Paper Conference

Proceedings of uSim Conference 2018: 1st uSim Conference of IBPSA-Scotland

     

PREDICTING LOCAL WIND PRESSURE COEFFICIENTS FOR OBSTRUCTED BUILDINGS USING MACHINE LEARNING TECHNIQUES

Ioanna Vrachimi, Daniel Costola

Abstract: Wind pressure coefficients (Cp) are important elements in the simulation of natural ventilation in urban environments, where it tends to be less effective. Cp values can be obtained by wind tunnel experiments or CFD simulations, but these methods are not always available to building simulators due to cost and time constraints. Cp values also can be obtained by inexpensive methods, such as databases and analytical models, but these values are usually surface-averaged and introduce major errors in the calculation. This paper reports early results on the use of machine learning techniques to derive more accurate models for obstructed buildings with the potential use for urban stock modelling. Artificial neural networks (ANN) were applied to the empirical data from wind tunnel experiment in order to predict local (non surface-averaged) values of Cp. The cases used were obstructed, flat-roofed buildings with different area density values and surrounding buildings’ height. One ANN was developed per wind attack angle using the statistical package R and consists of 5 inputs, three hidden layers and the output. Results obtained indicate than an ANN can predict the local Cp in obstructed buildings with uncertainty of ± 0.05 for a confidence level of 95%. This paper demonstrates promising results in the use of machine learning techniques to model complex input required by urban building performance simulation. Cp values by ANN show major improvements when compared to current practice sources.
Pages: 22 - 29
Paper:
usim2018_013