Examining the performance change of inverse surrogate models with building energy model time series data
Date
2024
Authors
Jowett-Lockwood, Liam
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
A building Surrogate Model (SM) is a Machine Learning (ML) model trained to reproduce the outputs of a building energy model at a much smaller computational cost. While a SM will traditionally accept Building Energy Modelling (BEM) parameters for its inputs to predict BEM outputs, a building Inverse Surrogate Model (ISM) suggests doing the opposite. Inverse modelling provides potential in determining unknown building thermal characteristics of existing structures. The task of deriving inputs from outputs is more difficult as multiple input combinations can result in the same output, thereby necessitating the need for comprehensive outputs allowing for more information to be extracted. With the rise of deep learning models and methods, ML practitioners have a greater array of tools available to handle increasingly complex tasks. This has enabled ISMs with a stronger opportunity to excel in parameter prediction.
The papers in this thesis focus on the ability of the ISMs to accurately predict parameter values. The first paper (Chaper 2) examines prediction performance of an ISM with synthetic data from a BEM model based on a single-family home. Performance changes were investigated when data was decreased by reducing the amount of time series provided, the duration of time series, or both. The second paper (Chapter 3) primarily focused on the generalizability of ISMs to be applied for multiple projects without having to retrain on new data each time. Several different ISM models were tested with predicting parameters for different BEM building shapes with varied geometry in addition to multiple locations.
The key finding of this research is that there is potential for ISMs to be used with building data. While all data used in this thesis was synthetic data generated from BEM simulation runs, ISMs were shown to not only successfully predict some parameters, but also hold solid degree of generalizability depending on the ML model used. If ISMs can successfully predict characteristics of an actual building, then it allows for new approaches for applications such as retrofit planning.
Description
Keywords
Machine Learning, Building Energy Modelling, Surrogate Modelling, Inverse Modelling