S SPE 163704 B Benefit Evaluation n of Keep ping an In ntegrated d Model D During Re eal-Time E ESP O Operation ns T T. Denney, SP PE, B. Wolfe, SPE, D. Zhu u, Baker Hugh hes Inc. C Copyright 2013, Society y of Petroleum Enginee ers T This paper was prepare ed for presentation at the 2013 SPE Digital Energy E Conference and d Exhibition held in Thee Woodlands, Texas, U USA, 5–7 March 2013.. mittee following review of information containned in an abstract subm mitted by the author(s)). Contents of the pape T This paper was selected for presentation by an a SPE program comm er have not been re eviewed by the S ociety y of Petroleum Engine eers and are subject to o correction by the a utthor(s). The material ddoes not necessarily re eflect any position of t he Society of Petroleu um Engineers, its officers, or members. Electronic E reproduction n, distribution, or st ora age of any part of thiss paper without the wrritten consent of the S Society of Pet roleum E Engineers is pr ohibite ed. Permission to illustrations mayy not be copied. The abbstract must contain co onspicuous acknowled dgment of SPE copyrig re eproduce in print is res stricted to an abstract of o not more than 300 words; w ght. Abstract A p (ESP) ap pplications, a great g deal of tim me and effort iss spent upon thhe initial sizingg of the IIn most electricc submersible pump ESP – projectin ng which equip pment is most appropriate a forr the physical pproperties, reseervoir delivery,, and operator eeconomics E A the ESP is i designed, ordered, and insttalled, howeverr, the sizing is often placed inn a oof the particulaar application. After n referred to again until thee ESP has faileed and the nextt ESP is being ddesigned. By iintegrating thee design rrepository and not model into a real-time monito oring environm ment, more valu ue can be demoonstrated. Inforrmation that is particularly vaaluable m oint, ESP diagnnostics alarms,, virtual monitooring and alarm ms, and inncludes the evaluation of opeerating point veersus design po monitoring of sub-component s t design thresh holds. In this paaper, we will reeview ESP dessign software inn general, the bbenefits of m mplementing a real-time mod del comparison n tool, demonsttrate several caases, and concllude with a disscussion of the challenges im h managing thee models. Addittionally, we wiill examine how w sizing can be used to optim mize productionn and aassociated with make adjustments for system enhancement. m IIntroduction Electric submerrsible pumps (E E ESPs) have become one of th he most populaar and widely ddeployed formss of artificial liift in the world. ESPs offfer operators a large degree of w o durability an nd reliability accross a host of applications, including slim--hole oil wells, high-production water wells, and coal bed methane wells. As techhnology has im w mproved, ESPs hhave been pushhed to eeven more challlenging enviro onments, includ ding steam assiisted gravity drrains (SAGD) and deepwaterr applications. When reviewin W ng ESP failuress, it has been fo ound that an alaarming rate off failures could have been prevented. Exam mples of ppreventable faillures include in ncorrect sizing g, improper setu up, operator errror, and failuree to follow maiintenance proccedures. The costs assocciated with prev T ventable failurres are significaant, and includde those from ddeferred producction, replacem ment eequipment, rig//work over costs, and significcant opportunitty costs from enngineering andd field personnnel. All of the aforeementioned preesent strong inccentives for reaal-time operatioonal surveillannce. Operationaal surveillance is not new A ffor ESPs. Most operators hav ve incorporated d their ESPs into a Supervisoory Control andd Data Acquisittion, or SCAD DA, m ESP manu nnetwork, and most ufacturers have begun offering g after sale moonitoring servicces. While tradditional SCADA A methodologies,, such as set po m oint alarming an nd trending, caan offer signifiicant value to thhe operation, aadditional valuue can be ggained by using g some non-SC CADA techniqu ues. This paperr will explore tthe benefits of integrating a m mathematical m model into thhe SCADA f an ESP operattion. 2 SPE ESP Design and Modeling In most of today’s operations, a software model is used to approximate the physical conditions the ESP will experience. In Table 1, Baker Hughes (Baker Hughes Centrilift, 2008) presents the data normally included in a typical ESP sizing model. Table 1: ESP design considerations Data Value Static Casing or liner size and weight X Tubing size, type and thread X Perforated or open hole interval X Pump setting depth X Dynamic Wellhead tubing pressure X Wellhead casing pressure X Test production rate X Producing fluid level (Flow Pressure) X Static fluid level (Static Pressure) X Bottom hole temperature X Desired production rate X Gas-oil ratio X Water cut X Specific gravity of produced fluids X Bubble point pressure of gas X Viscosity of oil X PVT data X Power sources: primary voltage, frequency, power source capabilities Other considerations: sand, deposition, corrosion, paraffin, emulsion, gas, temperature X X As shown above, the majority of the tags necessary are dynamic. During sizing, this is normally dealt with by developing several different static ‘cases’. The cases produced generally represent the best and worst scenarios for production rates to ensure the ESP is sized correctly. The sizing software gives the application engineer many tools to select the ‘right’ pump for the application. The most powerful tools within the sizing software are the equipment ‘flags,’ which notify the application SPE 163704 3 engineer when equipment constraint is being encroached. Equipment constraint is typically a physical constraint such as mechanical, hydraulic, or electrical issues. One of the biggest challenges associated with ESP design is the level of uncertainty. For the average brownfield well, a deep knowledge of the reservoir productivity and fluid characteristics is difficult to obtain at time zero, with further uncertainty introduced as the well is produced. However, by integrating the model with the real-time data, the operator can obtain the insight and care that was taken during the design process, and carry that knowledge further to represent a complete operating envelope for the ESP (in addition to the well and pump behavior changes). Furthermore, the design model can be used as a diagnostic tool to further our understanding of adverse conditions. ESP Diagnostics An ESP is a complicated system to be mathematically modeled. It involves domain knowledge of mechanical, hydraulic, and petroleum engineering, thermodynamics, and electrical engineering. Fuzzy logic is an innovative technology that can circumvent the need for rigorous mathematical modeling. Before the advantages of fuzzy logic are presented, a typical SCADA alarm system for ESP operation will be examined (Thornhill & Zhu). To protect the ESP motor, it is important to monitor the motor’s operating temperature. The conventional SCADA system defines an inflexible limit for a temperature alarm (for example, the ESP motor cannot operate above 300°F). There are two flaws to this approach: 1. 2. What if there is an issue with the temperature sensor? For example, the real temperature is 302°F, but the reading from sensor is only 298°F because of inaccuracies in the sensor’s measurement. In this case, the conventional SCADA system will not issue the alarm for an overheated motor. The traditional SCADA system is not robust enough to handle signal noise or a false sensor reading. Additionally, SCADA systems tend to treat alarms as black or white. Either the system is in a state of alarm or all systems are clear. From a mechanical and electrical point of view, an ESP with a motor temperature of 298°F is not in much better condition than an ESP operating at 300°F. Both situations require an ESP engineer be notified to take proactive action to prevent a catastrophic failure. The SCADA system, however, will only issue an alarm when the temperature reaches 300°F, ignoring the other situation. One could argue that the alarm condition could be made looser. Overall, however, this only produces more alarms (often false ones) and still produces a black/white boundary between alarms and all-clear conditions. In contrast, fuzzy logic performs reasoning based on uncertain or imprecise information. Table 2 outlines the differences between fuzzy logic and conventional crisp logic in an example of shutting down an ESP when high temperatures are observed. Table 2: Comparison of Fuzzy Logic vs. Crisp Logic for ESP shutdown How to describe hot? True or false Proactive action support Noise handling Fuzzy Logic Approach 290°F will be close to hot, but to a degree of 0.9; 299°F will be close to hot to a degree of 1. Partial true or false exists Yes Good Conventional Crisp Logic Approach 300°F exactly is hot, 299.99°F is not hot Nothing between true and false No Poor Another advantage of using fuzzy logic is its tendency to mimic human behavior. In the example above, a person knows to shut down the motor when the temperature is too high. In fuzzy logic terms, we can set up an “IF—THEN” rule to declare “if the motor temperature is too high, we need to shut down the motor.” The beauty of fuzzy logic lies in the linguistic term “too high.” Note that with fuzzy logic, we do not say “if the motor temperature is above 300°F, we need to shut down the motor.” 4 SPE The linguistic term “too high” allow us to assign a degree of ‘hot’ beyond 0 and 1. In reality, “too high” will allow a system to shut down a motor when temperatures reach 298°F, because from a fuzzy logic point of view, 298°F is close to “too high” to a degree of 0.98. The diagram illustrates the components that make up a fuzzy logic system: Fig. 1: Components of a fuzzy logic system Inputs Inference Engine Outputs Rule base During ESP operation, a SCADA reading such as motor temperature will be one of the inputs. The If-Then rule (“If motor temperature is too high, we need to shut down motor”) will be one of the rules stored in the rule base. The inference engine of the fuzzy logic system works much like the human brain. People use common sense and knowledge (inference engine) to make decisions every second of the day. For example, a human being that feels the air temperature can tell if they need to wear a coat or not. The air temperature does not have to drop below 50°F for a human to receive an alarm. The human can feel degrees of cold and make a decision when a coat is necessary. An inference engine will first convert the linguistic term “too high” to a numerical number that mathematical systems can understand. If the degree of truth is high enough, this rule will trigger, and an action will be taken to shut down the motor. The real challenge to build a robust fuzzy logic system is not the rule itself (since it is basically assembled together by “expert knowledge”). In reality, the rule is not assigned a precise number. We would rather use linguistic terms like “too high.” The difficult part of building a robust fuzzy logic engine is the inference engine, because computers cannot understand a linguistic term such as “too high.” The expert must program the computer to know how to calculate the degree of truth for “too high.” To discern this, an expert must define a “membership function” so the software can calculate the degree of truth. A typical membership function is demonstrated as the following: SPE 163704 5 Fig. 2: Membership function example - degree of truth for “too high” 1 0 150F 300F Temperature In this example, when the temperature is below 150°F, the degree of truth for “too high” is zero. When the temperature is above 300°F, the degree of truth for “too high” becomes 1. The strength of fuzzy logic is apparent when the temperature is between 150°F and 300°F. The degree of truth will control how strong the rule will be applied. In order for the computer to understand the linguistic term “too high,” we need to define: 1. 2. 3. A normal value A low value A high value In the example, the normal value could be 225°F, the low value 150°F, and the high value 300°F. The sensible question at this point is how should the proper low/normal/high values be chosen? Once an ESP has been designed, and its likely operating envelope is known, there is an idea about what the theoretical motor temperature will be. In other words, there is strong reason to believe the real operational motor temperature will not be that far away from the theoretical value (if the assumptions are valid). A reasonable approach for selecting a normal value includes studying the sizing for theoretical motor temperature, then applying a factor to set the low and high values. This factor can be application dependent (for example, a low flow application may have a higher expected motor temperature). Once the fuzzy logic system has been deployed to field production, it is important for the engineer to review the low/normal/high value to overcome any errors in the assumptions or model discrepancies. The strategy is to tune the fuzzy logic system to provide the correct sensitivity level for the ESP application. A high-cost environment such as deep water may have tighter values than a brownfield environment, because of the economic implications of a premature failure. Below are several cases that demonstrate how fuzzy logic can be used for ESP operational troubleshooting. 6 SPE F Fig. 3: Fuzzy logic l case histo ory, tubing lea ak O On 24Jan. 2009 9, the SCADA data showed pump p intake pressure (Pip) inncreasing and fl flow (Qstk) deccreasing. Whenn the fuzzy loogic inference engine processed the combin nation of all SC CADA signals against the tubbing leak rule iin the rule basee, the ddegree of truth gradually rosee from 10% in January J to 51% % in February, and then 61% in March. T This reveals an nother advantag ge of using fuzzzy logic. As Piip keeps increaasing, and flow w keeps decreassing, the composited ddegree of truth for the “if” sid de of the tubing g leak rule conttinues to get biigger. From ann operational pooint of view, thhe relative cchange of the tu ubing leak cou uld be used as a trigger for an ESP engineer to evaluate a ssituation. S SPE 163704 4 7 F Fig. 4. Fuzzy lo ogic case histo ory: broken pu ump shaft O On 27 April, th he intake pressu ure began increeasing rapidly. The motor tem mperature begaan increasing w while the motorr amps and ddischarge presssure began to drop. d In this exaample, one of the t fuzzy logicc rules for the ““pump shaft brroken” was trigggered. L Later, field testts confirmed th he broken shaftt. F For fuzzy logicc to be effectivee, it is extremeely important to o set correct infference valuess for low/normaal/high. If these values aare not set correectly, fuzzy log gic will perform m reasoning ag gainst the wronng “base line”. In other words, if normal isnn’t known, thhen it will be difficult d to deteect what is not normal. U Unfortunately, most of the weells included in n ESP applicatiion have dynam mic behaviors. For example, water cut for a well ccould change, and a therefore, it i will affect pu ump intake pressure. If an infference engine still uses old ppump intake prressure ddata, it will incorrectly triggerr the rule base.. For this reaso on, updating thee low/normal/hhigh values bassed on the currrent ooperating condiition becomes critical to utilizing fuzzy logic in a meaninggful way. Thee engineer needds effective andd userffriendly tools to o assist in keep ping the ESP model m valid. Th his challenge, aand the methodds for overcomiing it, are presented later inn this paper. V Virtual Meterin ng A Another advanttage of an integ grated ESP mo odel is virtual meters. m By com mbining existinng signals from m surface and doownhole ssensors, in conjjunction with a properly tuneed model, vario ous other param meters can be m metered virtually (Denney & Crossley, 22011). This allo ows measurem ments where eith her physical (ee.g. casing size limitations) orr economic connstraints exist. One of thhe primary, an nd most useful, measurementss that can be acchieved is flow w rate. F Flow rate in an ESP applicatio on is typically received by a periodic p well ttest. In many E ESP applicationns, this well tesst is cconducted oncee a month; how wever, it can bee much more in nfrequent depennding on the loocal circumstannces. A real-tim me flow rrate offers the following f advaantages (Denneey & Crossley, 2011): 1. 2. 3. 4. Allow ws for optimizattion of ESP equ uipment perforrmance; particuularly operatinng point-on-thee-pump curve Gives information to o diagnose non-catastrophic equipment e and well issues Verification of meassurements taken n by test separaator Evaluaation of reservo oir response 8 SPE F Flow monitorin ng typically utiilizes commonlly available reaal-time informaation from an E ESP, includingg tubing pressuure (Ptbg), ppump dischargee pressure (Pdp p), pump intakee pressure (Pip p), and controlller frequency. T The model is uused to make a rrepresentation of o the pump beehavior (the pu ump curve – mo odified to actuaal operating coonditions) and tthe reservoir ((productivity in ndex, reservoir pressures, fluiid characteristiics, etc.). Fig. 5 illustrates onne virtual flow m meter: 1. 2. Utiliziing the real tim me measuremen nts from Pdp an nd Pip, the heaad generated froom the pump ccan be calculatted (Pdp – Pip = head) h By plo otting head on the operational pump curve (at ( the current ooperating frequuency), the estimated flow raate of the pump can be achieveed F Fig. 5: Virtuall flow meter A As with the fuzzzy logic engin ne, the accuracy y of the virtual flow meter is directly depenndent on the quuality of the moodel used. F Fig. 6 illustratees this point (Denney & Crosssley, 2011). Th he Neuraflow rreading trackedd the flow readding from a muulti-phase fflow meter with h better than 95 5% accuracy frrom the period d between Septeember 2007 too May 2008 witthout the modeel needing too be recalibrated. In June 200 08, there was a water breakth hrough event w which caused thhe density of thhe fluid to channge, and thherefore, the behavior b of the pump change as well. If therre had been a m method to detecct the change inn water cut (test sseparator, surfaace flow meter,, etc.), a recalib bration would have h been issuued to keep the model matcheed. However, inn this eexample, the reecalibration waas not issued to illustrate the importance i of a valid well moodel for virtuall flow meter reeading. S SPE 163704 4 9 F Fig. 6: Virtuall flow meter ca ase history D Design Improveements T The primary methodology useed for sizing an n ESP tends to be selecting seeveral cases too reflect the dynnamic nature oof the ppump operation n or to reflect unknowns u (for example, the PI P in a newly ddrilled well). Inn most cases, thhese are just ‘bbest gguesses’ at a reepresentation of a static scenaario, and tend to o ignore the dyynamic nature of pump operaations. Some deesign ssoftware have simulators, s whiich allow the ESP E engineer to o simulate certtain dynamic evvents (such as startup, gas sluugging, eetc.), but simulation of all eveents would be impractical. i Du uring actual puump operationss, design softw ware can be flaggged in thhese dynamic situations. A An example off when flagging g is used would d be if during pump p operationn, the formationn of emulsionss caused a shafft to be ooverloaded. Wh hile installing the t subsequentt ESP, or an ES SP in an offset well, the ESP engineer may decide to utilizze a high sstrength shaft in n their design. The formation n of emulsion may m not have bbeen one of thee cases that the engineer tested when m making his orig ginal design. P Production and d System Optim mization T The initial sizin ng of ESP equiipment is based d on the best av vailable data, bbut often the innformation is only partially acccurate. F For this reason,, it is paramoun nt that the “theeoretical modell” is continuouusly updated wiith real data. B By matching thhe real data w with the theorettical model, the system and production p can be improved. T To boost produ uction, a steady y state of produ uction must firsst be reached too have an accuurate match. Once the well haas reached a steady state, the t pump intak ke pressure (Pip p), motor temperature (Tm), aand pump disccharge pressuree (Pdp) from thhe down hhole gauge and d motor amps, tubing t pressuree, and flow ratees from the surrface will be ussed to start the matching proccess. This innformation willl be used with hin the fuzzy lo ogic engine to determine d wherre the pump is operating on tthe pump curvee. The ffirst match thatt is performed will w become th he baseline from m which all datta will be comppared to in the future (on the current E ESP install). C Continuously updating u the mo odel within thee fuzzy logic en ngine will enabble the engineeer to determinee when it is neccessary to m make changes to t the operating g parameters of o the ESP to op ptimize producction and reducce lifting costs ($/bbl/1000’ oof lift). U Using nodal an nalysis, the engineer will be ab ble to ascertain n where in the system the botttlenecks are occcurring, and w with the aaid of a matcheed sizing file, be b able to deterrmine where the bottlenecks aare within the E ESP system. T The changes made could bbe as simple as adjusting chok ke settings or more m complicatted such as whhere multiple isssues are causinng production 10 SPE deficiencies. Some of these issues could be related to scale build-up within the pump, or tubing or emulsion issues causing decreased ESP efficiency. With an accurately matched sizing, the engineer will save time by skipping the trial and error approach, and reach the correct conclusion with a more scientific approach. Additionally, the software can be used to highlight opportunities for increased production. Because the software has knowledge of its’ operating constraints, it can virtually step through new operating frequencies and their resulting production increases until a new operating constraint is reached. At this point, the engineer is presented a classic ‘cost versus benefit’ analysis - do I want to increase my production by X% at the cost of exceeding Y constraint and perhaps cause a premature failure? While this could be done manually, automating the process allows for quicker decision making and more time vetting the opportunity than uncovering that the opportunity exists. Improving the system is also important to decrease lifting costs because of the expense associated with running ESP with electricity. An average system’s electrical cost can be as much as 5% of the original cost of the equipment every month. For this reason, it would be beneficial to evaluate the ESP’s electricity use. The matched sizing file helps engineers determine the correct ESP input voltage from field data. Because the data used to size ESP is only partially accurate, when the ESP is being installed, the field setup reflects the theoretical model and not actual field data. Once the well reaches a steady state, the optimal operation point can be determined. Depending on the outcome of the matched file, changes could be made to the drive setup, transformer taps, or re-rate or de-rate a motor. These tasks ensure the system operates the most efficiently. Also, as the well production follows the decline curve of the well, it will be important to make changes as production changes. For example, two years after the ESP is installed, production could have declined by 10%-meaning it is no longer operating the most efficiently. At that point, the engineer would decide where the new best efficiency point of the system is located and change the setup to match those conditions. Challenges with Model Management To realize the benefits listed above, an accurate ESP model must be maintained. The traditional method for maintaining such a model required ESP engineers to manually run the ESP sizing software, and use a trial-and-error approach to adjust the ESP design until the performance data (such as intake pressure, discharge pressure and flow rate) match the measurements from the SCADA system. This process is tedious and error prone-especially when scaled to the thousands of ESP application models which may need to be updated. A new method that includes soft-computing techniques is being investigated to allow a computerized system to perform the recalibration automatically with limited human intervention. Preliminary testing suggests this method can be used to perform ESP application model tune-ups for multiple ESP/wells. One can think of automatching as “reverse sizing.” In sizing, the user enters the well condition parameters outlined in Table 1 to size an ESP system. In automatching, we try to find optimized data sets of unknowns (or dynamic variables) such as water cut, gas to oil ratio, and pump modifiers to match SCADA measured values such as Pip, Pdp, flow, and motor amps. Automatching presently uses these five values as inputs. The following are values which are currently captured by most SCADA systems: 1. 2. 3. 4. 5. Pip (Intake Pressure) Pdp (Discharge Pressure) Flow Motor Amps Frequency Variables which will be modified to create a match for automatching include static pressure (Pr), productivity index (PI), water cut (WC), and gas oil ratio (GOR). An experienced ESP application engineer may already notice that to match a certain pump intake pressure (Pip) value acquired from a field, one can change Pr (static pressure) or PI (productivity index), or a combination of Pr and PI to match a Pip value. Without more information, the engineer and software are just making a guess as to what has changed. SPE 163704 11 To solve the problem, a customer’s production testing data should be analyzed. Periodically, a well test should be available to verify production data such as water cut and GOR. This information can be used as a constraint to help to narrow the solutions to more closely reflect reality. The algorithms also allow a progressive automatch of an ESP well. Automatching constantly uses the latest SCADA data to recalibrate the ESP model, in an attempt to keep it functional. When automatching finds a good solution, it sends the information to the ESP engineer, as shown in Table 2 below. Table 2: Automatching output APC Inputs Old APC Input New APC Input PI 0.14 0.17 Pr 2255 2255 WC 71.6 70.63 GOR 2431.47 1029.71 For this particular well, automatching suggests changing productivity index (PI) from 0.14 to 0.17, and GOR from 2431 to 1029 to match to the most recent SCADA data set. From this automatching result, the ESP engineer can validate whether the automatching system has made valid assumptions. If the automatch is correct, the ESP engineer can accept the changes with one click, and the model will be valid until the SCADA data no longer matches the model. Additionally, the ESP engineer might go to the field and get more data on one of the unknowns (for example, GOR). The ESP engineer could then constrain GOR to a known value and the automatching system will only pick from the other variables to generate a match. The automatching feature gives the ESP engineer a suggested match and can constrain the unknowns appropriately to reduce nonsensical matches from being displayed. This will improve the quality of matching and reduce the amount of time that the ESP engineer spends maintaining models. Another benefit is allowing the ESP engineer insight into how changing well parameters will impact the equipment. In the original sizing, the engineer may not have seen any equipment issues (based on his assumptions about water cut or GOR, for example). As the well is produced, however, and water cut and GOR changes, the load on the shaft or the internal motor temperature may change to a point that an alarm could be raised from the ESP sizing software. Integrated models have proven beneficial in many ESP applications. There are, however, some ESP applications that are not good candidates for utilizing an integrated model. As fields mature, additional measures will be taken to increase production by implementing enhanced oil recovery/ incremental oil recovery (EOR/IOR). By the very nature of these projects, most wells will experience very dynamic conditions downhole that will create additional challenges to accurately modeling well performance. One EOR technique that has been extremely challenging to model is water-alternating-gas (WAG) floods. A typical WAG flood consists of alternating water and carbon dioxide (CO2). To accurately model an ESP in these conditions, the engineer would have to match the actual field data every week because of the changing bottomhole pressures and fluid density. There are two issues with having to make changes every week to accurately match the well. One is the time required to rematch every well, and the other is making changes to the model every week will mask the downhole issues, making it impossible to tell if operational changes need to be made. To accommodate these dynamic wells, additional research will be needed to 12 SPE determine if there is a way to accurately model these wells, and will still provide practical information to make operational changes to the ESP system. Furthermore, there are some ‘real world’ challenges which make implementation at scale difficult. Some of these challenges include sensor reliability, availability of peripheral information (well test data, flow meter, field work, etc.), competitor information, and system noise. All of these factors contribute to the challenge of keeping a model functional. There needs to be further research into data processing and validation techniques to ensure that the automatching systems are provided untainted data. Conclusion There are many benefits of having a functional ESP model. The challenges associated with keeping a valuable model are substantial, and can be tedious. After all, a model is only a representation of the situation – we cannot expect the models to perfectly capture the true complexity of the reservoirs and mechanical equipment at play. While there is significant operational value in keeping a model valid, much effort is spent in the model management process. Further evaluation and research needs to be performed on different methods, including case-based reasoning and other artificial intelligence methods, which may allow for further scalability and reduce the overhead associated with model maintenance. SPE 163704 13 Works Cited Baker Hughes Centrilift. (2008). Submersible Pump Handbook Eight Edition. Claremore: Baker Hughes. Denney, T., & Crossley, A. (2011). Model-Derived Flow Monitoring in an Electrical Submersible Pump (ESP) Application. ESP Workshop. The Woodlands: SPE. Thornhill, D., & Zhu, D. (n.d.). Fuzzy Analysis of ESP Performance. 2009 Annual Technical Conference and Exhibition . SPE.