Abstract

Wind pressures on buildings and other structures, pedestrian level winds, and wind-induced dispersion of pollutants in urban locations depend, among other factors, on the velocity profile and turbulence characteristics of the upcoming wind. These, in turn, depend on the roughness and general configuration of the upstream topography. Consequently, wind standards and codes of practice typically assume simplified upstream topography conditions of homogeneous roughness or provide explicit corrections only for a limited number of specific topographies such as single hills, valleys, or escarpments. For all other more complex situations, the practitioner is referred to physical simulations in a boundary layer wind tunnel (BLWT). This thesis evaluates the effect of upstream topography on wind flow using two mathematical approaches: Computational Fluid Dynamics (CFD) and Neural Network (NN) chosen in lieu of traditional BLWT test. CFD-based numerical simulation usually consists of discretizing and solving a set of partial differential equations (the so called Reynolds-averaged Navier-Stokes equations and standard k-[varepsilon] turbulence model) describing the wind flow over a number of different topographic elements. For this purpose a robust Grid Generator, suitable for geometries characterized by curves and slopes, and a specialized CFD tool have been designed using an object-oriented approach and implemented in C++ programming language. Emphasis has been given to several numerical details and to the incorporation of influential parameters such as ground roughness. The associated accuracy of the proposed CFD tool has been quantified and validated against the results of the extensive review carried out. Despite continuous progress in hardware/software technology resulting in fewer resource requirements, the practical utilization of CFD-based numerical simulations to predict design wind load parameters is rather limited. To address this issue, a new combined CFD-NN approach in which the NN model is trained with CFD-generated data has been proposed and developed. Since the NN approach is defined on the basis of connections between system state variables (input, internal and output variables) with only limited knowledge of the "physical" behavior of the system, output variable values (speed-up ratio in the present case) are produced following input of simple geometrical parameters describing the topography and roughness of the ground. In this combined approach, the domain expert first carries out the complex numerical simulations and validations, then trains the NN model and makes it available for use by the end user who avoids therefore dealing with complex numerical simulations. (Abstract shortened by UMI.)