Keywords

29.1 Introduction

Under the Paris Agreement, the UK government has committed to net-zero carbon emissions by 2050. But steps need to be taken in order to reach this target [1, 2]. The Committee on climate Change (CCC), an advisory board to the government of the UK suggests that although UK is progressing well, further steps will need to be taken at the earliest for this target to be met. According to CCC, of the UK’s greenhouse-gas emissions, the residential housing accounts for about 14% [3]. The net zero strategy updated in October 2021 mentions insulation only once [4]. The government also scrapped its Green Homes Grant scheme [5]. One reason for this might be because of the data from the latest English Housing Survey; according to which in 2019, 7% of residents dwelling in residential homes reported that at least one part of their home got uncomfortably hot. Of these, 11% lived in homes built in 2003 or later [6]. This overheating invariably leads to window opening behaviour which in turn contributes to the stochastic nature of energy usage in buildings.

The International Energy Agency (IEA) is an organisation of 26 countries around the world, that came together to shape energy policies for a secure and sustainable future. The IEA project Energy in Buildings and Communities (EBC) is a collaborative research and development project among the member countries. Annex 53 of EBC, employed an interdisciplinary approach, integrating building science, architectural engineering, computer modelling and simulation, and social and behavioural science to develop and apply methods to analyse and evaluate the real energy use in buildings considering the six influencing factors namely climate, building envelope, building services and energy systems, building operation and maintenance, occupants’ activities and behaviour, and indoor environmental quality and found that better prediction of building and energy- related behaviour may result in benefits for energy savings, cost saving and better thermal comfort of occupants [7]. Pilkington et al. [8] examined how occupant behaviour affected the energy efficiency of six passive solar homes with sunrooms and discovered that space heating demand varied by a factor of 14 between homes, with evidence suggesting that it may vary by a factor of up to 45. The energy demand varied with different locations, ranging from 96 to 171 kWh/m2/year, according to Jang and Kang's study of individual apartments in a high-rise building. They created a model integrating and implementing the unit specific consumption variation [9]. When Van den Brom et al. [10] compared the average energy consumption of 1.4 million Dutch households living in social housing with theoretical values, one of their conclusions was that rehabilitated buildings did not perform as well in practise as predicted. The study by Yan et al. [11] suggests that occupant behaviour significantly affects the energy performance of buildings and emphasises the need to include OB representation in building simulation models for more accurate findings. Delzendeh et al. [12] studied the literature to find gaps in the evaluation of energy demand and its utilisation in buildings and observed an alarming EPG, with a variance of up to 300%. In their analysis of the existence and size of the energy gap in residential buildings in Switzerland, Cozza et al. [13] found that while buildings with high energy ratings (A and B) consumed more energy than expected, those with low energy ratings (Energy ratings C to G) consumed less energy than expected.

Accurate evaluation of energy-related occupant’s behaviour is a key factor for bridging the gap between predicted and actual energy performance of buildings [14]. One of the key energy-related human behaviours is window control behaviour that has been modelled by different probabilistic modelling approaches. Identifying the factors that impact residential energy consumption is a key factor to be considered when designing of models and in the implementation of policies for energy efficiency. To analyse this, both the contextual and behavioural factors should be considered. Contextual factors include the local climate, building characteristics; and the behavioural factors include the user demography, their behavioural differences and energy usage pattern. Esmaeilimoakher et al. [15] conducted a survey which collected information about several building and occupant related factors, including floor area, household size, household income in Perth, West Australia. Perth is warm throughout the year with temperatures ranging in between 15 and 30 °C during most of the year. The survey showed that floor area, household size, and income were significant factors affecting energy consumption, rather than window opening behaviour. This might be because the survey did not include any temperature data or windows opening behavioural data, and because it is a warm country where heating was not a factor contributing to high energy consumption. Since the last decade, building energy performance simulation models have been seen to consider occupant behaviour by including stochastic models of occupant behaviour in relation to energy efficiency of buildings [16,17,18,19]. However, validation of these models has been sporadic. Developed models must be validated with similar but not the same data, and the results analysed to understand the usability of a model. Haldi et al. [16] compared simulated model values to actual values to get an estimate of the forecast realism. They proposed a new procedure called ‘validation by simulating’ wherein the combined predictive accuracy of two existing behavioural models of window opening and thermostat adjustments were estimated and compared to actual values taken in sensors fitted in apartments in Copenhagen, Denmark. Data were collected for two months and compared with data from a simulated model of the same building. Building energy performance simulations (BEPS) IDA ICE tool was used for simulation. It was found that although the predicted and actual values were in the same range, the model was unable to predict the actual indoor environmental conditions, which meant the model needed to be improved. For the above discussion, further research is still needed in relation to windows opening behaviour and the effect on energy consumption and the behavioural prediction using artificial intelligence.

29.2 Methodology

To monitor windows opening behaviour of occupants and to understand its relation to house heating and thereby energy efficiency, data were collected in peak winter of 2020, from end of Nov/20 to March/21 in social housing in Nottingham. This report analyses data collected in one of the three considered houses. Data in this house was collected between Nov 27 and Dec 17, 2020.

Sensors are fitted near windows in the dining room, the radiator in the dining room and in the hall to monitor the downstairs window temperature and humidity, downstair radiator temperature and downstairs room temperature and humidity respectively. In total, six sensors were fitted, and data samples were collected every minute, see Fig. 29.1. The sensors used were temperature and humidity data loggers Omega OM-CP-RATemp101A. Temperature is measured in degree Celsius. For this study, the humidity measurement is not considered.

Fig. 29.1
Two layouts of a house indicate the locations of the sensors. Sensors include the downstairs and upstairs radiator temperature, downstairs and upstairs window temperature, and downstairs and upstairs room temperature.

House 1 layout with location of sensors

Full size image

House A is an end-of-terrace house, floor plan shown in Fig. 29.1. It has two floors, downstairs consisting of the kitchen, a w/c and an open dining/sitting room layout, and the upstairs consists of a landing, three bedrooms and a bathroom. The external wall was insulated 8 years ago and since then the residents feel considerable improvement in the heat retention property of the house. However, they tend to open the windows more often, for air circulation, even though they have vents, which are kept open all the time.

This paper focuses on the analysis of data from the selected house, which is used to develop a model which will help to predict the windows opening status and hence the factors influencing such decisions. As shown in Fig. 29.2, a block diagram presents the developed Artificial Intelligence models to predict the windows status from inputs based on outside ambient temperature, time of the day, radiator temperature and room temperature.

Fig. 29.2
A block diagram presents the proposed model to predict the downstairs windows' status from the inputs based on the outside ambient temperature, time of the day, radiator temperature, and room temperature.

The proposed model to predict windows status from other variables

Full size image

Selecting a model involves considering all factors to be considered, and a trade-off between specific characteristics of the algorithm like speed, memory usage, transparency etc. The basic structure of a decision tree consists of three basic parts called root node containing data, internal nodes (branches), and end nodes (leaves). The fundamental idea behind creating a decision tree structure may be summed up as asking a series of questions about the data and acting on the results as quickly as possible by using the attribute information of the training data. In this manner, the decision tree gathers the responses to the inquiries and develops the guidelines for making decisions. To classify the data and create the tree structure, questions are first asked of the root node, the tree's initial node. A coarse tree is where the classification is broad, and the branching is minimum. A fine tree is one where the main node or root branches to form several nodes, based on the classification criteria for the problem in hand [20, 21].

The study aims to understand the window opening behaviour of occupants; to find the conditions under which window is opened. Measurements of window opening and room temperature, radiator temperature, time of day and outside ambient temperature are taken. In the proposed mode, window status of the downstairs room is the ‘response’ and the outside ambient temperature, time of the day, downstairs radiator temperature and downstairs room temperature are the ‘predictors’. Based on 1 day of data (selected randomly) taken for training and testing, ‘Fine decision tree’ was found to provide the best solution with 99% accuracy. Hence the fine tree is chosen to test the rest of the days data. Support Vector Machine, Logistic Regression, Gaussian Naive Bayes are some of the other models tested on the Matlab classification Learner app.

The percentage error of the model is calculated. The total error is the number of times the window status (open or closed) is wrongly predicted (see Eq. 29.1). Therefore, the error is the sum of inequality between the two logical arrays, as shown in equation. The percentage error is the total error over the total number of observations (Eq. 29.1). In this study, data is collected every minute giving observations of 1440 per day.

$$Error = \left( {predicted\;window\;status\sim = measured \, window \, status} \right)$$
(29.1)
$$Percentage\;error = \frac{{\mathop \sum \nolimits_{i = 0}^{n} Error}}{n} \times 100$$
(29.2)

where n = 1440.

The average error for the Model A1 to Model A15 is calculated as shown in Eq. (29.3)

$$Average\;error\;Model\;A_{i} = \frac{{{\text{Model}}\;A_{i} \;day_{i + 1} \;{\text{error}} + \cdots + {\text{Model}}\;A_{i} \;day_{n} \;{\text{error}}}}{n - i}$$
(29.3)

where I is the number of days taken for training and n is the total number of days (n = 1440 in this case).

29.3 Results and Discussion

Figure 29.3 presents the data of outside ambient temperature, downstairs radiator temperature and room ambient temperature, upstairs radiator and room ambient temperature, window temperature, (measured using the sensor placed as shown in Fig. 29.4) between 1st December and 8th December 2020, as an example of the data captured. The other factor considered in this study is the outside ambient temperature. A daily average value of the nearest met weather station nearest to the location is obtained [22].

Fig. 29.3
6 line graphs of temperature recorded from various sensors versus date and time. All have a fluctuating trend. The graphs of each sensor present the weekly average temperature of 3.2, 35, 20.6, 16.2, 17.2, and 12.3 degree Celsius, respectively.

House 1 sensor values recorded from December 1 to December 8

Full size image
Fig. 29.4
15 line graphs represent the window open status using the fine tree model A 1 for days 2 to 16. They have 2 lines for real and predicted values in a square wave pattern and the lines overlap with each other.

Model A1 testing—actual values and predicted values from day 2 to day 16

Full size image

It should be noted that this study focuses on the temperature values and window status and does not include other factors that might influence window opening behaviour like ventilation type, noise level, security issues etc. Window temperature is the raw sensor data obtained to indicate the window status. The window temperature from the sensor is converted into binary values 0 indicating closed widow and 1 indicating open window (the degree of opening is not considered in this study). This is done taking into consideration, the window temperature, outside ambient temperature, room temperature and radiator temperature. The degree of opening of window is not considered in this study.

Based on the Block diagram of the proposed machine learning model shown in Fig. 29.2, Fine Tree algorithm is used to develop different training and testing models.

Models A1 to A15 are developed based on the training process presented in Table 29.1. In Model A1 the available data of House 1 (16 days data) is split as one day data set for training and the other 15 days data for testing the model. Hence Model Ax, is the Fine Tree algorithm training, where x is the number of training days, and the number of testing days will be (16 − x). In this study, weekends and weekdays were not considered separately, since the occupants of the house were not working and did not have a different pattern of living for weekends/weekdays.

Table 29.1 Models trained and tested as part of iteration process
Full size table

Figure 29.4 shows the graph of actual values and predicted values from day 2 to day 16 for Fine Tree Model A1 (training using day 1 and testing using days 2 to 16). In Model A1, one day data is taken for training and the remaining 15 days data are used for testing the model. In the same way the process is repeated by taking two days data for training and 14 days data for testing the model (Model A2); 3 days data for training and 13 days data for testing (Model A3) and so on. The percentage error between the predicted and actual window opening status is calculated using (1) and (2).

The predicted window status is calculated using the machine learning models. Model A1 uses one day data to train and day 2 to day 16 data is used to test the performance of the developed model. The results and the difference between the actual and predicted data is shown in Fig. 29.5.

Fig. 29.5
15 bar graphs represent the percentage error versus time in days for Fine Tree models A 1 to A 15. They plot the bars for days taken for training and days tested with the trained model. The error percentage is 0 as the days tested with the trained model increase from A 1 to A 15.

Percentage error for Model A1 to Model A15

Full size image

The testing data is used to validate the developed model. This is done by testing the data with the developed model and comparing the result with the actual measured value. Model A2 uses 2 days of data to train the model and the developed model is tested with the rest of the days (day 3 to day 16), and so on. Figure 29.6 presents the training data and the error levels for the test models, Models A1 to A15.

Fig. 29.6
A bar graph of percentage error versus Fine tree models A 1 to A 15. The respective error values are as follows, 24, 37, 40, 43, 47, 49, 53, 31, 33, 33, 40, 41, 37, 36, and 30.

Average percentage error for Model A1 to Model A15 (rounded to the nearest integer)

Full size image

Model A1 is found to have the lowest percentage prediction error of 23.8%. In this case, the proposed model of using temperature of room, radiator, external temperature, and time is found useful in predicting the reason for opening or closing of windows with 76.2% accuracy.

29.4 Conclusion

Temperature data were collected from one house in Nottingham. Data were collected in peak winter of 2020, from end of Nov 2020 to March 2021. Sensors were fitted to measure radiator temperature, room temperature, room humidity and window temperature for both upstairs and downstairs. Outdoor ambient temperature for the same period was also collected. Stochastic models were implemented, using machine learning of Fine Tree model, with datasets of window opening behaviour as response. The inputs to the model were the radiator temperature, room temperature, outside ambient temperature, and time. Results  have shown a reasonable prediction capability of windows opening behaviour, as the output of the model . Results have shown that Fine Tree model with 1 day data taken for training had the smallest percentage error of 23.8%. Further investigation needs to be undertaken to explore whether the presence of humidity plays a role in the window opening behaviour of occupants. This is a pilot study to investigate the potential applicability of classification models to find the relationship between energy efficiency of a building and windows opening behaviour of house occupants. Findings from this investigation can be used to identify the factors that contribute to ‘performance gap’ in energy efficiency of a building.