Numerical Optimisation of Building Thermal and Energy Performance in Hospitals

This thesis details the development and testing of a metamodel-based building optimisation methodology dubbed thermal building optimisation tool (T-BOT), designed as an information gathering framework and decision support tool rather than a design automator. Initial samples of building simulations are used to train moving least squares regression (MLSR) meta-models of the design space. A genetic algorithm (GA) is then used to optimise with the dual objectives of minimising time-averaged thermal discomfort and energy use. The optimum trade-off is presented as a Pareto front.
Adaptive coupling functionality of the building simulation program ESP-r is used to augment the dynamic thermal model (DTM) with computational fluid dynamics (CFD), allowing local evaluation of thermal comfort within rooms. Furthermore, the disconnect between simulation and optimisation induced by the metamodeling is exploited to lend flexibility to the data gathered in the initial samples. Optimisations can hence be performed for any combination of location, time period, thermal comfort criteria and design variables, from a single set of sample simulations; this was termed a “one sample many optimisations” or OSMO approach. This can present substantial time savings over a comparable direct search optimisation technique. To the author’s knowledge the OSMO approach and adaptive coupling of DTM and CFD are unique among building thermal optimisation (BTO) models.
Development and testing was focussed on hospital environments, though the method is potentially applicable to other environments. The program was tested by application to two models, one a theoretical test case and one a case study based on a real hospital building. It was found that variation in spatial location, time period and thermal comfort criteria can result in different optimum conditions, though seasonal variation had a large effect on this. Also the sample size and selection of design variables and their ranges were found to be critical to meta-model fidelity.