A Review of the Troll A Model Test A PRELIMINARY SUMMARY Classification: Statoil Internal Status: Draft A Review of the Troll A Model Test A PRELIMINARY SUMMARY 1. Why did we do a model test? 2. Why did we decide to go for a very complex model of the platform? 3. The basin model 4. Test plans – brief review? 5. Some important dynamic properties of Troll A! 6. Preliminary results of test program? 7. Effect of currents. 8. Full scale data 9. Uncertainties Sverre Haver, Statoil, Stavanger, February 2008

1. Why did we do a model test? Ringing was discovered as detail design was done for Troll A. NC initiated a lot of activities in order to establish reliable loads accounting for this new phenomenon. It is impressing how fast they were able to present adequate conclusions. A model test was performed in Denmark. The test did clearly show that ringing should be accounted for. Major ringing events were observed, but full weight was not given to the most severe events due to a somewhat questionable shape of the most extreme waves. A computer program, Nirwana, was ”calibrated” to the model test results. This program was used for obtaining characteristic loads for design. The program handle drag-type non-linear loading and a non-Gaussian sea surface (approximately). However, it is not obvious that the program captures the load mechanism causing ringing. In May 1996, we did measure 50% of the 100-year load in one of the shafts for a sea state of 5-6m significant wave height. We were not able to explain this event fully.

1. Why did we do a model test? Were we originally concerned about the May 1996 episode? To some extent – we were – in particular since we were not able to explain it. However, since the as-installed natural period was much less (4s) than the period used in design (5.5s), we did not see a reason for major concern. A reduction of the natural period will typically reduce dynamics caused by linear mechanisms. We expected this to compensate for a possible underprediction of the ringing response. Were we concerned autumn 2006 when we did hear about the major planned weight increase for top side weight? Some of us became very concerned! Because this weight increase was so large that it would bring the natural period up to the natural period used in design. If this happen, the linearly forced dynamics will increase to the level used in design. A critical question thus became: How accurate was the original prediction of the ringinging response (non-linear forced dynamics)?

1. Why did we do a model test? We are still not able to handle ringing by numerical calculations?  The only robust approach is to carry out an extensive and well planned model test! Marin and Marintek were invited to bid. Marintek was chosen. Marintek has done an outstanding job! In addition, throughout both the planning of the test, the introductory part of the test (model building) and the execution of the test program in the basin, people from AkerKværner have participated actively. The good and close cooperation betweeen all involved was crucial for the quality we seem to have achieved.

2. Why did we decide to build such a complex model? The uncertainty is related to the wave induced loads – in particular – their detail behaviour in time and space close to the instantaneous surface. We had two options: Measure load density (force pr length unit) over the whole structure and then calculate the corresponding response by numerical methods. or Measure the response in the platform with sufficient accuracy for some selected platform intersections. From these measurements we could directly estimate 10**(-2) – annual probability (100-year) values and compare with original values. This requires a flexible model with same bending stiffness, axial stiffness and shear stiffness (scaled properly) as the prototype standing at Troll field. The last option was chosen!

3. The Basin Model Did we succeed to match protype properties by the basin model The model building was an enormous challenge for Marintek! In parallel with the model builing, a complete finite element model of the basin model had to be developed. This for the purpose of having something to control the physical basin model properties against. Various load cases were introduced to the finite element model. The same loads were thereafter introdused to the physical model and the results in terms of displacements and bending moments were compared. From my point of view, the basin model became of very high quality. With an element of luck – the model went from a very good model to an outstanding model! The model is equipped with about 200 channels of measurements. During the 4-5 week testing period only one – 1- channel fell out – impressing!!

3. The basin model Laboratory arrangement Drawing: Marintek

3. The Basin Model Documentation of bending stiffness Photo: Marintek

3. The Basin model Force transfer from outer shell to inner core of shaft Photo: Marintek

3. The Basin model Gluing of strain gauges Photo: Marintek

3. The Basin model Ballasting shafts with divinycell Photo: Marintek

3. The Basin model Mounting outer shell Photo: Marintek

3. The Basin model Riegel Photo: Marintek

3. The Basin model Caisson installed on foundation springs Photo: Marintek

3. The Basin model Platform installed in the basin Photo: Marintek

3. The Basin model Deck model Photo: Marintek

3. The Basin model A subsea view Photo: Marintek

3. The Basin model The platform in waves from 90o Photo: Marintek

3. The Basin model The platform in waves from 67.5o Photo: Marintek

4 Test plans – a brief review Deck acceleration 0o 67.5o M_A2 M_E2 Chosen wave directions 90o M_C2 Quantities for preliminary assessment of condition BS OTM

4. Test plans – a brief review? We are (acc. to Norwegian Regulations) going to estimate loads with an annual exceedance probability of 10-2 (ULS-control). This requires, in principle, a long term response analysis: For a very non-linear response problem, it is quite some challenge to find the short term distribution. For a ringing problem we have to do this my model testing and we can not test all important sea states.  We have to select an approximate method. We went for the environmental contour line method.

4. Test plans – a brief review. Contour method: Worst sea state for Troll A response Identify wort sea state along the 10-2 – annual probability contour. Generate a sufficient number of realizations of the 3-hour extreme value for the worst sea state. (Here 24 3-hour realizations of worst sea state were tested.) Fit a probability model to the observed 3-hour extremes. Estimate 90-percentile. Adopt this value as an estimate for 10-2 – response. (This is why we have used such a high number (24) of realizations.)

Some important dynamic properties of Troll A Some important dynamic properties of Troll A! Transfer functions for 90o og 67.5o obtained by: i) WADAM (quasi-static), ii) Estimated from model test (dynamic) , AkerKværner. Note difference around the natural frequency between waves from 67.5o and 90o.

Preliminary results of test program? The basin test went very smooth. The physical model has functioned according to specification. We have changed deck mass, changed soil-structure stiffness and changed direction. But as we came back to an earlier tested conditions, natural period estimates (5 largest periods) have been very well reproduced. We have performed the following tests (in total about 250 tests): 67.5grader: Base case (M1+K1): 24 3-hour tests Case 1(M2+K1): 24 3-hour tests Case 2 (M2+K2): 24 3-hour tests 90 grader: Base case (M1+K1): 24 3-hour tests Case 1 (M2+K1): 24 3-hour tests Case 2 (M2+K2): 24 3-hour tests Case 3 (M1 + K2): 24 3-hour tests Case 4 (M2+K3): 24 3-hour tests (Case 3: 20 3-hour 10**(-4) tests)) Case 4 : 2 3-hour tests with current 0, 67.5 and 90 degrees: Screening tests for various sea states.

6. Preliminary results of test program? Base shear – an extreme case

6. Preliminary results of test program 6. Preliminary results of test program? Resonance response in first sway mode

6. Preliminary results of test program 6. Preliminary results of test program? A challenging problem – extremes of composed processes M1+K1 M2+K2

6. Preliminary results of test program 6. Preliminary results of test program? Estimated distribution function for 3-hour extreme response Extreme response: Here it is taken as the largest of the 3-hour largest maximum and the absolute value of the 3-hour smallest minimum.

6. Preliminary results of test program 6. Preliminary results of test program? Comparison between results for 0o, 67.5o and 90o

6. Preliminary results of test program. Preliminar estimated 10 6. Preliminary results of test program? Preliminar estimated 10**(-2) annual probability values using the 90-percentile Green: No problem Yellow: Further work must be done Red: May be difficult to withstand with proper margin

6. Preliminary results of test program. Preliminar estimated 10 6. Preliminary results of test program? Preliminar estimated 10**(-2) annual probability values using the 85-percentile

6. Preliminary results of test program. Preliminar estimated 10 6. Preliminary results of test program? Preliminar estimated 10**(-2) annual probability values using the 95-percentile

7. Effect of current No current With current

7. Effect of current No current With current

7. Effect of current No current With current

8. Fullscale measurements

9. Uncertainties Adequacy of model test sea states. (Can be slightly conservative or slightly non-conservative.) Long crested versus short crested sea. (Most like a conservative effect.) Model test sea states may be low regarding spectral density for a period band around 5-8s. (Non conservative, but not necessarily important.) Choice of percentile – 90% or 95% or something else? (Can be conservative, most likely slightly non-conservative.) Effects of current. (Conservative or non-conservative? Most probably neutral.) My best guess: Estimates in summary report can be associated with uncertainties of +/- 15%.

9.1 Uncertainties in empirical distribution from 24 obs Values of true model: 0.90 (2.25) = 3330MNm 0.95 (2.97) = 3617MNm From simulations: 90% band of 0.9-value: 3150 – 3650MNm 90% band of 0.95-value: 3350 – 4250MNm Fitting Gumbel distributions to the data will narrow the uncertainty bands somewhat.