Proceedings of the Sixteenth (2006) International Offshore and Polar Engineering Conference San Francisco, California, USA, May

28-June 2, 2006 Copyright © 2006 by The International Society of Offshore and Polar Engineers ISBN 1-880653-66-4 (Set); ISSN 1098-6189 (Set)

Extreme Response Analyses of Marlin TLP Tendon Tension during Hurricane Ivan
Guang Li
EPTG, BP Houston, TX, USA

Ronald N. Perego
DWP, BP Houston, TX, USA

David L. Garrett
SES Houston, TX, USA

Table 1: Marlin Principal Hull and Tendon Dimensions

ABSTRACT
In order to verify design factors of Marlin TLP related to extreme tension response, a statistical study on the tension data recorded during Hurricane Ivan is conducted. We discuss the effect of the bandwidth and nonstationarity on the extreme tension estimate and conclude that the direct application of the Weibull method overestimates the extreme tension. Then, a normalization method is proposed to minimize the effect of nonstationarity. We apply this method to both stationary and nonstationary simulated data to show its robustness and effectiveness. We then apply this method to recorded tension data to calculate the observed peak factors1 (PF) during each time interval. Finally, the peak factor distributions (histograms) are calculated and compared with those of Gaussian processes. The study results are also compared with design factors. Column Diameter Column Centerline Spacing Column Height Pontoon Width Pontoon Height Draft Tendon OD Tendon Wall Thickness 50 ft (15.24 m) 160 ft (48.8 m) 140 ft (42.7 m) 25 ft (7.6 m) 22.5 ft (6.9 m) 71 ft (21.6 m) 28 in (0.71 m) Stepped 1.10”-1.05”-1.10” (0.028 – 0.027 – 0.028 m)

KEY WORDS: TLP; tension; peak factor; Weibull; Ivan; nonstationarity; bandwidth. INTRODUCTION
The Marlin tension leg platform (TLP) was installed in the Gulf of Mexico in 1999 on Viosca Knoll Block 915. Marlin was designed for BP by ABB (1997) using an uncoupled, frequency domain numerical solution, with tendons modeled as springs, incorporating a number of empirical design factors derived from model tests. Fig. 1: Marlin TLP Photo The water depth at the platform site is approximately 3,250 ft (991 m). The TLP consists of a four column hull connected by ring pontoons and is moored using two tendons per corner. The TLP has a displacement of approximately 52,000 kips (23,608 tons). Table 1 summarizes the principal hull and tendon dimensions. Figure 1 is a photo of the platform. A fitness study was commissioned in 2003 to evaluate the platform’s payload capacity limits in light of additional development opportunities in the area. Payload capacity was found to be governed by the limit state of minimum tension failure that could lead to tendon unlatching (Li, Banon, and Perego, 2005). The acceptance criterion for this case was the requirement to maintain a positive most-probable-minimum (MPM) tension for any single hull compartment flooded case during a maximum operating storm environment. The MPM tension is calculated using the following formula

Peak factor is obtained by normalizing the difference between tension peak and mean tension against the RMS tension.

1

TMPM = Tmean − γ MPM Trms

(1)

Paper No. 2006-JSC-356

Li

73

1

Ivan passed over the Viosca Knoll area late on September 15.. Ivan was an intense hurricane that reached Category 5 strength three times during its life. However. 4: RMS Tensions during Hurricane Ivan Paper No. the most probable value of a Rayleigh distribution (Ochi. 6 and 7 was recorded at 4 Hz during Hurricane Ivan by the IMMS (Integrated Marine Monitoring System).5 in the plots) on September 15th. 2: Track of Hurricane Ivan DATA PROPERTIES Tendon Tension Monitoring System (TTMS) status channels indicated that some of the individual sensors were out of tolerance after 12:00 CST (date/time of 15. 1973).e. Figs. only data between 0:00 CST on September 14th and 12:00 CST on September 15th is used in our analysis. 2). During this time. 3. 3 and 4 show the mean and RMS tensions calculated at 1800-second non-overlapping intervals during Hurricane Ivan. Hence. Nonstationarity Design tensions are typically based on the assumption of stationary seastate. Fig. The tension data of tendon 2. 9/15 12Z -89 -88 -87 -86 Longitude (deg) Fig.5%. A one-second moving average was applied to the raw data to remove the local noise.where γ MPM denotes the absolute value of the peak factor (PF) associated with exceeding probability of 62. It is important to verify that analytical tools and corresponding assumptions used for the design and analysis of these systems represent the true response of the floating production system. the mean and RMS tensions are assumed to be independent of time. 3: Mean Tensions during Hurricane Ivan Buoy 42040 9/16 06Z Latitude (deg) MARLIN 29 28 27 26 -90 9/16 00Z Closest Model Grid Point 9/15 18Z NOTE: Circle indicates approximate radius to maximum winds. We will discuss our treatment in detail in the next section. the environment and the facility. 2004. the tension response is also assumed to be stationary. Therefore. It caused extensive damage and loss of life while in the Caribbean. it rapidly weakened from Category 5. and was at Category 3 at its closest approach to Marlin at approximately 18:00 CST on September 15th (Fig. Perego et al (2005) compared the measured platform response with the predicted values. the trend of RMS tension requires careful treatment when the extreme tension response is studied. and for future expansion. To this end. for operations. wave. As the necessary conditions. This requires real time field monitoring of the FPS. and current environment. Sustained flight level winds at this time were 110 knots. The trend of mean tension can be removed by subtracting the linear fit to the tension trace. 31 30 Fig. Understanding how a floating production system (FPS) such as Marlin responds to the environment is critical for protection of personnel. Their trends with respect to time clearly illustrate the nonstationarity of the recorded data. 2006-JSC-356 Li 74 2 . BP has extensively monitored the motions and tendon tensions of the Marlin TLP along with the wind. Extensive environment and platform response data was recorded during Hurricane Ivan in 2004. i.

the PFs predicted based on model test results are often conservative. As shown in Fig. Fig. 2006-JSC-356 Li 75 3 . A narrowband process has a single dominant frequency component while a broadband process has its components more uniformly distributed. Then the distribution can be extrapolated to target probability level to determine the corresponding PF. 8. the tension responses are not exactly narrowband. In the subsequent section. neither the seastates during a hurricane nor the corresponding platform responses are stationary. The nonstationarity elevates the tail of the peak distribution as compared to the underlying stationary process. we simulated a 40-minute zero-meaned tension history based on a power spectrum calculated from recorded tension data.9. only the peaks of a narrowband Gaussian process follow the Rayleigh distribution (a special case of Weibull). the α values of tension histories are generally between 0. 6: Weibull Fit to Observed Peak Factors Fig. we plotted the PF (z) distributions for both processes. a direct application of the Weibull method to the data should be avoided. The upper and the lower plots show Weibull fits to the maxima and the minima. As shown in the last section. In addition. their bandwidths increase during Hurricane Ivan. 1997). Fig. 5. a Weibull distribution can be fitted to a subset of observed peaks (maxima or minima) from a time series of limited duration (e. 6 illustrates this method. However. 7 shows the shapes of the conditional distributions of positive peaks for Gaussian processes with different bandwidth. 20 minutes). The slope of the trend was chosen to increase the RMS tension by 50% over the 40-minute interval. Fig. 2000). Then. Hence. Because the tension data from model tests is not necessarily narrowband. respectively.6 and 0. Because actual peak probability distribution is something between a Gaussian ( α = 0 ) and a Rayleigh ( α = 1 ) distribution (Bendat and Piersol. the more overestimation the Weibull method creates. 7: Peak Distributions of Gaussian Processes with Different Bandwidth Additionally. where ‘+’ symbols denote the PF observed in a 30-minute interval of tension history. Paper No. we normalized the tension history against its RMS value and multiplied the simulated data by a linear trend so that it becomes nonstationary. 5: α Values during Hurricane Ivan PEAK FACTOR CALCULATION Weibull Method As an approach commonly used in model tests (Mercier et al. nonstationarity in the recorded data can also cause this method to overestimate the MPM PF. As an illustration. we will examine the impact of bandwidth on PF determination. the Weibull fit method often overestimates the MPM PF The larger the subset of data is used for the Weibull fit.Bandwidth Bandwidth of a tension history indicates the distribution range of its periodic components in the frequency domain. Finally. The peak distribution of a random Gaussian process asymptotically approaches a Weibull distribution as the bandwidth of this process approaches zero ( α approaches unity). Fig.g. Increase of the bandwidth of a random process will increase its mean-crossing rate and decrease the ratio between the number of mean-crossings and the number of peaks (defined as α ). a Weibull fit to the peaks of a nonstationary process will overestimate the MPM PF. as a result. as shown in Fig.

0100 0.00 6.1.00 0.00 z 2. Paper No.00 4. all PFs in the subsequent text refer to ones obtained from normalized processes.0010 -3.0000 Normalization Method In order to remove the trend in the mean and RMS tension.00 -1.1000 Prob > Normalized Stationary 0.00 3.00 0. 1.00 z 2. Because the nonstationarity elevates the tail of the peak distribution as compared to the underlying stationary process (as shown in Fig. Figure 11 shows the peak distribution of the normalized process against the stationary process that underlies the nonstationary process.00 5.00 -1.00 1. 10 shows the peak distributions of the stationary and the normalized processes. The normalization method had little effect on the peak distribution of a stationary process. 1.0000 Normalized Stationary 0.00 -1.0000 NonStationary Stationary 0. Fig. We further average the four maxima and four minima in each interval and plot the averaged values in Fig. We applied this normalization method to the nonstationary process we created in the last subsection and obtained a normalized process. 9: Peak Distributions of Nonstationary and Normalized Processes To show the robustness of this method. approximately.00 z 2.) All tension data is normalized using the aforementioned method. 8). Therefore.0010 -3.1000 Prob > 0.00 6.00 -1.00 Fig. 11: Peak Distributions of Stationary and Normalized Processes In order to avoid the error in MPM PF calculation due to broadbandness.00 5.00 4.0100 0.0000 NonStationary Normalized 0.00 3. The normalized process u has zero mean and unit STD.00 1.00 4. unless noted otherwise.1000 Prob > 1.00 0.00 1.00 Fig.0100 0.00 5.00 2. respectively.1000 Prob > 0. 8: Peak Distributions of Stationary and Nonstationary Processes Fig. 0. we propose a normalization method based on the following formula ( − ) u (t ) = T (t ) mT (t ) σ T (t ) (2) where mT (t ) and σ T (t ) are the average and STD of tension T over a 10-minute moving window centered at time t.00 z 3. we calculate the maxim and minim observed PFs in each half-hour interval during Hurricane Ivan (one maximum and one minimum from each tendon in each interval.00 0.00 5. 2006-JSC-356 Li 76 4 .0100 0. Fig.0010 -2.00 Fig. The normalization method effectively recovers the peak distribution of the underlying stationary process.00 4.0010 -3. 12. 10: Stationary Peak Distributions Before and After Normalization To show the effectiveness of this method. we conclude that our normalization method can reduce the nonstationarity effect on peak distributions.00 3. 9 compares the peak distributions of these two processes and reveals that our normalization method can lower the tail of peak distribution of the nonstationary process.00 -2.00 -2.00 -2.00 0. we applied it to a stationary process.00 1. we applied it to a nonstationary process with 50% increase of its RMS tension over 40 minutes.

48) than the absolute values of the minim PFs. The MPM PF of a Gaussian process can be calculated using the following equation (Ochi.2 Fig. All tension data is normalized using the aforementioned method.6 3. We also calculate the PF corresponding to 62. Also. The Ochi Factor approximates the observed MPM PF closely. minimum tension is the governing condition for the platform. A similar formula has been adopted in API RP 2T (American Petroleum Institute.0 14. together with that of a Gaussian process. We further apply Eq.6 Ochi Peak Factor 3. 1981) Peak Factor Fig. consistent with our assumption made earlier that the normalized tension process is stationary throughout the storm. as the effect of nonstationarity is removed through the normalization process. The design ranges (ABB.5% exceeding probability based on the observed histograms. we can conservatively combine the observed maxim and minim PFs (absolute values) from all four tendons during Hurricane Ivan and treat them as realizations of a single random variable (denoted generally as PF). commonly used for design.4 3. Fig.2 5.8 4. we plot the observed 3-hour PF distribution and those obtained using various Weibull fits to half-hour Paper No. (We will discuss the accuracy of this assumption in the next subsection.6 5.4 5. The design ranges are obtained based on Weibull fit to model test results.4 3.8 6. the peak distribution of a Gaussian process agrees well with observed distributions.6 4. denoted as observed MPM PF.0 2. 14. we average these observed PFs over the analysis period and find that the maxim PFs are slightly larger (3. the Ochi Factor range predicts well the observed MPM values. 15: 3-Hour Peak Factor Distribution To illustrate the bandwidth effect.0 Fig. 3. 1997) and the Ochi MPM range are also shown for reference.2 3. Again.8 2. Fig. 15 shows the observed 3-hour PF distribution.0 3. half-hour) intervals.0 T2 T3 T6 T7 Histogram Method As discussed in the Introduction.0 6.0 2004 Sep Date (day) 15.8 3.4 2. 14: Ochi Peak Factor during Ivan In order to study PF variation across storms.5 15.8 5. 12: Tendon-Averaged Observed Peak Factors during Ivan The mean values of the observed maxim or minim PFs exhibit little correlation to storm intensity.4 4. Therefore.6 2. 30% Marlin Design Range Marlin Design Range for 100-Yr to 1000-Yr Hurricanes Observed (10-Min Moving) Gaussian 25% for Winter Storms 20% % in bin Ochi MPM Range 15% 10% 5% 0% 3.2 4.0 5.5 16. 4.γ MPM = 2 ln(n) (3) where n is the number of zero up-crossings of the normalized process. 13: Peak Factor Distribution for Ivan. shown in Fig. 13. The observed peak distribution for Ivan compares very well with that of a Gaussian process.) Hence.2 3. we compare observed PF distributions for other storms together with that of an equivalent (in terms of bandwidth) Gaussian process (not shown here). 2006-JSC-356 Li 77 5 .2 2. 1997). The histogram of 30minute observed PFs is shown in Fig. 14. The predicted MPM FPs vary little throughout the storm.64 vs. Our histogram method can also be applied to 3-hour (vs. 3 to every 3-hour non-overlapping time interval for each tendon during Ivan and calculate the Ochi Factor. therefore. The peak of the observed histogram also agrees well with that of a Gaussian process. The PDF (probability density function) of Gaussian peaks is also plotted.0 4. we make the assumption that the normalized tension process is stationary throughout the storms.

Additionally. Conservatism in the design range of the MPM PF may result from application of the Weibull fit method to broadband and/or nonstationary data from model test.6 3. OMAE 2005-67083. OTC 17335. 15% 10% 5% 0% 3.8 4. RN (2005). JS and Piersol.2 3. MK (1981). respectively) in Fig. etc. VA. API RP 2T.0 5.2 4. Halkidiki. data analysis. We also applied a histogram method to the normalized process to calculate the observed PF distributions.6 5. John Wiley & Sons. G.” Proc Offshore Technology Conference. Report No. The curves labeled Weibull-95% or 50% Peaks are calculated based on Weibull fit to the subsets of 95% and 50% data. Mercier. Arlington.6 4. 29-37 Ochi. 3. Application of conventional stationary analysis techniques to nonstationary data also results in overestimation of the MPM PF. respectively. American Petroleum Institute (1997). 6. New York. 17: Nonstationarity Effect on 3-Hour Peak Factor Distribution As seen in Fig. “Marlin TLP: Measured and Predicted Responses during Hurricane Ivan. USA.0 4. AG (2000). the Ochi PF range predicts the observed MPM value well.2 4. as commonly applied to model test data. Li.2 Peak Factor Fig. The Gaussian PF distribution approximates the observed distributions very well. 2006-JSC-356 Li 78 6 .0 5.) are commonly used in their designs.6 4. USA.4 5. CONCLUSIONS Full scale tension data was collected for the Marlin TLP during Hurricane Ivan.0 6. “Recommended Practice for Planning. 25% for Winter Storms 20% % in bin Ochi MPM Range 4.” Proc Extreme Loads Response Symposium. “Random Data Analysis and Measurement Procedures. RN. Ochi.6 5.2 3.2 5. The Ochi peak factor serves as an excellent predictor of observed MPM PF. our conclusions are applicable to other TLPs because the same methodologies (model tests. pp. 5. Weibull fit method. Paper No.” Journal of Ship Research. and Perego. 10% Ochi MPM Range REFERENCES ABB (1997). March 1973. “On the Prediction of Extreme Values. Houston. Data noise was discussed and treated by a movingaverage technique. Also. “Principles of Extreme Value Statistics and their Application. Bendat.8 5. Our normalization method effectively minimizes the nonstationarity effect and has little effect on a stationary process. MK (1973). Greece.” Proc Offshore Mechanics and Arctic Engineering Conference.8 6.8 4. 5% 0% 3. Houston. The conservatism seen in the design ranges may result from application of these methods to model test data. we plot the observed distributions with and without normalization (denoted as 10-Min Moving and 3-Hour Fixed. 16: Bandwidth Effect on 3-Hour Peak Factor Distribution 30% Marlin Design Range Marlin Design Range for 100-Yr to 1000-Yr Hurricanes Although our study was focused on Marlin. BMT Scientific Marine Services provided Marlin’s Integrated Marine Monitoring System (IMMS). Banon. API. USA.” Prepared for Amoco.8 6. USA.2 Peak Factor Fig. BBN Technologies provided Marlin’s Tendon Tension Monitoring System (TTMS). We proposed a normalization method to minimize the nonstationarity effect.4 3.2 5. Perego.4 5.” second edition. The tension data are clearly nonstationary and broadband. OTC 8354. The Weibull fit method and the methods without normalization each shift the observed PF distribution to the right. ACKNOWLEDGEMENTS for Winter Storms 25% 3-Hour Fixed 10-Min Moving 20% % in bin 15% We wish to thank the BP Gulf of Mexico Deepwater Production Business Unit for funding this work.0 6.” Proc Offshore Technology Conference. et al (1997). 16. 30% Marlin Design Range Marlin Design Range for 100-Yr to 1000-Yr Hurricanes Weibull-95% Peaks Weibull-50% Peaks Observed (10-Min Moving) The conclusions from our study are 1. and Constructing Tension Leg Platforms. “Mars Tension Leg Platform – Use of Scale Model Testing in the Global Design. OH685-09252. the observed 3-hour PF distribution approximates that of a Gaussian process. “TLP Reliability Study Based on the Limit State of Tendon Unlatching.4 3. overestimates the MPM PF of a broadband process. RS.data samples in Fig. “AMOCO Marlin TLP – Phase II. The resulting effects on tension PF distribution are discussed.4 4. H. 2. 13. Global Performance Analysis. et al (2005). Designing.” third edition.8 5.6 3.0 4.4 4. 17 to illustrate the nonstationarity effect.

Sign up to vote on this title
UsefulNot useful