize trends or patterns, it can be employed in ESP surveillance and hence once it has learned from training data sets (containing ESP failure or conditions which led to ESP failures), it may be a valuable resource for effectively surveying the ESP operations and generate Alerts/Alarms for experienced experts and petroleum engineers to validate the warnings raised by ANN ESP surveillance system. Some design of ANN for ESP surveillance consists of a three layer neural network – input layer will take a mixture of surface and downhole data variables, hidden layer will process the input variables and present an input to the output layer which will be a set of vector containing output values for common ESP failures.
Fault tree analysis
Fault tree analysis (FTA) is an important analysis method for fault diagnosis. Various faults are described in an inverted tree structure according to the internal cause and relationship. In fault analysis, the direct causes and indirect causes of the fault can be found out from a point of failure. This analysis method can accurately locate the fault, and it is easy to understand.
The complexity and variety of the fault condition have put forward higher request to the accurate diagnosis of the electric pump well condition, but there is a certain causal relation between these faults, which can be described by the fault tree. The most important two elements in the fault tree are as follows:
Event: Mainly divided into top event, basic event and intermediate event. The top event is the starting point for the current fault diagnosis, that is the fault to investigate; Basic event is the ultimate causes of other events, and there is no need to further explore the reason for the occurrence; Intermediate event is the event that between top event and basic event.
Logic gate symbol: Used to describe the logical relations between events, including AND, OR and NOT gates.
In the diagnosis of motor working condition, as the motor efficiency is an important indicator to measure the motor condition, three levels of the motor working condition can be obtained by this index: good (efficiency is greater than 70%), the general (efficiency is between 40% to 70%) and poor (efficiency is less than 40%). Therefore, these three levels can be used as the top event in the fault tree, the following is the “motor condition is poor” as an example, the establishment of the fault tree, as shown in the following figure:
Figure 2: Example of fault tree diagnosis for motor condition of electric pump well Reference
As shown in Fig. 2(a), T1 is top event; A1, B1 and B2 are intermediate events; X1 to X6 are the basic events. According to the fault tree, the combination of the basic events leading to the occurrence of the top event can be obtained based on qualitative analysis. The minimal cut set of the fault tree is one of the most important methods used in the qualitative analysis, which is based on the Boolean algebra rules.
The slove process of the minimum cut set for the top event T1 is as follows:
As shown in Eq.1, fault tree minimal cut set is converted to: K1 = (X1 X2 X6), K2= (X3 X4 X5 X6).Once the minimum cut set is obtained, which means that when a set of basic events occur, it may cause the occurrence of the top event, so the fault tree is simplified as shown in Fig. 2 according to the minimum cut set.
The knowledge in expert system is used to simulate the way of thinking of experts, and the knowledge base is used to store knowledge. In the field of artificial intelligence, knowledge expression has a variety of ways, such as production, semantic web, etc. Production rule is widely used in expert system, its basic form is IF… THEN…, if a particular condition is met, a result is produced. Similar to programming language syntax, IF can be followed by a number of conditions that connect with the logical operator AND, OR, NOT.
Each of the minimum cut sets corresponds to a failure pattern of the top event, and for the knowledge base, every minimum cut set corresponds to knowledge. When building a knowledge base based on the production rule, each cut set is converted into a rule. Such as R1 that converted from K1: IF X1 AND X2 AND X6 THEN T1.
Wellhead pressure Buildup Diagnosis: (real time monitoring paper)
SCADA-driven surveillance systems organize and store key operating parameters in a centralized location for remote monitoring, analysis and control. The ESP’s downhole sensor provides data to a remote telemetry unit on the surface that transmits real-time data into the surveillance system via a satellite or cellular modem. Field operators and engineers are automatically notified of operating problems and are able to review well performance and make remote adjustments to equipment. This allows operators to optimize production, identify data anomalies and correct problems before they lead to costly system failures
Once the ESP suffers problems, the pressure buildup curve appears different and deviates from the one in normal conditions. Therefore, the analysis of wellhead shut-in pressure buildup should be a practical diagnosis of ESP malfunction.
Fig. 1 illustrates the well shut-in process. Fig. 1a represents the well before being put on production. Fig. 1b illustrates the normal production and the flowing fluid is a mixture of liquid and gas. Fig. 1c depicts the early period after well is shut-in. The gas and liquid phases start to segregate. The liquid level rises to the bottom red dash line because of the reservoir pressure support, and the top red dash line is achieved after gas is compressed as the ESP head increases in Fig. 2. Fig. 1d shows the final shut-in scenario.
Figure 3: Wellhead shut-in pressure buildup test procedures. Reference
To estimate the wellhead shut-in pressure change with shut-in time, pump throughput is required. Based on the evolution of wellhead shut-in pressure as shown in Fig. 3, the pump outlet pressure head can be mathematically described by a quadratic equation related to pump throughput in Eq. 6. The constant C is determined by the ESP performance curve, and the curve should be provided by manufacturers or field bench test.
pump outlet pressure head can be also calculated from wellhead pressure, which is
The universal gas constant and water density are known. Using the wellhead pressure to approximate the average pressure of gas column and substituting gas and water densities into Eq. 7 we have
Combining Eqs. 6, 14, and 15, we have
According to mass conservation, the mass of pump throughput is equal to the sum of mass changes of liquid and gas in wellbore, which is
Substituting Eq. 16 into Eq. 5, we obtain a correlation between wellhead shut-in pressure change and shut-in time
Eq. 17 gives the theoretical wellhead pressure with time during buildup test, which can be represented by the red solid line in Fig. 3.
The theoretical maximum wellhead shut-in pressure is given by
Figure 4: The evolution of wellhead pressure with shut-in time during wellhead shut-in pressure buildup test.
The red solid line in Fig. 3 represents the theoretical pressure buildup curve after well shut-in, as derived from Eq. 10 in the previous section. We use two pressure gradients, Gradeth_1 and Gradeth_2 in two half periods to capture the curve characteristics. Applying the same approach of Fig. 3, we can obtain the actual maximum wellhead shut-in pressure pa_max. The workflow chart to diagnose ESP malfunction using wellhead pressure buildup is shown in Fig. 4. The threshold values (0.78, 1.5, and 1) used in this model were determined from field experience provided by subject matter experts. This method, wellhead pressure buildup diagnosis, is completely based on the above developed equations. It is specifically useful for detecting tubing/ESP chamber leakage.
Figure 5: The work-flow chart to diagnose ESP malfunctions by wellhead shut-in pressure buildup.
Back Propagation (BP) neural network analysis of ammeter card
In the current field practice, the analysis of ammeter card is performed by human pattern recognition. The individual knowledge, background, and experience can vary significantly among engineers. This leads to materially different intepretations for the same ammeter card. Another shortcoming of human interpretation is the low efficiency in terms of time taken, amount of data reviewed, and consistency of diagnosis. To overcome these disavantages, Back Propagation (BP) neural network analysis is applied to diagnose ESP malfunctions. The development of a BP neural network tool for diagnosis of ESP ammeter cards requires the following steps:
Collect the field data that are corresponding to different operating conditions;
Develop a database to capture different operating mulfactions –to be elaborated in the section of “Sample Database Development”;
Characterize different operating mulfactions by eigenvalues –to be elaborated in the section of “Extraction of Sample Characteristics”;
Train the neural network and generate a weighted-factor matrix based on the developed database mulfactions –to be elaborated in the section of “Neural Network Training Procedures”;
The recognition model is consequently used to diagnose ESP malfunctions on field ammeter cards.
Sample Database Development
To develop a comprehensive database, it is important to collect sample ammeter cards for different malfunctions and it should avoid taking too many similar samples for any specific one. Sample ESP ammeter cards can come from two sources:
Field cases –equal sample number for each malfunction is preferred.
Simulated ammeter cards – because some ESP malfunctions do not have enough sample ammeter cards in the available field data and public domain, an ammeter card generation program has been developed separately to create different shaped cards based on the required malfunction characteristics. Those were used to complete and improve the pattern recognition capacity.
Extraction of Sample Characteristics
The ammeter card characterization description is critical in pattern identification. The input eigenvalues,corresponding to the weighted-factor matrix of netrual networking, must be (1) distinguishable; (2) numerically quantifiable; (3) unique, which is only owned by itself; (4) independent. Note in any ammeter card of Appendix B, the outside circle represents 24 hours, and the marked radius represents the electric current range. The red line is the recorded current change in of the day at interest. According to this information, we analyzed all the possible causes for ESP malfunctions and suggested the following diagnosis workflow. Consequently, we identified 45 characterization parameters (E1-E45), which are indeed the input eigenvalues in the neural network analysis. Those eigenvalues can satisify the four criteria mentioned above:
E1 is the number of pump stop, which is equal to Noff.
E2 is the number of electric (voltage) fluctuation, which is equal to Nfluc.
E3 is the minimum period of a SINGLE fluctuation, tfluc in 24 hours.
E4 is the maximum period of a SINGLE fluctuation, tfluc in 24 hours.
E5 is the summation of ALL the lasting fluctuation time, _tfluc in 24 hours.
Next, we defined the minimum of electric current change rate in the ammeter card in each 2-hours period, so 12 (E6 to E17) eigenvalues were defined.
Similarly, we defined the maximum of electric current change rate in the ammeter card in each 2-hours period, so 12 (E18 to E29) eigenvalues were defined.
To improve the recognition resolution, we also defined the minimum and maximum of electric current change rates in the ammeter card in each 4-hours period, which are, E30 to E35, and E36 to E41, respectively.
E42 is the ratio of fluctuation, tfluc, to smoothness periods, tsmooth.
E43 is the ratio of circle perimeter, L, to area, S.
E44 is the maximum well flowing period, which means the electric current is always positive during this period.
On the other side, E45 is well no-flowing period, which means the electric current is 0.
These 45 eigenvalues compose the input layer, and the final output layer is the probabilities of 10 ESP working status. Therefore, for the 2-layer neural network, the weighted-factor matrix is in the dimension of 45 rows and 10 columns (45X10). During our training process, the convergency was not satisfied. The we employed a 3-layer neural network analysis with a hidden layer of 15 intermediate nodes in Fig. 6.
Figure 6: The layout of BP neural network in the model used in this study.
Neural Network Training Procedures
In the neural network training process, a good selection of feeding sample batch not only shortens learning time, but also avoids possible training instability, and it further improves identification accuracy. Therefore, the samples used to train this neural network comply with the following rules:
The neural networking involves multiple rounds with different sample populations. It started with a small volume of samples and achieved convengency first, and then we gradually increased the sample populations to strengthen the recognition ability. In the favor of the aforementioned ammeter card generation program, the training samples can be completed theoretically.
In each training round, all samples are fed in and used equally.
In each training round, all samples affiliated to a specific ESP malfunction are fed in together and continuously.
The learning rate is set at 0.4 in a balance of the training time and converging rate.
The workflow chart of the neural network training is shown in the Fig. 7.
ESP pattern recognition analysis
Figure 6: The workflow chart of neural network training.
ESP pattern recognition analysis
By developing an agreed analytical technical and mathematical tested Pattern to all the common failures and trips in PDO’s ESP operation, the failures will be identified and even be able to be predicted faster for a quick reaction and optimum solution. In addition, pattern Recognition will enable the engineers to monitor huge data in a particular field for a group of wells.
As the failures are repeating, and with the vast real time data, the failure can be predicted through ESP EBS approach once the Pattern Recognition analysis cases have been tested using Nodal Analytical Tools while cross checking with the Real Time trend, the common trips and failures can be predicted for pro active trouble shooting and remedial work.
During ESP operations, there are a number of changes can happen downhole or at surface which can disturb the normal operation of ESP which at the end can cause a trip or failure. These changes can be mechanically; hydraulically or electrically, and can be at the surface; in the tubing / ESP equipment or even in the reservoir. Some of these deviations can actually improve the ESP performance in term of production. After operating these fields for many years with the same reservoirs and fluid characteristics; and after experiencing various types of failures while matching with the real time data trending, the Pattern was derived from various common cases.
PDO have identified 13 cases which can derive a specific unique pattern in term of the change of the measured parameters (variables) which can be used to analyze the ESPs as part of ESP diagnosis. These identified pattern cases are listed as follow:
Hole in Tubing
Blockage at Pump Intake
Blockage at Perforations
Increase in Water Cut
Shut in at Surface
Blockage in Pump Stages
Increase in Reservoir Pressure
Increase of free gas at Pump Intake
Wearing Stages (erosion)
Increase in Frequency
Open Choke (decrease in WHP)
The arrows indicate the rate of change of the variable. The coloured boxes indicate the unique characteristics of the response. The best trip parameters are indicated by Trip.
Figure 7: Pattern Recognition Analysis checklist
Case study using (Wellhead pressure Buildup Diagnosis and BP neural network analysis)
An offshore well has experienced sharp production decline in field, but the well electric power consumption keeps constant. The operation team desired to find the root cause, and we applied this systematic diagnosis. Firstly we applied the neural network analysis on the ammeter card recognition. Fig. 8 shows the real-time surveillance data of electric current along with time, which were transimitted from offshore platform to onshore central office.
Figure 8: Real-time plot of current vs. producing time
Fig. 8 represents a 24-hours electric current variation, which can be plotted in the form of ammeter card as shown in Fig. 9.
Figure 9: Real-time plot of current vs. producing time in form of ammeter card
Artificial Lifting Systems play a very important part in today’s oil industry. Therefore it is only natural that the race for new ….
1 Hofstätter, H. 2014. “Artificial Lift Systems” Lecture Notes. Chapter 1. Montanuniversität Leoben
2 Ghareeb, M., Ellaithy, W.F., Zahran, I.F. 2012. Assessment of Artificial Lifts for Oil Wells in Egypt. SPE 160719. Prepared for presentation at the SPE Annual Technical Conference and Exhibitiom held in San Antonio, Texs, USA, 8-10 October 2012