4 Frst Internatonal Conference on Systems Informatcs, Modellng and Smulaton A eural etwork Predctve Model of Ppelne Internal Corroson Profle Gula De Mas, Roberta Vch, Manuela Gentle, Roberto Brusch ADVE Dept. Sapem Spa Fano (PU), Italy Gula.demas@sapem.com Govanna Gabetta SIAV Dept. EI Spa Mlan, Italy Abstract Internal corroson s a crucal ssue for the safe operaton of ol&gas ppelnes. Ths s a phenomenon due to nteracton of dfferent mechansms. Water and electrochemstry, protectve scales, flow velocty, steel composton and localzed bactera attacks are relevant. Despte the large number of models proposed n lterature, the corroson process s very complex and rarely reproduced by exstng models. For ths reason, an artfcal neural network (A) based model s nvestgated, wth the am to correctly predct the presence of metal loss and corroson rate along a ppelne. In ths paper, a case study s consdered, based on real feld data. The model ntegrates the geometrcal profle of a real ppelne, flow smulatons and the most mportant determnstc corroson models. It s shown that the A model outperforms the determnstc ones. Keywords-nternal corroson predcton, ol&gas ppelne, neural network, partal dervatve senstvty method I. ITRODUCTIO Multphase transport wll have a major mpact on offshore development durng the next decade. In the past, emphass was placed on preprocessng the multphase well stream through separaton on platforms or even subsea, close to the wells. Drastc reducton n both nvestment and operatng costs can be acheved when unprocessed, multphase well streams can be transported over longer dstances n carbon steel ppelnes from subsea wells to man platforms, exstng nstallatons on neghborng felds or onshore processng facltes. Ppelne cost s a consderable part of the nvestment n subsea projects. For long-dstance and large-dameter ppelnes, cost can become prohbtvely hgh f the corrosvty of the flud necesstates the use of corrosonresstant alloys nstead of carbon steel. Better understandng and control of the corroson of carbon steel can help to ncrease applcablty and therefore have a large economc mpact[]. In ths paper we focus on predcton of generalzed and localzed corroson generatng metal loss. A dfferent mechansm of corroson s nstead stress corroson crackng (.e. H S stress corroson crackng) that should be addressed durng desgn stage, through a rght selecton of base materal and welds fabrcaton process. Internal corroson s one of the man causes of deteroraton of ppelnes, partcularly n presence of water. The water content nsde the ppelne s partally due to formaton water, partally to upsets, partally to sea water entrance as a consequence of falures. Corroson of carbon steel may be nfluenced by many factors: CO (sweet corroson), H S (sour corroson), water chemstry, flow velocty, ol or water wettng and composton and surface condton of the steel[]. A small change n one of these parameters can change the corroson rate consderably, due to changes n the propertes of the thn layer of corroson products that accumulates on the steel surface. For nstance, a ppelne not affected by corroson for many years, can be subjected to hgh corroson (several mm per year) after changng flow characterstcs. The corroson rate can be reduced substantally under condtons where ron carbonate (FeCO 3 ) can precptate on the steel surface and form a dense and protectve corroson product flm. When n some localzed regons ths depost s no more adherent to steel surface, localzed corroson can appear. Ths may happen due to oxde layers formaton wth subsequent formaton of underlyng crevces where fast corroson (even tmes hgher than unform corroson) can occur (crevce corroson), or due to turbulent/slug flow drven eroson (mesa corroson). Another source of localzed corroson s due to acton of bactera (MIC, mcrobologcally nduced corroson) that occurs where stagnant water s present. In order to control corroson n ppelnes, t s mportant to understand the underlyng degradaton mechansms and to predct whether corroson wll be ntated, whch sectons of ppelne wth have hgher rsk of corroson, and how t can be prevented. Several models to predct CO corroson can be found n lterature[3][4][5][6]. These models are manly based on laboratory data and, n some cases, are valdated wth feld data. Moreover, they can be classfed as mechanstc models, sem-emprcal models and emprcal models[]. They may gve markedly dfferent corroson rate predctons for the same feld case, and whch models are most successful n ther predcton vary consderably from case to case. In ths paper, a nondetermnstc artfcal ntellgence model s proposed wth the am to ncrease the accuracy of predcton of occurrence of corroson and of corroson rate. Ths model s based on neural network technque. Ths knd of models outperforms determnstc models when they have 978--7695-598-/4 $3. 4 IEEE DOI.9/SIMS.4.4 8
to represent very complex hghly non-lnear phenomena. The applcaton of artfcal neural networks (A) has been already proposed n lterature to predct the average corroson of ppelnes[7][8][9]. In the present study nstead, the model s focused on predctng the corroson profle along the ppelne, to dentfy the ppelne sectons more exposed to corroson rsk. The neural network model here proposed ntegrates geometrcal characterstcs of a ppelne (an applcaton case s consdered), corroson determnstc models and smulatons of multphase flow velocty and transport, as schematzed n Fgure. 4 5 3 4 3 3 3 4 4 5 4 6 8 4 6 8 KP(Km) 4.5 3.3...3 3 concavty nclnaton( ) 4.5 4 6 8 4 6 8 ( ) Fgure Elevaton (black lne), nclnaton (red lne, top panel) and concavty (red lne, bottom panel) Fgure Scheme of artfcal ntellgence model Ths tool represents an ntegrated process of corroson analyss, useful both for ppelne desgn and for ntegrty management. II. METHODOLOGY A. Geometrcal characterzaton The ppelne has been characterzed by ts geometrcal features: elevaton, nclnaton and concavty. Inclnaton s demonstrated to play an mportant role n corroson process [], because above certan crtcal angles water holdup and therefore rsk of corroson ncrease. Concavty s expected to be mportant, determnng water accumulaton. Inclnaton I s related to ppelne elevaton e n the pont as: () Whereas concavty C s defned as: () In Fgure elevaton, nclnaton and concavty are shown. B. Multphase flow parameters Multphase flow modellng s based on OLGA software[]. Ths program provdes nformaton on temperature profle along the ppelne, pressure profle, velocty profles of each phase, phase hold-ups and flow regmes, gven boundary pressure, temperature values and flow composton. Water plays a crucal role for corroson, enhancng corroson rate dependng on ts hold-up and velocty, gas flow rate, pressure and temperature and ppelne nclnaton[]. In our specfc case, water can be consdered a phase separated from gas, at the bottom of ppe. The multphase flow smulator can help to dentfy locatons where varaton n flow regme, flow velocty and water accumulaton may ncrease the rsk of corroson damage[]. Flud regme s descrbed by a dscrete number as follows: : stratfed flow : annular flow 3: slug flow 4: bubble flow As evdent from Fgure 3, n the present case the flow regme s usually stratfed or slug. Fgure 4 and Fgure 5 report gas velocty and water velocty along the ppelne, as provded by OLGA smulator. 3 3 4 5 5 Fgure 3 Flud regme and ppelne elevaton 4 3 IDfludregme 9
3 3 Layer Layer gasvelocty(m/s) nput IW b HW + f f b + 3 4 5 5 Fgure 4 Gas velocty and ppelne elevaton 3 3.5.3...3 4.5 5 5 Lqudvelocty(m/s) Fgure 6 Fttng neural network (F) block dagram The F ntegrates all the above quanttes as nput values. Therefore, nput varables are of three types: Geometrcal ppelne characterstcs (elevaton, nclnaton and concavty) Flud dynamc multphase varables (flow regme, pressure, gas flow, total flow, lqud velocty, gas velocty) Determnstc models (de Waard and ORSOK) Each network has only one. Three alteratve varables are consdered: Corroson rate (CR) Metal loss Area of defects The network structure s reported n Fgure 7. Elevaton Input layer {x } Fgure 5 Water velocty and ppelne elevaton C. Determnstc models Two determnstc models are ntegrated n the artfcal ntellgence model here proposed. The frst one s the de Waard model [4], whch correlates corroson rate to temperature t and CO partal pressure (pco ), by the followng relatonshp: (3) Inclnaton Concavty Flow regme Hold-up Pressure Gas flow Total flow Lqud velocty Hdden layer {h j } Output layer {o k } Bas CR/ metal loss/ defect area The second model s proposed by ORSOK[6]: CR s an emprcal functon of temperature t, CO, ph, wall shear stress. For temperature between C and C: (4) For t=5 C (5) At temperature 5 C (6) D. Artfcal neural network In the present study, a fttng neural network s used[3]. Fttng networks (F) are feedforward neural networks used to ft an nput- relatonshp [4], as shown n Fgure 6. Gas velocty De Waard model ORSOK model Fgure 7 F archtecture wth all nputs and alternatve s (ntermedate arrows are not ndcated) Several tranng algorthms were tested; fnally the Levenberg-Marquardt back propagaton algorthm was selected as the one producng best predcton[5]. Two (or more) layer fttng networks can ft any fnte nput- nonlnear relatonshp arbtrarly well, gven enough hdden neurons: n the present case hdden neurons are demonstrated to obtan the best network performance. Bas
E. Senstvty analyss eural network s a black box type model and t s not evdent the partcpaton of each of the nput varables. In ths study, a smple method based on the use of the partal dervatves of the network response wth respect to each nput s used [6]. The lnk between nputs modfcaton, x, and s varaton, ok, s the Jacoban matrx dok/dx. It represents the senstvty of the network s accordng to small nput perturbatons. For a network wth I nputs, one hdden layer wth J nodes, and one (K=) (6. years dstant each other). The of nspectons conssts of: defect poston (both dstance from nlet and orentaton n ppelne crcumference), depth, wdth and length. In Fgure 8 the defect dstrbuton along the ppelne s reported: as evdent, 76% of defects are located n the last 8km of the ppelne length. (7) (8) Therefore, the gradent vector of ok wth respect to x s gven by (9) where f s a logstc sgmod functon, f s a lnear functon, n the present case. The senstvty of the F wth respect to nput x s gven by the sum over the p observatons of the square of Eq.9: () Ths fnal SSD s normalzed as: () The dervatve can be effcently computed as mnor extenson to the backpropagaton algorthm used for tranng. III. APPLICATIO In the present study an applcaton to a ppelne km long n Medterranean Sea s nvestgated. It conssts of around 7 bars, each m long. It has been bult around 4 years ago. Ths s an old ppelne, wth a complex hstory. Formaton water was frstly njected n the ppe n the frst years, then separated at the nlet wth an effcency of 95%. The temperature and pressure profle and flow velocty changed durng the lfetme of the ppelne. For nstance, operatng pressure decreased to about half from 4 to. Moreover, several reparatons have been performed (partcularly near the end part of the lne). The man mechansms of corroson are CO and MIC. Moreover, bottom of lne (BoL) corroson s much more probable than top of lne (ToL) corroson (94% of defects are between 4: and 8: angles). A. Corroson measurements In order to reduce falure ncdents caused by nternal corroson, pggng nternal lne nspectons (ILI) are performed to montor corroson and nspect crtcal parts of ppelnes. In partcular, two ILI have been performed for the ppelne of nterest, one durng 5 and one durng Fgure 8 Defect dstrbuton along the ppelne A comparson between the two ILI has been done, selectng the most relevant defects, wth a depth larger than 35% of ppe thckness. For these defects the corroson rate (CR) has been calculated as () The paper consders both sets of data. Moreover, for each bar, the total number of defects s calculated as well as average corroson rate. To each bar are assocated: a relatve area of defects, defned as the sum of defect areas dvded by bar area and a value of metal loss, defned as the sum of metal loss volumes measured n the bar. IV. RESULTS Three quanttes are predcted by F: CR, metal loss and area of defects. For each varable, a F s mplemented. CR value derves from the dataset of comparson between 5 and. Metal loss and relatve area of defects derve from dataset. These dataset are preprocessed selectng defects wth depth larger than 3% of ppelne thckness. Then, the quanttes ntegrated on each bar are calculated, as explaned above. Fnally, an upper bound correspondng to 9 th percentle and a lower bound correspondng to th percentle are fxed, and only the data between these values are processed. The sample fnally conssts of only 5 bars. Only nputs wth SSD larger than one half of maxmum SSD are mantaned as network nputs. As evdent, flow characterstcs play a crucal role (n partcular gas velocty, lqud velocty and hold-up) as well as geometrc ppelne features. In ths sense, both knds of nputs have to be consdered as network nputs. Tranng s also mproved by feedng the F wth mechanstc de Waard model.
ORSOKmodel DeWaardmodel Gasvelocty Lqudvelocty Totalflow Gasflow Pressure Holdup Regme Sector Concavty Inclnaton Elevaton...4.6.8...4 SSD RelatveArea.9.8.7.6.5.4.3.. 5 5 target Fgure 9 Senstvty analyss for CR predcton. Results of CR predcton are shown n Fgure and Fgure. In partcular, Fgure shows the CR profle along the ppelne: black dots are the F s, whle the red dots are the observed CR values (targets). Fgure reports the correspondng scatter plot. Fgure shows also the s of the two determnstc models above descrbed: de Waard model (black lne) and ORSOK model (red lne). As evdent, the F model gves a predcton far more accurate than determnstc models. CR(mm/y).8.7.6.5.4.3.. 5 5 Fgure CR profle along the ppelne: black dots are the F s, whle the red dots are the observed CR values (targets). Also two determnstc models are reported: de Waard model (black lne) and ORSOK model (red lne) CRtarget(mm/y).8.7.6.5.4.3.. Fgure CR scatter plot: on x-axs the CR predcted by the model, on the y-axs the CR observed after the comparson of the two ILI. In a perfect predcton, all dots should be postoned along the dentty lne (black lne) Fgure reports the relatve corroded area profle along the ppelne, whereas Fgure 3 the metal loss profle along the ppelne. target DeWaard ORSOK...3.4.5.6.7.8 CR(mm/y) Fgure Relatve corroded area profle along the ppelne: black dots are the F s, whle the red dots are the observed CR values (targets) Metalloss(mm^3) Fgure 3 Metal loss profle along the ppelne: black dots are the F s, whle the red dots are the observed CR values (targets) The F predcton performance s evaluated by four measurements: correlaton coeffcent (R), rangng from to, root mean square percentage error (RMSPE), mean absolute percentage error (MAPE) both lower bounded to : R 6 5 4 3 5 5 p po o p p o o (3) p o p RMSPE (4) p o MAPE (5) o Fnally, the scatter ndex (SI) s calculated. It s lower bounded to : closer values to ndcate better agreement between the observed and forecasted tme seres. RMSE SI (6) o Statstcal measurements of F performance are provded n TABLE, comparng wth the same quanttes from de Waard and ORSOK model results. target
TABLE Statstcal measurements of network performance, compared wth two determnstc models (de Waard and ORSOK models) Model RMSPE MAPE SI R F 5 3.34.66 de Waard 95 95.3.8 ORSOK 95 95.3.8 The F model outperforms standard determnstc models, showng better statstcal measurements. The no excellent performances of the F model are explaned by the very small sample to tran the network model, whch s the man drawback of ths case study. V. COCLUSIOS Corroson s the man cause of deteroraton of ppelnes. Therefore, predcton of nternal corroson along the ppelne profle s a crtcal ssue for the Ol&Gas sector, partcularly for new fronters of ultra-deep waters, where remote treatment s performed on floatng processng unts: n ths case, also flowlnes and rsers are subjected to corroson. A correct corroson assessment mpacts metallurgy (for nstance, the choce between carbon steel, stanless steel or alloys) and therefore ppelne costs. A relable predcton of ppelne sectons more exposed to corroson rsk would help also the ppelne ntegrty management, reducng the economc mpact. Furthermore, gven the worldwde ncreasng number of old ppelnes, ths ssue s partcularly relevant also to avod ppelne falures and to reduce envronmental mpact. In ths paper, the predcton of nternal corroson along the ppelne profle s performed by a data-drven model, gven the avalable measurements derved from two nternal lne nspectons. In the best of our knowledge, ths s the frst applcaton of A to predcton of local corroson along a ppelne. Gven the complexty of the phenomenon, ths s a very hard task. A case study has been consdered relatvely to a ppelne km long, bult n 997, where the corroson s due to CO contrbuton and bactera actvty. Beng the corroson phenomenon due to dfferent mechansms, a determnstc approach s not able to reproduce the corroson rate and the defects dstrbuton observed durng pggng actvty. Therefore, an artfcal ntellgence model has been nvestgated, consderng several contrbutons to corroson as network nput. In partcular: Geometrcal ppelne features Flud dynamc multphase varables Determnstc models By a senstvty analyss, t has been demonstrated that all these three components play an mportant role n network tranng and smulaton. Ths strategy strongly mproves the results obtaned by determnstc models, usually consdered n lterature. A mean absolute percentage error equal to 3% s reached. Gven the hgh uncertanty nherent to real nternal corroson problem, ths can be consdered a good result. Predctons can be further mproved n the future consderng larger datasets (wth several real ppelne cases and dfferent flow condtons) that would allow to mprove the generalzaton of the model and to extend t to dfferent condtons. REFERECES [] R. yborg, Controllng nternal corroson n Ol and Gas ppelnes, busness brefng : exploraton & producton: the ol & gas revew, Issue, 5 [] S. esc, "Key ssues related to modellng of nternal corroson of ol and gas ppelnes A revew", Corroson Scence 49, 438 4338, 7 [3] C.De Waard, D.E.Mllams: Carbonc Acd Corroson of Steel, Corroson975, Paper 3, 975. [4] C.De Waard, U.Lotz, Predcton of CO Corroson of Carbon Steel, Corroson93, Paper 69, 993. [5] C.De Waard, U.Lotz, Dugstad: Influence of Lqud Flow Velocty on Corroson: a Sem-Emprcal Model, ACE, Corroson 95 conference, Paper 8, 995. [6] "CO corroson rate calculaton model", ORSOK STADARD M- 56, Rev., June 5 [7] S. esc, M. ordsveen,. Maxwell, and M. Vrhovac, Probablstc modellng of CO corroson laboratory data usng neural networks. Corroson Scence, 43 7: 373-39,. [8] S. Hernández, S. esc, G. Weckman, V. Gha, "Use of Artfcal eural etworks for Predctng Crude Ol Effect on CO Corroson of Carbon Steels", ACE Corroson 5 conference, Paper o. 5554, 5 [9] G.Gabetta, S.P.Trasatt, Analyss of CO corroson model by eural etworks, Proceedngs of EUROCORR 6, Maastrcht, The etherlands, 6 [] OLGA, Multphase Flow Smulator, by ScandPower Petroleum technology. [] P. O. Gartland,..Bch, Internal corroson of dry gas ppelnes durng upsets, ACE Corroson 4 conference, Paper o.499, 4 [] P. O. Gartland, R. Johnsen, I. Ovstetun, Applcaton of Internal Corroson Modelng n the rsk assessment of Ppelnes, ACE Corroson 3 conference, Paper o.379, 3 [3] S. Haykn, eural etworks, A comprehensve foundaton, Pearson, Prentce Hall, 999 [4] MATLAB Verson: 8...783 (Rb), eural etwork Toolbox Ver. 8. [5] M.T. Hagan, and M. Menhaj, "Tranng feed-forward networks wth the Marquardt algorthm," IEEE Transactons on eural etworks, Vol. 5, o. 6, pp. 989 993, 994, 999. [6] Y. Dmopoulos, P. Bourret, S. Lek, Use of some senstvty crtera for choosng networks wth good generalzaton ablty, eural Processng Letters Vol., p. -4, 995. 3