Constellation, le dépôt institutionnel de l'Université du Québec à Chicoutimi

Résumé

Accurate and efficient modeling of the wind field is critical to effective mitigation of losses due to the tropical cyclone-related hazards. To this end, a knowledge-enhanced deep learning algorithm was developed in this study to simulate the wind field inside tropical cyclone boundary-layer. More specifically, the machine-readable knowledge in terms of both physics-based equations and/or semi-empirical formulas was leveraged to enhance the regularization mechanism during the training of deep networks for dynamics of tropical cyclone boundary-layer winds. To comprehensively appreciate the high effectiveness of knowledge-enhanced deep learning to capture the complex dynamics using small datasets, two nonlinear flow systems governed respectively by 1D and 2D Navier-Stokes equations were first revisited. Then, a knowledge-enhanced deep network was developed to simulate tropical cyclone boundary-layer winds using the storm parameters (e.g., spatial coordinates, storm size and intensity) as inputs. The reduced 3D Navier-Stokes equations based on several state-of-the-art semi-empirical formulas were employed in the construction of deep networks. Due to the effective utilization of the prior knowledge on the tropical cyclone boundary-layer winds, only a relatively small number of training datasets (either from field measurements or high-fidelity numerical simulations) are needed. With the trained knowledge-enhanced deep network, it has been demonstrated that the boundary-layer winds associated with various tropical cyclones can be accurately and efficiently predicted.