Paper Conference

Proceedings of BSO Conference 2020: Fifth Conference of IBPSA-England


Surrogate Optimisation of Housing Stock Retrofits using Deep Neural Networks

James Michael Hey, Peer-Olaf Siebers, Paul Nathanail, Ender Ozcan, Darren Robinson

Abstract: Surrogate modelling can greatly reduce the computational time required to perform building simulation, by trading accuracy for execution speed. We propose a surrogate optimisation extension to this process, capturing the wider optimisation loop in a second surrogate model to predict the optimisation output from runs of a conventional surrogate model; applying this to optimise of building energy retrofit strategies. In this we model the housing stock of part of Nottingham (UK), representing the c. 95,000 dwellings using a combination of Ordnance Survey and English Housing Survey data. We use established simulation and optimisation methods to create a sample of 5000 near-optimal retrofit solutions for buildings in Nottingham. Using this sample we train a set of DNN models to form a Surrogate Optimiser to predict retrofits for the remaining building stock. Using this method, a cost efficient whole house retrofit solution was found in 16.7% of the housing stock, compared with 19.2% identified by the base optimiser. The solutions identified by the optimiser scored 11% worse than those identified by the base optimiser, but the surrogate optimiser was approximately 100,000 times faster; although this improvement drops to c.20x when considering the time required to generate the training data.
Pages: 64 - 71