Accelerated optimisation methods for low-carbon building design
Date
Authors
Advisors
Journal Title
Journal ISSN
ISSN
DOI
Volume Title
Publisher
Type
Peer reviewed
Abstract
This thesis presents an analysis of the performance of optimisation using Kriging surrogate models on low-carbon building design problems. Their performance is compared with established genetic algorithms operating without a surrogate on a range of different types of building-design problems. The advantages and disadvantages of a Kriging approach, and their particular relevance to low-carbon building design optimisation, are tested and discussed. Scenarios in which Kriging methods are most likely to be of use, and scenarios where, conversely, they may be dis- advantageous compared to other methods for reducing the computational cost of optimisation, such as parallel computing, are highlighted. Kriging is shown to be able, in some cases, to find designs of comparable performance in fewer main-model evaluations than a stand-alone genetic algorithm method. However, this improvement is not robust, and in several cases Kriging required many more main-model evaluations to find comparable designs, especially in the case of design problems with discrete variables, which are common in low-carbon building design. Furthermore, limitations regarding the extent to which Kriging optimisa- tions can be accelerated using parallel computing resources mean that, even in the scenarios in which Kriging showed the greatest advantage, a stand-alone genetic algorithm implemented in parallel would be likely to find comparable designs more quickly. In light of this it is recommended that, for most lowcarbon building design problems, a stand-alone genetic algorithm is the most suitable optimisation method. Two novel methods are developed to improve the performance of optimisation algorithms on low-carbon building design problems. The first takes advantage of variables whose impact can be quickly calculated without re-running an expensive dynamic simulation, in order to dramatically increase the number of designs that can be explored within a given computing budget. The second takes advantage of objectives that can be !Keywords To Be Included For Additional Search Power: Optimisation, optimization, Kriging, meta-models, metamodels, low-energy design ! "2 calculated without a dynamic simulation in order to filter out designs that do not meet constraints in those objectives and focus the use of computationally expensive dynamic simulations on feasible designs. Both of these methods show significant improvement over standard methods in terms of the quality of designs found within a given dynamic-simulation budget.