A Neural Network Predictive Model of Pipeline Internal Corrosion Profile

Giulia De Masi, Roberta Vichi, Manuela Gentile, Roberto Bruschi

ADVEN Dept. Saipem Spa

Fano (PU), Italy Giulia.demasi@saipem.com

Giovanna Gabetta SIAV Dept.

ENI Spa Milan, Italy

Abstract—Internal corrosion is a crucial issue for the safe operation of oil&gas pipelines. This is a phenomenon due to interaction of different mechanisms. Water and electrochemistry, protective scales, flow velocity, steel composition and localized bacteria attacks are relevant. Despite the large number of models proposed in literature, the corrosion process is very complex and rarely reproduced by existing models. For this reason, an artificial neural network (ANN) based model is investigated, with the aim to correctly predict the presence of metal loss and corrosion rate along a pipeline. In this paper, a case study is considered, based on real field data. The model integrates the geometrical profile of a real pipeline, flow simulations and the most important deterministic corrosion models. It is shown that the ANN model outperforms the deterministic ones.

Keywords-internal corrosion prediction, oil&gas pipeline, neural network, partial derivative sensitivity method

I. INTRODUCTION

Multiphase transport will have a major impact on offshore development during the next decade. In the past, emphasis was placed on preprocessing the multiphase well stream through separation on platforms or even subsea, close to the wells. Drastic reduction in both investment and operating costs can be achieved when unprocessed, multiphase well streams can be transported over longer distances in carbon steel pipelines from subsea wells to main platforms, existing installations on neighboring fields or onshore processing facilities.

Pipeline cost is a considerable part of the investment in subsea projects. For long-distance and large-diameter pipelines, cost can become prohibitively high if the corrosivity of the fluid necessitates the use of corrosion-resistant alloys instead of carbon steel. Better understanding and control of the corrosion of carbon steel can help to increase applicability and therefore have a large economic impact[1].

In this paper we focus on prediction of generalized and localized corrosion generating metal loss. A different mechanism of corrosion is instead stress corrosion cracking (i.e. H2S stress corrosion cracking) that should be addressed during design stage, through a right selection of base material and welds fabrication process.

Internal corrosion is one of the main causes of deterioration of pipelines, particularly in presence of water.

The water content inside the pipeline is partially due to formation water, partially to upsets, partially to sea water entrance as a consequence of failures. Corrosion of carbon steel may be influenced by many factors: CO2 (sweet corrosion), H2S (sour corrosion), water chemistry, flow velocity, oil or water wetting and composition and surface condition of the steel[2]. A small change in one of these parameters can change the corrosion rate considerably, due to changes in the properties of the thin layer of corrosion products that accumulates on the steel surface. For instance, a pipeline not affected by corrosion for many years, can be subjected to high corrosion (several mm per year) after changing flow characteristics. The corrosion rate can be reduced substantially under conditions where iron carbonate (FeCO3) can precipitate on the steel surface and form a dense and protective corrosion product film. When in some localized regions this deposit is no more adherent to steel surface, localized corrosion can appear. This may happen due to oxide layers formation with subsequent formation of underlying crevices where fast corrosion (even 100 times higher than uniform corrosion) can occur (crevice corrosion), or due to turbulent/slug flow driven erosion (mesa corrosion). Another source of localized corrosion is due to action of bacteria (MIC, microbiologically induced corrosion) that occurs where stagnant water is present.

In order to control corrosion in pipelines, it is important to understand the underlying degradation mechanisms and to predict whether corrosion will be initiated, which sections of pipeline with have higher risk of corrosion, and how it can be prevented. Several models to predict CO2 corrosion can be found in literature[3][4][5][6]. These models are mainly based on laboratory data and, in some cases, are validated with field data. Moreover, they can be classified as mechanistic models, semi-empirical models and empirical models[2]. They may give markedly different corrosion rate predictions for the same field case, and which models are most successful in their prediction vary considerably from case to case.

In this paper, a nondeterministic artificial intelligence model is proposed with the aim to increase the accuracy of prediction of occurrence of corrosion and of corrosion rate. This model is based on neural network technique. This kind of models outperforms deterministic models when they have

2014 First International Conference on Systems Informatics, Modelling and Simulation

978-0-7695-5198-2/14 $31.00 © 2014 IEEE

DOI 10.1109/SIMS.2014.14

18

2014 First International Conference on Systems Informatics, Modelling and Simulation

978-0-7695-5198-2/14 $31.00 © 2014 IEEE

DOI 10.1109/SIMS.2014.14

18

2014 First International Conference on Systems Informatics, Modelling and Simulation

978-0-7695-5198-2/14 $31.00 © 2014 IEEE

DOI 10.1109/SIMS.2014.14

18

2014 First International Conference on Systems Informatics, Modelling and Simulation

978-0-7695-5198-2/14 $31.00 © 2014 IEEE

DOI 10.1109/SIMS.2014.14

18

2014 First International Conference on Systems Informatics, Modelling and Simulation

978-0-7695-5198-2/14 $31.00 © 2014 IEEE

DOI 10.1109/SIMS.2014.14

18

to represent very complex highly non-linear phenomena. The application of artificial neural networks (ANN) has been already proposed in literature to predict the average corrosion of pipelines[7][8][9]. In the present study instead, the model is focused on predicting the corrosion profile along the pipeline, to identify the pipeline sections more exposed to corrosion risk.

The neural network model here proposed integrates geometrical characteristics of a pipeline (an application case is considered), corrosion deterministic models and simulations of multiphase flow velocity and transport, as schematized in Figure 1.

This tool represents an integrated process of corrosion analysis, useful both for pipeline design and for integrity management.

II. METHODOLOGY

A. Geometrical characterization The pipeline has been characterized by its geometrical

features: elevation, inclination and concavity. Inclination is demonstrated to play an important role in corrosion process [2], because above certain critical angles water holdup and therefore risk of corrosion increase. Concavity is expected to be important, determining water accumulation.

Inclination I is related to pipeline elevation e in the point �� as:

����� � ��� �������������� � (1)

Whereas concavity C is defined as:

����� � ������������������������ (2)

In Figure 2 elevation, inclination and concavity are shown.

Figure 2 Elevation (black line), inclination (red line, top panel) and concavity (red line, bottom panel)

B. Multiphase flow parameters

Multiphase flow modelling is based on OLGA software[10]. This program provides information on temperature profile along the pipeline, pressure profile, velocity profiles of each phase, phase hold-ups and flow regimes, given boundary pressure, temperature values and flow composition.

Water plays a crucial role for corrosion, enhancing corrosion rate depending on its hold-up and velocity, gas flow rate, pressure and temperature and pipeline inclination[2]. In our specific case, water can be considered a phase separated from gas, at the bottom of pipe.

The multiphase flow simulator can help to identify locations where variation in flow regime, flow velocity and water accumulation may increase the risk of corrosion damage[12]. Fluid regime is described by a discrete number as follows:

1: stratified flow 2: annular flow 3: slug flow 4: bubble flow As evident from Figure 3, in the present case the flow

regime is usually stratified or slug. Figure 4 and Figure 5 report gas velocity and water velocity along the pipeline, as provided by OLGA simulator.

Figure 3 Fluid regime and pipeline elevation

Figure 1 Scheme of artificial intelligence model

�5

�4

�3

�2

�1

0

1

2

3

4

5

�40

�30

�20

�10

0

10

20

30

40

0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000

incl

inat

ion(

°)

elev

atio

n(m

)

KP(Km)

�0.05

�0.03

�0.01

0.01

0.03

0.05

�40

�30

�20

�10

0

10

20

30

40

0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000

conc

avity

elev

atio

n(m

)

( )

0

1

2

3

4

�40

�30

�20

�10

0

10

20

30

0 5000 10000 15000 20000

ID�fl

uid�

regi

me

elev

atio

n(m

)

KP(m)

1919191919

Figure 4 Gas velocity and pipeline elevation

Figure 5 Water velocity and pipeline elevation

C. Deterministic models Two deterministic models are integrated in the artificial

intelligence model here proposed. The first one is the de Waard model [4], which correlates corrosion rate �� to temperature t and CO2 partial pressure (pCO2), by the following relationship: ������� � ��� � ! "

�!#�$ % &�'( ) ���*�+�,�� (3)

The second model is proposed by NORSOK[6]: CR is an empirical function of temperature t, CO2, pH, wall shear stress. For temperature between 20°C and 120°C:

�-$ � .$ ) /01�"�2� ) 3 4�

"� 52�"�"#�5)678*�901�� ) /�+:�$ (4) For t=15°C

�-$ � .$ ) /01�"�#2 ) 3 4�

"� 52�"�"#�5)678*�901�� ) /�+:�$ (5) At temperature 5°C

�-$ � .$ ) /01�"�#2 ) /�+:�$ (6)

D. Artificial neural network In the present study, a fitting neural network is used[13].

Fitting networks (FNN) are feedforward neural networks used to fit an input-output relationship [14], as shown in Figure 6.

Figure 6 Fitting neural network (FNN) block diagram

The FNN integrates all the above quantities as input values. Therefore, input variables are of three types:

� Geometrical pipeline characteristics (elevation, inclination and concavity)

� Fluid dynamic multiphase variables (flow regime, pressure, gas flow, total flow, liquid velocity, gas velocity)

� Deterministic models (de Waard and NORSOK)

Each network has only one output. Three alterative output variables are considered:

� Corrosion rate (CR) � Metal loss � Area of defects

The network structure is reported in Figure 7.

Figure 7 FNN architecture with all inputs and alternative outputs (intermediate arrows are not indicated)

Several training algorithms were tested; finally the Levenberg-Marquardt back propagation algorithm was selected as the one producing best prediction[15]. Two (or more) layer fitting networks can fit any finite input-output nonlinear relationship arbitrarily well, given enough hidden neurons: in the present case 20 hidden neurons are demonstrated to obtain the best network performance.

0

1

2

3

�40

�30

�20

�10

0

10

20

30

0 5000 10000 15000 20000

gas�v

eloc

ity�(m

/s)

elev

atio

n(m

)

KP(m)

�0.5

�0.3

�0.1

0.1

0.3

0.5

�40

�30

�20

�10

0

10

20

30

0 5000 10000 15000 20000

Liqu

id�v

eloc

ity�(m

/s)

elev

atio

n(m

)

KP(m)

f1

IW

input

b1

+ f2

HW

b2

+

Layer 1 Layer 2

output

Bias

Output layer{ok}

Hidden layer{hj}

Input layer {xi}Elevation

Bias

Inclination

Concavity

CR/metal loss/defect area

Flow regime

Hold-up

Pressure

Gas flow

Total flow

Liquid velocity

Gas velocity

De Waard model

NORSOK model

2020202020

E. Sensitivity analysis Neural network is a ‘black box’ type model and it is not

evident the participation of each of the input variables. In this study, a simple method based on the use of the partial derivatives of the network response with respect to each input is used [16]. The link between inputs modification, xi,and outputs variation, ok, is the Jacobian matrix dok/dxi. It represents the sensitivity of the network outputs according to small input perturbations. For a network with I inputs, one hidden layer with J nodes, and one output (K=1)

�; � /�<= >;?@?A B (7) @? � / <= >A?�AA B (8)

Therefore, the gradient vector of ok with respect to xi is given byCDEFC�GF

� = CDEFC�HF

)?C�HFC�GF

� = >;? ) /�I? ) >A? ) / I (9)

where f1 is a logistic sigmoid function, f2 is a linear function, in the present case. The sensitivity of the FNN output with respect to input xi is given by the sum over the pobservations of the square of Eq.9:

JJKA � = LCDEF

C�GFM

�N (10)

This final SSDi is normalized as: JJKA � 33OG

= 33OGG (11)

The derivative can be efficiently computed as minor extension to the backpropagation algorithm used for training.

III. APPLICATION

In the present study an application to a pipeline 20 km long in Mediterranean Sea is investigated. It consists of around 1700 bars, each 12m long. It has been built around 40 years ago. This is an old pipeline, with a complex history. Formation water was firstly injected in the pipe in the first 20 years, then separated at the inlet with an efficiency of 95%. The temperature and pressure profile and flow velocity changed during the lifetime of the pipeline. For instance, operating pressure decreased to about half from 2004 to 2012. Moreover, several reparations have been performed (particularly near the end part of the line).

The main mechanisms of corrosion are CO2 and MIC. Moreover, bottom of line (BoL) corrosion is much more probable than top of line (ToL) corrosion (94% of defects are between 04:00 and 08:00 angles).

A. Corrosion measurements In order to reduce failure incidents caused by internal

corrosion, pigging internal line inspections (ILI) are performed to monitor corrosion and inspect critical parts of pipelines. In particular, two ILI have been performed for the pipeline of interest, one during 2005 and one during 2012

(6.2 years distant each other). The output of inspections consists of: defect position (both distance from inlet and orientation in pipeline circumference), depth, width and length. In Figure 8 the defect distribution along the pipeline is reported: as evident, 76% of defects are located in the last 8km of the pipeline length.

Figure 8 Defect distribution along the pipeline

A comparison between the two ILI has been done, selecting the most relevant defects, with a depth larger than 35% of pipe thickness. For these defects the corrosion rate (CR) has been calculated as

�- � P�QR��P�QQS2��TUVWX (12)

The paper considers both sets of data. Moreover, for each bar, the total number of defects is calculated as well as average corrosion rate. To each bar are associated: a relative area of defects, defined as the sum of defect areas divided by bar area and a value of metal loss, defined as the sum of metal loss volumes measured in the bar.

IV. RESULTS

Three quantities are predicted by FNN: CR, metal loss and area of defects. For each variable, a FNN is implemented. CR value derives from the dataset of comparison between 2005 and 2012. Metal loss and relative area of defects derive from 2012 dataset. These dataset are preprocessed selecting defects with depth larger than 30% of pipeline thickness. Then, the quantities integrated on each bar are calculated, as explained above. Finally, an upper bound corresponding to 90th percentile and a lower bound corresponding to 10th percentile are fixed, and only the data between these values are processed. The sample finally consists of only 150 bars.

Only inputs with SSDi larger than one half of maximum SSDi are maintained as network inputs. As evident, flow characteristics play a crucial role (in particular gas velocity, liquid velocity and hold-up) as well as geometric pipeline features. In this sense, both kinds of inputs have to be considered as network inputs. Training is also improved by feeding the FNN with mechanistic de Waard model.

2121212121

Figure 9 Sensitivity analysis for CR prediction.

Results of CR prediction are shown in Figure 10 and Figure 11. In particular, Figure 10 shows the CR profile along the pipeline: black dots are the FNN outputs, while the red dots are the observed CR values (targets). Figure 11 reports the corresponding scatter plot.

Figure 10 shows also the outputs of the two deterministic models above described: de Waard model (black line) and NORSOK model (red line). As evident, the FNN model gives a prediction far more accurate than deterministic models.

Figure 10 CR profile along the pipeline: black dots are the FNN outputs, while the red dots are the observed CR values (targets). Also two deterministic models are reported: de Waard model (black line) and NORSOK model (red line)

Figure 11 CR scatter plot: on x-axis the CR predicted by the model, on the y-axis the CR observed after the comparison of the two ILI. In a perfect prediction, all dots should be positioned along the identity line (black line)

Figure 12 reports the relative corroded area profile along the pipeline, whereas Figure 13 the metal loss profile along the pipeline.

Figure 12 Relative corroded area profile along the pipeline: black dots are the FNN outputs, while the red dots are the observed CR values (targets)

Figure 13 Metal loss profile along the pipeline: black dots are the FNN outputs, while the red dots are the observed CR values (targets)

The FNN prediction performance is evaluated by four measurements: correlation coefficient (R), ranging from 0 to 1, root mean square percentage error (RMSPE), mean absolute percentage error (MAPE) both lower bounded to 0:

� �� �

� � � �� �

�

� �

�

��

���

N

i

N

iii

N

iii

oopp

ooppR

1 1

22

1 (13)

Np

op

RMSPE

N

i i

ii��

��

���

�

�� 1

2

100 (14)

��

���

N

i i

ii

oop

NMAPE

1

1100 (15)

Finally, the scatter index (SI) is calculated. It is lower bounded to 0: closer values to 0 indicate better agreement between the observed and forecasted time series.

oRMSESI � (16)

Statistical measurements of FNN performance are provided in TABLE 1, comparing with the same quantities from de Waard and NORSOK model results.

0.000 0.020 0.040 0.060 0.080 0.100 0.120 0.140

ElevationInclinationConcavity

SectorRegimeHold�upPressureGas�flowTotal�flow

Liquid�velocityGas�velocity

De�Waard�modelNORSOK�model

SSD

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 5000 10000 15000 20000

CR(m

m/y

)

KP(m)

output

target

De�Waard

NORSOK

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

CR�ta

rget

�(mm

/y)

CR�output�(mm/y)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 5000 10000 15000 20000

Rela

tive�

Area

�

KP(m)

output

target

0

10000

20000

30000

40000

50000

60000

0 5000 10000 15000 20000

Met

al�lo

ss�(m

m^3

)

KP(m)

output

target

2222222222

The FNN model outperforms standard deterministic models, showing better statistical measurements. The no excellent performances of the FNN model are explained by the very small sample to train the network model, which is the main drawback of this case study.

V. CONCLUSIONS

Corrosion is the main cause of deterioration of pipelines. Therefore, prediction of internal corrosion along the pipeline profile is a critical issue for the Oil&Gas sector, particularly for new frontiers of ultra-deep waters, where remote treatment is performed on floating processing units: in this case, also flowlines and risers are subjected to corrosion. A correct corrosion assessment impacts metallurgy (for instance, the choice between carbon steel, stainless steel or alloys) and therefore pipeline costs. A reliable prediction of pipeline sections more exposed to corrosion risk would help also the pipeline integrity management, reducing the economic impact. Furthermore, given the worldwide increasing number of old pipelines, this issue is particularly relevant also to avoid pipeline failures and to reduce environmental impact.

In this paper, the prediction of internal corrosion along the pipeline profile is performed by a data-driven model, given the available measurements derived from two internal line inspections. In the best of our knowledge, this is the first application of ANN to prediction of local corrosion along a pipeline. Given the complexity of the phenomenon, this is a very hard task.

A case study has been considered relatively to a pipeline 20km long, built in 1997, where the corrosion is due to CO2contribution and bacteria activity. Being the corrosion phenomenon due to different mechanisms, a deterministic approach is not able to reproduce the corrosion rate and the defects distribution observed during pigging activity.

Therefore, an artificial intelligence model has been investigated, considering several contributions to corrosion as network input. In particular:

� Geometrical pipeline features � Fluid dynamic multiphase variables � Deterministic models

By a sensitivity analysis, it has been demonstrated that all these three components play an important role in network training and simulation. This strategy strongly improves the results obtained by deterministic models, usually considered in literature. A mean absolute percentage error equal to 30% is reached. Given the high uncertainty inherent to real internal corrosion problem, this can be considered a good result.

Predictions can be further improved in the future considering larger datasets (with several real pipeline cases and different flow conditions) that would allow to improve the generalization of the model and to extend it to different conditions.

REFERENCES

[1] R. Nyborg, “Controlling internal corrosion in Oil and Gas pipelines”, business briefing : exploration & production: the oil & gas review, Issue 2, 2005

[2] S. Nesic´, "Key issues related to modelling of internal corrosion of oil and gas pipelines – A review", Corrosion Science 49, 4308–4338, 2007

[3] C.De Waard, D.E.Milliams: “Carbonic Acid Corrosion of Steel”, Corrosion1975, Paper N°31, 1975.

[4] C.De Waard, U.Lotz, “Prediction of CO2 Corrosion of Carbon Steel”, Corrosion93, Paper N°69, 1993.

[5] C.De Waard, U.Lotz, Dugstad: “Influence of Liquid Flow Velocity on Corrosion: a Semi-Empirical Model”, NACE, Corrosion 95 conference, Paper N°128, 1995.

[6] "CO2 corrosion rate calculation model", NORSOK STANDARD M-506, Rev. 2, June 2005

[7] S. Nesic, M. Nordsveen, N. Maxwell, and M. Vrhovac, Probabilistic modelling of CO2 corrosion laboratory data using neural networks. Corrosion Science, 43 7: 1373-1392, 2001.

[8] S. Hernández, S. Nesic, G. Weckman, V. Ghai, "Use of Artificial Neural Networks for Predicting Crude Oil Effect on CO2 Corrosion of Carbon Steels", NACE Corrosion 2005 conference, Paper No. 05554, 2005

[9] G.Gabetta, S.P.Trasatti, “Analysis of CO2 corrosion model by Neural Networks”, Proceedings of EUROCORR 2006, Maastricht, The Netherlands, 2006

[10] OLGA, Multiphase Flow Simulator, by ScandPower Petroleum technology.

[11] P. O. Gartland, N. N.Bich, “Internal corrosion of dry gas pipelines during upsets”, NACE Corrosion 2004 conference, Paper No.04199, 2004

[12] P. O. Gartland, R. Johnsen, I. Ovstetun, “Application of Internal Corrosion Modeling in the risk assessment of Pipelines”, NACE Corrosion 2003 conference, Paper No.03179, 2003

[13] S. Haykin, Neural Networks, A comprehensive foundation, Pearson, Prentice Hall, 1999

[14] MATLAB Version: 8.0.0.783 (R2012b), Neural Network Toolbox Ver. 8.0

[15] M.T. Hagan, and M. Menhaj, "Training feed-forward networks with the Marquardt algorithm," IEEE Transactions on Neural Networks, Vol. 5, No. 6, pp. 989–993, 1994, 1999.

[16] Y. Dimopoulos, P. Bourret, S. Lek, “Use of some sensitivity criteria for choosing networks with good generalization ability,” Neural Processing Letters Vol. 2, p. 1-4, 1995.

TABLE 1 Statistical measurements of network performance,compared with two deterministic models (de Waard and NORSOKmodels)

Model RMSPE MAPE SI R FNN 52 31 0.34 0.66

de Waard 95 95 1.03 0.08 NORSOK 95 95 1.03 0.08

2323232323