Optimisation of deepwater riser using bio-inspired algorithm

Abstract

The search for oil and gas in deeper water has been ongoing since the early 1990s; this search has also affected the structures used for such search and exploration and improved structure design capacity in those regions due to the severe challenges of such increased water depth. Over decades, free hanging/steel catenary riser (SCR) has been regarded as a cost-effective and straightforward system for riser solution in deepwater and ultra-deepwater by National oil companies (NOCs), International oil companies (IOCs), and independent players. However, the free-hanging SCR comes with its challenges, which can result in system failures if not correctly evaluated during the design stage. The two main challenges facing free-hanging SCR usage are increased self-weight and high fatigue damage, especially at the touchdown and hang-off points. Fatigue damage is the failure resulting from stress accumulation over time. The involvement of technologies has helped to improve the design method for these systems over the years. This research aims to optimise a deepwater riser using a bioinspired design while ensuring its structural integrity is maintained. The study notes the challenges facing the application of bioinspired design on riser optimisation in West Africa, including the non-integration of FE solver with bioinspired optimisation algorithm for the search for the global optimal solution. Also noticed is the non-inclusion of fatigue limit state criteria as a constraint in the search for the global optimum solution of the riser, as well as the non-use of coupled analysis in the conduct of riser optimisation. Another area noted is the need for more standard practice on the type of GA operator for such an analysis. The study observed the zero application of bio-inspired machine learning algorithms such as artificial neural network (ANN) on the optimisation of riser design in offshore West Africa as well as the non inclusion of fatigue limit state criteria derived from ANN prediction for the conduct of riser optimisation. The study noted the non-availability of a standard activation function and the number of hidden neurons for such an analysis. Also noted is the non-implementation of the WASP recommendation for the modelling of the wave body, which should be the case in the conduct of dynamic analysis, fatigue life/damage estimation, and the conduct of optimisation, as well as in the development of the ANN algorithm for offshore West Africa. To address these gaps, a bioinspired optimisation method is developed and implemented with GA. The developed method is positioned to infuse a finite element solver, OrcaFlex, into GA. In the instance of GA, data are simultaneously transmitted into various program sections to develop the GA fitness function and constraint. The method has positioned the optimisation algorithm GA to transmit input data into OrcaFlex, where it will run and generate response data which the optimisation algorithm GA will pick to develop the constraint or fitness function. This data exchange will continue until the program ends when it meets the predetermined termination criteria. The exchange of data between the GA and OrcaFlex is made possible through an application programming interface (API), which enables the control of OrcaFlex within the same platform, such as MATLAB, for the performance of dynamic analysis and fatigue life/damage in coupled time domain nonlinear procedure. The study taught replacing the FE solver with ANN to conduct the optimisation. On this basis, the study developed a copycat method of the earlier developed for integrating GA and FE solver OrcaFlex. However, in this instance, GA is integrated with the ANN algorithm while ensuring the fatigue limit state criteria is captured as one of its constraints for searching the riser global optimum solution. The methodologies developed herein are implemented with the programming of a GA code in MATLAB to optimise the riser design weight, during which a performance evaluation of the GA operator combination is conducted for selection, crossover and mutation are combined into twelve (12) different layouts to ascertain the most effective layout. Results from the performance study indicate that the roulette selection combined with the heuristic crossover and uniform mutation RHU produces the most minor fitness function performance at the least time. From the results, the GA search for the global optimum weight of the SCR based on the existing parameters is equal to 744.6354ton, with the optimised design variable for declination angle equal to 172.0321o , riser length equal to 3473.9212m, and wall thickness equal to 0.02435m. Integrating GA and FE solver OrcaFlex results in a 35.76 per cent weight reduction compared to the worst GA-generated fitness function and a processing time of 410,040sec. The study observed that optimisation of the riser using GA and OrcaFlex produced good results. The employed network comprises three layers: input, hidden, and output layers. The study aims to alternate the type of hidden layer activation function between logistic and hyperbolic tangent and also alternate the hidden number of neurons between 5, 10, and 15 so as to ascertain the ANN configuration that will produce the best performance. The obtained results demonstrate that in the case of effective tension, the network architecture that produced the best performance was the hyperbolic tangent activation function, with the number of neurons in the hidden layer equal to 5. The network gave the least MSE of 9.41∙10-6 with a COV equal to 0.55606 and showed a good correlation equal to 0.99995 and R-square equal to 0.99991. The results obtained in the case of the bending moment showed that the network architecture that produced the best performance was the hyperbolic tangent activation function, with the number of neurons in the hidden layer equal to 5. The network gave the least MSE of 1.8955∙10-4 with a COV equal to 0.67498 and showed a good correlation equal to 0.99881 and R-square equal to 0.99763. Accordingly, in the case of the maximum von Mises stress, the best network architecture is the logistic activation function with the number of neurons in the hidden layer equal to 10. The network gave the least MSE of 3.7895∙10-4 , with a COV equal to 0.71497 and showed a good correlation equal to 0.99767 and an R-square value equal to 0.995355. Finally, in the case of fatigue damage, the network architecture that produced the best performance is the hyperbolic tangent activation function with the number of neurons in the hidden layer equal to 15. The network gave the least MSE/performance of 2.4013∙10-4 with a COV equal to 0.56208; this shows a good correlation equal to 0.99784 and an R-square equal to 0.99569. The study compared single-output and multi-output networks to illustrate the performance deficiency attached to using a multi-output network. The results showed a performance decline of 197.24 per cent, thereby discouraging its use, as the concern here is the accurate and reliable prediction of the riser response. The result obtained from the single output network shows an average time reduction of about 96 per cent when comparing ANN to the FEA analysis. In implementing GA integrated with ANN, the study used the preferred choice of the GA operator configuration, which comprises roulette selection combined with the heuristic crossover and the uniform mutation technique. For ANN, the study used the single-output preferred choices of each of the response parameters in Chapter 8. The GA search results for the best and mean global optimal weight solutions for the SCR are found to be equal, and the best, worst, and mean scores are the same, which shows a near convergence of the searched fitness value. The obtained result indicates that the GA search for the global optimal solution illustrates the fact that the initial solution is randomly generated, from which the fittest solution is automatically ranked and deployed for the next population; this continues, as shown in the results obtained until the search for the global optimal weight solution is finally achieved. Results from the study show that the GA search for the global optimum weight of the SCR based on the existing parameters is equal to 723.23ton, with the optimised design variable for declination angle equal to 171.42o , riser length equal to 3479.4m, and wall thickness equal to 0.0235m. This results in a 37.60.10 per cent weight reduction compared to the GA-OrcaFlex worst fitness function. The study's results show an agreement in the performance evaluation of both proposed methods: the integration of GA and ANN technique and GA and FEA software OrcaFlex technique, respectively. The study confirmed the results: the integration of GA with ANN is highly commendable as it indicates a 97.76 per cent reduction in computation time when compared to using GA with an FE solver.