A Fast and Accurate Predictive Model for Thermal Environments Using Machine Learning

Motive: Heating and cooling systems are large contributors to energy consumption. In order to save energy and cost, the controls of

these systems must be optimised. This provides an opportunity for a

machine learning approach, which has the potential to provide fast

thermal predictions from measurements within a thermal environment. Unfortunately, state-of-the-art simulators are either too slow

or provide little resolution of the thermal environment. Furthermore,

the modelling of thermal systems is often approached differently depending on the environment or sector. This thesis aims to bridge a

gap in the thermal modelling literature between sectors, namely the

automotive and residential sectors, and introduce a common, fast

approach for modelling various thermal aspects of an environment

using machine learning algorithms.

Method: Two case studies are investigated. The first is a car cabin,

with data from 5 different experiments on the same vehicle in a controlled environment (climatic wind tunnel). The second is a house,

which provides a larger dataset with real-world observations from 18

different houses, resulting in 18 models. The data used for analysis

is split into vectors of states and controls, which are measurable variables and do not include knowledge of the structure of the space, for

example, insulation levels and geometry. A range of machine learning

approaches are applied and compared for the two case studies, making

use of both hyperparameter search and cross validation methods.

Results: The top performing models for both case studies were

found to be based on linear regression. The resulting linear regression

model for the car cabin is fast (0.007 milliseconds per predicted second), and yields good accuracy (NRMSE 0.3%) for multi-step ahead

predictions, which exceeds the performance of the traditional physicsbased model. For the house, a regularised regression (lasso) model

provides a good accuracy across the 18 house models that were built

(average NRMSE 6.8%), again providing a fast result (average 0.0014

milliseconds per predicted second). Furthermore, the models are able

to differentially predict the thermal environment in various locations

(for example, footwell vs. head for the car and kitchen vs. bedroom

for the house).

Conclusion: The resulting fast and accurate predictions based on

machine learning can be utilized to optimise thermal systems and its

controllers. Implementing a fast, accurate thermal model such as the

ones proposed in this thesis can accelerate the adoption of techniques,

such as deep reinforcement learning, for climate control in various

settings.