GAS LIFT OPTIMIZATION USING ARTIFICIAL NEURAL NETWORK

Article pubs.acs.org/EF

Gas Lift Optimization Using Artificial Neural Network and Integrated Production Modeling Eissa Mohamed El-M. Shokir,*,† Mazen M. B. Hamed,‡ Azza El-S. B. Ibrahim,§ and Ismail Mahgoub∥ †

Mining, Petroleum, and Metallurgical Department, Faculty of Engineering, Cairo University, Giza 12613, Egypt Gulf of Suez Petroleum Company (GUPCO), New Maadi, Cairo 11311, Egypt § Electronics Research Institute (ERI) El Dokki, Giza, 12622, Egypt ∥ Petroleum Engineering Department, Future University in Egypt, New Cairo, Cairo 11835, Egypt ‡

ABSTRACT: The well flowing bottom-hole pressure and fluid rates must be known for different applications in the oil and gas industry. The accurate values of these parameters are necessary for different calculations such as gas lift optimization, well monitoring, reservoir performance evaluation, and field development plans. Estimating the values of these parameters without any well intervention is in great demand to minimize the numbers of intervention jobs, operation risk, time, and money. Many correlations are available in the literature to predict these parameters; however, these correlations required knowledge of different variables that are not usually available in accurate values. Searching for a more robust tool for predicting accurate values for these parameters is in demand. Therefore, an artificial neural network (ANN) model was developed from an extracted data set from PROSPER1 software, production logging tool (PLT), and test separator data. First the ANN model was trained and tested by synthetic data (extracted from PROSPER1 software). Then, the ANN model was tested by a group of test points collected from the PLT reports. The developed ANN model yielded an accurate prediction of the well flowing bottom-hole pressure and well fluid rate. The values of these parameters of each well are used to build an integrated production model (IPM) using GAP2 software to perform different gas lift optimization scenarios.

1. INTRODUCTION In the oil and gas industry, millions of dollars are spent on improving measurement instruments such as down-hole gauges and wireless surface sensors. Additionally, there is a large amount of measured data and interpretation reports regarding well intervention jobs available in the wells history files, but unfortunately they are scattered in an unstructured format. This makes it hard for data to be analyzed and studied by engineers to determine a certain relation. In some cases, engineers could not know the production rates of each well during production because a group of wells are connected to one separator. Therefore, the total flow rates of these wells are measured, not the rate of fluids of each well. In the offshore environment, performing the jobs and taking the measurements is complicated and expensive. In the case of gas lift wells, the situation will be worse. For reservoir and production engineers, it is important to know the fluid rate and the bottom hole flowing pressure of each well in order to optimize the production, allocate the gas lift injected volumes, and monitor the reservoir and well performance. Therefore, this work presents an approach for how to utilize this large amount of structured and unstructured data to find the variables related to the bottom-hole pressure and well fluid rate. Then, these variables are used to build different PROSPER1 models for different wells from the Egyptian offshore field (group of wells connected to one separator) for generating a huge a amount of synthetic data. This synthetic data was used to develop an ANN model to predict the bottomhole flowing pressure and the fluid rate of each well in that field. After that, these predicted bottom-hole flowing pressures and fluid rates were used to build an updated PROSPER1 model for © XXXX American Chemical Society

each well to generate an integrated production model using GAP2 software for gas lift optimization in the selected field.

2. OPTIMIZATION METHODS AND TECHNIQUES Many variables are involved in a successful gas lift operation. Gas lift optimization means specifying these variables in such a way that the production and the operation’s net present value (NPV) are maximized. The most important parameter in a gas lift operation is the amount of gas injected into the well. The increase of injected gas will first increase the gas−liquid ratio (GLR) which will decrease the bottom hole flowing pressure. Therefore, larger oil production rates are likely to be achieved. On the other hand, there is a limit value of GLR, called “limit GLR”. Above this limit, the decrease in the hydrostatic pressure will be offset by the increase in the friction pressure. Therefore, finding the optimum amount of gas injected into the well is important because injecting an extra amount of gas increases slippage between liquid and gas and consequently reduces production.3 Poettmann and Carpenter4 introduced the “Gas-Lift Performance Curve (GLPC)″ which has been used extensively in the petroleum industry. GLPC is a graphical plot locating the intersection points of the IPR (Inflow Performance Relationship) curve with VLP (Vertical Lift Performance) curves for various values of gas injection rate and various GLR (Gas Liquid Ratio) in given tubing. Simply, the GLPC is a plot of the gas injection rate versus the production rate. The maximum production rate corresponds to the gas injection rate for the limit GLR. Practically, the degree to which production can be increased through increased gas injection depends on many variables, but generally the higher the water cut of the well and the lower the associated gas in the fluid, the more gas lift will be needed to increase Received: June 20, 2017 Revised: July 31, 2017 Published: August 4, 2017 A

DOI: 10.1021/acs.energyfuels.7b01690 Energy Fuels XXXX, XXX, XXX−XXX

Article

Energy & Fuels Table 1. Minimum and Maximum Values of the Input and Output Variables for of the Developed ANN Variable

GOR SCF/BOPD

I.G MMSCF/BOPD

D/S psi

Pr psi

WC %

WHP psi

WHT oF

Qf BOPD

Pwf psi

minimum maximum

164 1517

0.3 5

700 1450

1600 3500

0 98

70 300

65 228

35 10101

1330 2960

liquid production. Therefore, for any gas lift optimization program the well needs to be multirate-tested. The petroleum engineer can either optimize for the maximum production rate if there are no constraints or select a gas injection rate that gives the maximum economic return. Nodal analysis could be used to generate the gas-lift performance curve of a single-well based on actual pressure and temperature surveys along with a suitable multiphase flow correlation.5,6 Although this model has been established well, the single-well configuration, considered in isolation of other wells, is not a field gas-lift optimization solution. Vazquez and Hernandez7 examined a single-well case to determine the injection depth, pressure, and the amount of gas injection. A more accurate well model, based on mass, energy, and momentum balance, was proposed and the results were reported as being more accurate and therefore better suited for field-wide simulation studies than the standard nodal approach. However, no results were reported for a field-wide application. For better accuracy, it is recommended to use compositional models over simple black oil models.8 The IPM tool is a powerful simulator and useful mean for simulating actual production systems and assessing their responses to different production scenarios, challenges, and the impact of various solutions on production systems before field implementation. It also improves the understanding of the overall production system performance and provides an analytical tool to assist in the optimization of the entire production system.9 PROSPER1 software is one of the most important packages of IPM that can help petroleum producers maximize their production by providing the means of critically analyzing the performance of each producing well. It assists production and/or reservoir engineers to make a model for each component of the producing well system separately, which contributes to overall performance. Once the system model has been tuned to real data, by performance matching for each model subsystem, PROSPER1 can be used confidently to model the well in different scenarios.

back-propagated through the system to adjust the weights, which control the network. Once a neural network is trained to a satisfactory level, it may be then used as an analytical tool on other hidden data. To do this, the user no longer specifies any training runs and instead allows the network to only work in forward propagation mode. Inputs are presented to the network through input layer and processed by the middle layers as the training is taking place; however, at this point, the output is retained, and no back propagation occurs.10−12 Nowadays, ANN models are the subject of study in areas as diverse as medicine, engineering, and economics, to tackle problems that cannot be easily solved by other more established approaches.

4. BUILDING ANN MODEL 4.1. Data Description and Quality Control. In the selected field, there are scattered down-hole data from different

Figure 1. Neural network architecture.

3. ARTIFICIAL NEURAL NETWORKS (ANN) ANN is technology initially growno from the full understanding of some ideas and aspects about how biological systems work, especially the human brain. Neural systems are typically organized in layers. Layers are made up of a number of interconnected nodes (artificial neurons), which contain activation functions. Patterns are presented to the network via the input layer, which communicates to one or more hidden layers where the actual processing is done through a system of fully or partially weighted connections. The hidden layers then linked to the output layer. Neural network contains some sort of learning rule that modifies the weights of the connections according to the input patterns.10,11 Neural networks have the capacity to learn, memorize, and create relationships between the input and output variables. There are many different types of neural network. The most widely used network is known as the Back Propagation Neural Network (which is in use in this paper). This type of neural network is excellent at prediction and classification tasks.10−12 Neural networks require the use of training patterns, and involve a forward propagation step followed by a backward propagation step. The forward propagation step sends an input signal through the neurons at each layer resulting in the calculation of an output value. This output is then compared with the desired output and the error is computed, which is

Figure 2. Performance of the developed ANN Model.

wells collected from old test separator and automation systems which represent the well head pressure and temperature, gas lift injected volume, and gas injection pressure. As mentioned before, the scope of work is to use these scattered data to B

DOI: 10.1021/acs.energyfuels.7b01690 Energy Fuels XXXX, XXX, XXX−XXX

Article

Energy & Fuels Table 2. PLT Points Used as a Secondary Test for the Developed ANN Model Point

WHP psi

WHT oF

I.G MMSCF/BOPD

D/S psi

WC %

Pr Psi

GOR SCF/BOPD

Qf BOPD

Pwf Psi

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

168 140 168 163 141 256 85 104 107 105 115 145 128 93 100 110 122 126 150 110 115 115 169 182 129 273 270 151 90 112 140 180 180 170 160 130 125 120 135 300

160 1185 152 130 131 150 130 154 156 139 132 123 128 126 150 130 180 140 125 152 165 162 162 179 121 140 175 166 143 164 100 167 140 145 160 120 180 187 180 140

3 3 3 2.2 2.5 3 1.5 2.5 3 2 1.2 1.5 1.5 0.9 3.2 3.5 4 3.5 4 1.2 2.3 3.5 2.5 2.2 2.8 2 4 3 2.8 4.5 2.6 4.4 3.5 3.5 1 1.5 5.5 4.5 5 2.1

1275 950 1275 924 1150 1117 900 1300 1135 1150 725 1074 930 1385 1170 1070 1226 1144 1450 920 1350 1310 1226 1294 1225 1125 1175 1310 998 1209 1250 1290 1210 1210 1280 773 1288 1300 1150 1250

65 89 79 28 55 68 81 10 61 63 6 17 6 6 86 85 90 77 84 29 90 86 73 75 70 45 92 86 62 82 53 90 78 62 23 4 85 95 91 74

2600 2500 2540 1540 1945 3000 2077 1910 2090 2700 1550 1620 1282 1720 2800 2780 2810 2600 2800 2310 1920 3000 2820 2600 2740 3150 3150 2730 2370 2600 2730 2800 2300 2700 2700 1640 2700 2575 2790 3100

359 279 279 474 359 819 359 359 279 359 359 359 359 359 279 279 279 279 279 359 359 279 279 279 279 819 819 279 359 279 164 279 164 164 164 164 279 279 279 819

1405 5200 2430 800 715 2460 855 1940 3350 1110 1833 1157 1500 1500 2343 1716 4000 2178 1378 1916 1875 2756 3231 4750 1265 2473 3430 3170 1086 2918 614 3262 1264 1610 2220 1411 4155 4415 4550 1690

2275 1990 2275 1400 1770 2650 2030 1771 1930 2230 1543 1216 1200 1720 2725 2660 2210 2575 2620 1720 1920 2380 2290 2410 2500 2520 2840 2344 2030 2350 2300 2695 2060 2100 2276 1484 2080 2415 2250 2570

predict the bottom flowing pressure and fluid rate. The quality of this approach was checked by getting the value of each variable from more than one source as a double check and comparing it with the offset well values on the same date. The reservoir characteristics are considered in collecting the data and the quality check process. For example, if reservoir pressure performance is strong water drive, and all the wells are connected, one can trust the reservoir pressure reading easily and take from offset wells. However, if the reservoir is depletion drive, heterogeneous, and compartmentalized, the reading will need more investigation to be confirmed. Some data was excluded based on observation of any well problems like gas lift valve leaks, communication between tubing and annulus, and tubing obstruction. 4.2. Neural Network Architecture and Training. First, the ANN model was built for predicting the bottom flowing pressure and fluid rate from synthetic data (extracted from PROSPER models) representing different wells from different reservoirs in the selected field. These well PROSPER models

were matched by using the available PLT data. Then, these matched PROSPER models were used to determine different sensitivities on different parameters.This sensitivity was done on well head pressure (WHP), injection gas volume (IG), injection gas pressure (D/S), static reservoir pressure (Pr), and water cut parameters (WC). For each sensitivity case the bottom hole flowing pressure Pwf, total well fluid rate, well head temperature (WHT), and reservoir gas oil ratio were also collected with the above parameters forming one completed sensitivity case. About 32259 synthetic sample points were generated from the sensitivity cases of the match PROSPER models. The minimum and maximum values of the input and output variables for the developed ANN are listed in Table 1. This synthetic data set was used in developing the artificial neural network (ANN) model. This data set was randomly divided into 70% for training, 15% for validation, and 15% for primary test. The resulting architecture of the neural network to predict the bottom flowing pressure and fluid rate contained seven input variables, one hidden layer with 6 neurons (as C

DOI: 10.1021/acs.energyfuels.7b01690 Energy Fuels XXXX, XXX, XXX−XXX

Article

Energy & Fuels

Figure 3. Actual vs predicted ANN fluid rate.

Figure 4. Actual vs predicted ANN bottom hole flowing pressure.

temperature, gas injection volume, and gas injection pressure. The developed ANN model was used to predict the missing data (bottom hole flowing pressure and well fluid rate) that is necessary to do gas lift optimization and reallocated the available volume of gas injection between the current 40 wells. The predicted bottom hole flowing pressure and total fluid rate were used to build updated PROSPER model for each well. After that the GAP2 software was used to build a surface network model for the field under study which consists of the wells, wellheads, flow lines from the wellhead to the production manifold, and the production line from the production manifold to the low pressure separator. This surface network model was linked to the generated IPR and VLP files of the PROSPER models of all the wells in the selected field forming an integrated production model (IPM). By this IPM, illustrated in Figure 5, the gas lift optimization process was ready to be run. The main challenge while running the gas lift optimization is the limited volume of the high pressure gas available in the field. The gas volume is around 104 MMSCF/D with pressure

shown in Figure 1). It was found that increasing numbers of hidden layer neurons greater than six are not efficient, because this increase led to memorization and overfitting of the ANN model which resulted in poor prediction quality. Figure 2 shows the performance of the developed ANN model from the synthetic data set. After performing the first test with the developed ANN model, extra test points which were not used in developing the model were used as a secondary test for final approval of the model efficiency in predicting the bottom flowing pressure and well fluid rate. These secondary test data are listed in Table 2. Figures 3 and 4 display a good match between the predicted and measured of bottom hole flowing pressure and total well fluid rate for the second test.

5. ESTIMATION OF OPTIMUM GAS-INJECTION RATE Currently the available data per each well in the field under study are only the water cut samples, reservoir pressure readings, reservoir gas oil ratio, well head pressure, well head D

DOI: 10.1021/acs.energyfuels.7b01690 Energy Fuels XXXX, XXX, XXX−XXX

Article

Energy & Fuels

Figure 5. Schematic of the GAP model.

equal to 1450 psi. The total volume of gas was reallocated between all the wells in the field according to different optimization scenarios: (a) optimize only the high oil production wells; (b) optimize all the wells in the field. The optimization processes aims to increase oil production from the current wells without any changes of gas lift design, well recompletion, or reservoir recovery method and/or stimulation techniques.

6. RESULTS OF GAP SIMULATOR RUNS 6.1. Case Study No. 1. The optimization run was done on wells which produce about 500 BOPD. These wells are

Figure 7. Saved gas injection volume in each case.

currently shut-in wells with oil gain around 50 BOPD. The total oil production increment after the optimization process is 1070 BOPD) in addition to 50 BOPD after using the 1.7 MMSCFD saved gas in the other well. 6.2. Case Study No. 2. In this case the optimization process is done on all the wells in the field. The purpose of this is to monitor the change on all the wells in the field and know if it is possible for the low oil production wells to produce a higher amount of oil or not. It was found that applying the optimization process on all the wells yielded significant oil production. Not only did two wells result in a significant increase in oil production, as in case 1, but 20 oil producers had an increase in oil production in this case. The overall results showed that there are around 2.2 MMSCFD have been saved from the optimization process and there is an oil gain of about 2260 BOPD. The 2.2 MMSCF/D saved gas can be easily directed to open one of the current shut-in wells with oil gain around 80 BOPD. The optimization process for all wells in the field has a better impact on field daily oil production and also on saving some amount of gas injection volume than optimizing the higher oil producer wells (as shown in Figures 6 and 7).

Figure 6. Oil gain in each case.

considered the top field producers and it is not easy to shut in or even try to change their operating conditions due to company production policy. It is found that only two wells are highly affected by the optimization process, which resulted in more oil production, while the others have no significant change in oil production. The overall results showed that there are around 1.7 MMSCF/D that have been saved from the optimization process, and there is an oil gain of about 1070 BOPD. The 1.7 MMSCFD saved gas can be easily directed to open one of the E

DOI: 10.1021/acs.energyfuels.7b01690 Energy Fuels XXXX, XXX, XXX−XXX

Article

Energy & Fuels

(11) Shokir, E.M.EL-M.; Alsughayer, A. A.; AL-Ateeq, A. Permeability Estimation from Well Log Responses. JCPT 2006, 45 (11), 1 DOI: 10.2118/06-11-05. (12) Shokir, E. M.; El-M; Al-Homadhi, E. S.; Al-Mahdy, O.; ElMidany, A. A. Development of artificial neural network models for supercritical fluid solvency in presence of co-solvents. Korean J. Chem. Eng. 2014, 31 (8), 1496−1504.

7. CONCLUSIONS To obtain the optimum gas injection and oil production rate, some wells had been modeled properly on PROSPER using trusted PLT data. All available well test data in the time of these PLT data had been considered for quality checking. These modeled wells on PROSPER are used to export a huge synthetic data set. ANN model was developed using the generated synthetic data. It consists of seven inputs variables (WHP, WHT, GOR, WC, IG, D/S, and GOR, one hidden layer with 6 neurons, and two output variables (Pwf and Qf). The developed ANN model is tested by separate PLT values which did not used in the training process and it yielded a high accurate results. Then the developed ANN model was used to predict Pwf and Qf values of each well in the field under study. An integrated production model (IPM) using PROSPER and GAP was built for the whole field. Then, by the predicted values of Pwf and Qf different gas lift optimization scenarios was done to all the wells in the field under study. The gas lift optimization processes resulted in increasing oil production by 2260 BOPD and saving 2.2 MMSCF/day gas injection volume in the field under study.

■

AUTHOR INFORMATION

Corresponding Author

*E-mail: [email protected]. ORCID

Eissa Mohamed El-M. Shokir: 0000-0002-7623-9646 Notes

The authors declare no competing financial interest.

■

REFERENCES

(1) Petroleum Experts. User Manual, PROSPER (PROduction and Systems PERformance analysis), version 11; 2009. (2) Petroleum Experts. User manual, GAP (Groups, Algorithms and Programming), version 11; 2009. (3) Ebrahimi, M. Gas Lift Optimization in One of Iranian South Western Oil Fields. In Proceedings of Society of Petroleum Engineers (SPE) Trinidad and Tobago Energy Resources Conference; June 27−30, 2010; paper SPE 133434. (4) Poettman, F. H.; Carpenter, P. G. Multiphase Flow of Gas, Oil and Water through Vertical Flow Strings with Application to the Design of Gas-Lift Installations. Drill. and Prod. Prac., API 1952, 257− 317. (5) Bahadori, S.; Ayatollahi, S.; Moshfeghian, M. Simulation and Optimization of Continuous Gas Lift System in Aghajari Oil Field. In Proceedings of Society of Petroleum Engineers (SPE) Asia Pacific Improved Oil Recovery Conference; Malaysia, Oct. 2001; SPE Paper 72169. (6) Denney, D. Simulation and Optimization of Continuous Gas Lift. JPT, J. Pet. Technol. 2002, 54, 60. (7) Vazquez-Roman, R.; Hernandez, P. P. A New Approach for Continuous Gas Lift Simulation and Optimization. In Proceedings of Society of Petroleum Engineers (SPE) Annual Technical Conference and Exhibition; Texas, United States; Oct. 2005; paper SPE 95949. (8) Bahadori, A.; Zeidani, K. Compositional Model Improves GasLift Optimization for Iranian Oil Field. Oil Gas J. 2006, 5, 42−47. (9) Campos, S. R. V.; Teixeira, A. F.; Vieira, L. F.; Sunjerga, S. Urucu Field Integrated Production Modeling in Occidental of Sultanate of Oman; In Proceedings of Society of Petroleum Engineers (SPE) Intelligent Energy Conference and Exhibition; Utrecht, The Netherlands; March 23−25, 2010; paper SPE 128742. (10) Shokir, E. M. El-M. Neural Network Determines Shaly-Sand Hydrocarbon Saturation. Oil Gas J. 2001, 99 (17), 37−41. F

DOI: 10.1021/acs.energyfuels.7b01690 Energy Fuels XXXX, XXX, XXX−XXX