Abstract
This study explores the potential of mesoscale weather forecasting models, such as WRF output, characterized by high spatiotemporal resolution, to enhance the precision of initial and boundary conditions for microscale simulation model development, particularly computational fluid dynamics (CFD) models. Conventional numerical weather prediction methods face challenges in maintaining accuracy and managing escalating computational costs, particularly in proximity to urban boundaries. To address these challenges, this paper proposes a fusion of model reduction techniques and machine learning-based methods, including Artificial Neural Networks (ANN) and Recurrent Neural Networks (RNN). The research illustrates that complex flow field patterns can be effectively identified using ANNs and RNNs. Moreover, it introduces a Reduced-Order Model (ROM) as a promising approach to generating low-dimensional surrogate models having high accuracy and computational efficiency. The presented ROM algorithms, leveraging RNNs and ANNs, are modeled specifically for higher spatiotemporal resolution simulations, highlighting their capability to maintain accuracy compared to reference solutions.
