Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Daylight design exploration using parametric processes and Artificial Neural Networks

Lorenz, Clara-Larissa 2020. Daylight design exploration using parametric processes and Artificial Neural Networks. PhD Thesis, Cardiff University.
Item availability restricted.

[img] PDF - Accepted Post-Print Version
Download (17MB)
[img] PDF (Cardiff University Electronic Publication Form) - Supplemental Material
Restricted to Repository staff only

Download (280kB)

Abstract

The integration of Artificial Neural Networks (ANNs) as surrogates for daylight simulation models within parametric design environments promises greater computational efficiency in the exploration and optimisation of design solutions. This thesis demonstrates how ANNs can be integrated in design exploration processes, specifically focusing on the investigation of design solutions for the central atrium of a school building. ANNs are validated as surrogates for climate-based-performance metrices including Daylight Autonomy (DA) and spatial Daylight Autonomy (sDA) for thresholds of 100 lux (DA100) and 300 lux (DA300). The presented work discusses the prediction accuracies and sensitivities of the developed ANN models, the efficacy of the method, and atrium design strategies aimed at improving daylight conditions in atrium adjacent spaces. The research also critically evaluates daylight performance metrices and their implications on the design outcome of optimisation. Contributions are made in terms of validating ANN prediction accuracies for annual climate-based-daylight metrices, presenting a workflow for the selection and optimisation of input features from parametric models, and identifying limitations of ANN predictions related to model complexity and number of design variables. The work also contributes to the field of atrium design research by analysing the impact of atrium design changes on daylight performance, and by employing and comparing multiple daylight performance metrices. Thesis results showed that robust predictions could be achieved by optimising the network architecture of ANN ensembles, optimising input features, and employing cross-validation and early stopping. Overall, high accuracies were achieved for performance metrices predicting both % of occupied hours in a year and the % of space. For %time metrices, mean absolute errors were around 0.6% DA MAE (for DA ranging from 0 to 100%) for the 100 lux and 300 lux thresholds. For %space metrices, mean absolute errors were around 0.3% sDA MAE for both the 100 lux and 300 lux thresholds (for sDA ranging between 0 and 100%). Daylight simulation time was reduced by up to 71% by integrating ANNs within the design process. The design results showed that optimum atrium design solutions varied between the sDA300/50% and sDA100/50% metric. Additionally, the favorable design solutions also varied depending on whether design solutions were explored via the %space results of the sDA metric or the %time visualisations of the DA metric. Hence, this work discusses both the target thresholds employed in daylight performance metrices and bias that can be introduced by careless implementation of them. In terms of design strategy, southward orientations of the atrium well and reducing WWR towards the top floors increased daylight in atrium adjacent spaces on lower floors, but was met by a tradeoff, as this also reduced daylight on upper floors. The interdependencies of atrium design changes and the value and interpretability of the applied daylight performance metrices are further elaborated on in this thesis.

Item Type: Thesis (PhD)
Date Type: Completion
Status: Unpublished
Schools: Architecture
Uncontrolled Keywords: Daylight Design Parametric; Artificial Neural Networks
Date of First Compliant Deposit: 8 July 2021
Last Modified: 08 Jul 2021 09:40
URI: http://orca.cardiff.ac.uk/id/eprint/142272

Actions (repository staff only)

Edit Item Edit Item

Downloads

Downloads per month over past year

View more statistics