CS32B - Concurrent Session 32B: Performance-based design of existing and future reservoir structures: An application of computational fluid dynamics modelling and machine learning techniques

Traditionally, structural design of reservoir structures, such as wave walls and crown walls, is performed by estimating return periods of the reservoir wave climate, isolating specific events of interest (say a 200-year wave event), and performing calculations of hydraulic and structural loads (forces and overtopping) by using simple empirical methods or more sophisticated methods, such as computational fluid dynamics (CFD) and/or physical modelling.

This traditional approach to testing and confirming the design has been widely utilized; however, it has several drawbacks. Perhaps most important, setting the return period at the level of wave climate does not mean that the wave loadings or overtopping rates are subject to the same probability of occurrence: i.e., a 200-year-return-period wave event does not necessarily lead to a 200-year-return-period overtopping event. In addition, it is likely that a return period estimated today may be revised in the future due to climate change and/or collection of additional data. Tests or calculations that are based on isolated events may have limited future value; it is likely that these tests will have to be repeated, adding to the cost and duration of future engineering studies.

This paper proposes a methodology to overcome these challenges. A workflow of CFD modelling and machine learning algorithms is used to estimate return periods for each individual hydraulic or structural load component by simulating more than 100 test cases and using these to train a machine learning application. The simulation output is presented in terms of return periods of wave overtopping and forces for various scenarios of relative sea level rise. The creation of the emulator also has the added benefit that the model can be used as a real-time forecasting system for overtopping or wave loading.

The selection of the CFD model to generate the simulations database is preferred, due to the low cost of CFD compared to physical modelling and the relatively high fidelity of these models, which, when implemented carefully, may yield results of equivalent value to physical models.