You are on page 1of 3

2 ARTIFICIAL INTELLIGENCE, ROBOTICS

ABSTRACTS AND ENERGY MARKETS

TRACK 2: ARTIFICIAL INTELLIGENCE, ROBOTICS


AND ENERGY MARKETS

A Machine-Learning Approach for Modelling Air-Conditioner Heating Behaviour

Ying Guo1, Mark Goldsworthy2

1Data61, CSIRO, Marsfield, New South Wales, Australia


2Energy, CSIRO, Mayfield, New South Wales, Australia

Introduction and literature review


Space heating is one of the main forms of energy consumption during winter, and understanding heating
energy consumption can lead to strategies for improving heating efficiency. For residential dwellings, reverse-
cycle air conditioners (ACs) are the most common type of heater in Australia (Your Energy Savings, 2016)
with a slightly higher penetration than gas heaters at just over one third of dwellings in 2014 (BIS Shrapnel,
2014). By building residential AC usage models, we can understand and analyse occupants’ heating behaviour,
such as predict loads for managing and designing the electricity network, understanding how certain types of
consumers’ comfort may be impacted by different changes, and eventually improving heating energy efficiency.

An earlier study of the AC heating behaviour (Goldsworthy 2017) using the Residential Building Energy
Efficiency (RBEE) dataset (Ambrose 2013, Guo 2017) revealed that occupants operate AC heating in certain
patterns, for example with increased likelihood between 7-9am and 8-10pm, and that around 65% of
operating periods are less than 2 hours in duration. In this paper we build a machine-learning model that aims
to capture these operating behaviours. The model is based on the same RBEE dataset. The output of the model
is a prediction of whether the AC is on or off over any given 30-minute interval given whole-of-house electricity
consumption and indoor temperature information for the current and preceding intervals. The approach taken
was to build one single model for all the dwellings involved, rather than the traditional approach of one model
for each dwelling (see for example Bojic 2001, Murphy 1982). The average prediction accuracy was 96% for
the test dataset with models with indoor temperature data. When the indoor temperature data is not available,
the average prediction accuracy remained above 87%.

Machine-learning approach for modelling AC heating behaviour


Modelling residential heating behaviour is challenging because space heating is related to occupant behaviour in
addition to building and weather characteristics, and there is substantial variability across all three. The approach
taken was to implement a supervised machine-learning method to allow for as much variation in the model as
possible. This also can reveal which forces are shaping household heating actions without the need to make prior
assumptions. The method that we explored is known as Support Vector Machine (SVM) (Vapnik 1998, Dehestani
2013). SVM has good generalisability, and is a stable and robust technique resistant to overtraining.

Here we trained a single SVM model for predicting AC heating actions. The sample population used to
develop and calibrate the method is the RBEE dataset, which includes the ambient temperature TA, the indoor
temperature TB, and the total dwelling energy consumption E, all with a 30-minute sample rate. It also has the
AC circuit energy consumption which is required for training the model. Data from 111 houses and covering

66
IMPROVING RESIDENTIAL ENERGY
EFFICIENCY CONFERENCE 2019

the period (June 2012 to August 2014) was used here. Data from days were maximum external temperature
was below a certain threshold was excluded to reduce the size of the dataset. The model was formulated to
predict the action for the next 30-minute interval as one of following: if AC is currently on, will the next action
be keep-on or switch-off; if AC is currently off, will the next action be keep-off or switch-on.

A previous work (Goldsworthy, 2017) identified that for the data set used in this study, the effect of ambient
temperature was markedly weak, while there was a strong relationship between the indoor temperature and the
actions of switching on and off AC. Other important factors included the local time t (within a 24-hour range),
and the duration of the on or off period τ. Considering that the internal temperature is usually not measured in
a typical dwelling, we explored how well the SVM predictor can work both with and without TB. Hence, we
chose the following combination as the feature space for SVM:
• Feature set X1: [TB, E, S, t, τ], where TB is the 2-hour prior indoor temperature, E the 2-hour prior whole-
of-home energy, S the predicted state of AC half an hour ago, t the current time of day, and τ the duration of
current keep-on/off period.
• Feature set X2: [E, S, t], where TB data is not available.

We divided the data set into test and training sets with half of the data from all the houses randomly sampled
to train the SVM and the other half data used to test the model. Cross-validation on the training dataset was
used to optimise the SVM predictor’s parameters. Results below are based on the performance of the model for
the testing data set.

Heating action prediction results


To assess the model performance we use the Accuracy and Precision Rate measures defined by Guo (2013)
and Zeng (2009). The results are summarised in Table 1. For the model using feature set X1, the accuracy on
the test dataset was higher than 93% for all four actions, and the precision rates were all higher than 80%.
As expected, there was a drop in performance for the model using feature set X2 (without indoor temperature
and duration of on/off period τ). However, the accuracy remained above 77% for all four actions while the
precision rates remained above 76%. The average performance drop was 8.7% and 8.8% respectively; that is,
the model using feature set X2 was still able to predict most heating use periods correctly.

Table 1. Prediction results for the test dataset with models with indoor temperature data (left) versus models without
indoor temperature information

Feature set X1 Feature set X2


Action Accuracy (%) Precision rate (%) Accuracy (%) Precision rate (%)
Switch-on 97.3 87.4 88.7 81.1
Switch-off 96.6 90.0 95.2 79.1
Keep-on 96.3 80.7 87.8 80.3
Keep-off 93.7 93.9 77.3 76.4
Average 96.0 88.0 87.2 79.2

Discussion and future work


There is a clear need to develop improved residential air-conditioner usage models to understand and predict
heating energy consumption in buildings. This paper presents a machine-learning approach for the residential
AC heating modelling that can help understand and predict heating behaviour in residential buildings. A
supervised machine-learning method, SVM, was deployed to predict AC actions, and achieves relatively high

67
2 ARTIFICIAL INTELLIGENCE, ROBOTICS
ABSTRACTS AND ENERGY MARKETS

accurate and precision rates. Along with quantitative data from which to build models, some other interesting
findings were the presence of two peak switch-on periods, and the much smaller variation of switch-on
probability with temperature as compared with space cooling behaviour.

Acknowledgement
This work was funded by the Department of the Environment and Energy under the Energy Use Data Model
project. Thank you to Dr Adam Berry for helpful comments and suggestions. We wish to acknowledge and
thank the Department of Climate Change and Energy Efficiency (which was abolished in March 2013 and its
energy efficiency functions transferred to the Department of Industry) for providing a proportion of the funding
for the project under which the RBEE data were collected. We wish to acknowledge and thank the following
organisations for their collaboration and substantial technical contributions to the project under which the
RBEE data were collected: Energy Makeovers Pty Ltd, Energy Monitoring Solutions Pty Ltd, and Association
of Building Sustainability Assessors.

Reference
Ambrose, M.D., James, M., Law, A., Osman, P., & White, S. (2013). The Evaluation of the 5-star energy
efficiency standard for residential buildings. CSIRO, Australia.
BIS Shrapnel. (2014). “The household appliances market in Australia 2014”, BIS Shrapnel, Sydney, Australia.
Bojic, M., Lee, M., & Yik, F. (2001). Flow and temperature outside a high-rise residential building due to heat
rejection by its air-conditioners. Energy and Buildings, 33, 737-751.
Dehestani, D., Su, S., Nguyen, H., & Guo, Y. (2013). Robust fault tolerant application for HVAC system based
on combination of online SVM and ANN black box model. Proceedings of European Control Conference,
2976 – 2981.
Goldsworthy, M., Toscas, P., Guo, Y., & Motlagh, O. (2017). Residential air-conditioner use behaviour &
building performance analysis, CSIRO Report, Newcastle.
Guo, Y., & Mahdavi, N. (2017). Machine-learning method for day classification to understand thermostatically
controlled load demand. 2017 IEEE Innovative Smart Grid Technologies - Asia (ISGT-Asia), Auckland,
New Zealand.
Guo, Y., Wall, J., Li, J., & West, S. (2013). Intelligent model based fault detection and diagnosis for HVAC
system using statistical machine-learning methods. ASHRAE 2013 Winter Conference, Dallas, USA.
Murphy, W., & Goldschmidt, V. (1982). A performance analysis of split versus package type residential air-
conditioners. ASHRAE Transactions, 88 (2), 346-358.
Vapnik, V. N. (1998). Statistical Learning Theory. New York: Wiley.
Your Energy Savings. (2016), Understanding heating and cooling, Department of the Environment and Energy,
June 2016. [Online]. Available: http://yourenergysavings.gov.au/energy/heating-cooling/understand-heating-
cooling.
Zeng, Z. (2009). Improving SVM classification with imbalance data set. Neural Information Processing,
Springer Berlin: Heidelberg.

68

You might also like