-Thanks to Building Energy and Comfort Managements (BECM) systems, it is possible monitor and control buildings with the aim to ease appliance management and at the same time ensuring efficient use of them from the energetic point of view. To develop such kind of systems, it is necessary to monitor users' habits, learning their preferences and predicting their sequences of performed activities and appliance usage during the day. To this aim, in this paper a system capable of controlling home appliances according to user preferences and trying to reduce energy consumption is proposed. The main objective of the system is to learn users' daily behaviour and to be able to predict their future activities basing on statistical data about the activities they usually perform. The system can then execute a scheduling algorithm of the appliances based on the expected energy consumption and user annoyance related with shifting the appliance starting time from their preferred one. Experimental results demonstrate that thanks to the scheduling algorithm energy cost can be reduced of 50.43% and 49.2% depending on different tariffs, just by shifting the use of the appliance to time periods of non-peak hours. Scheduling based on probability evaluation of preferred time of usage of the appliances allows to still obtain evident energy savings even considering the errors on predicted activities. Index Terms-Activity Recognition; Activity Prediction; Energy Management; Comfort Management; Smart Building
Smart buildings are characterised by the presence of sensors,
actuators and smart devices that give the opportunity to
monitor and control, either manually or automatically, key
equipment within buildings [
In such an intelligent scenario, one of the major goals is to
provide users with tools that support cost-effective solutions
to appliance management, which: i) demand the lowest effort
in terms of training and management, dynamically adapting to
user requirements, and ii) take into account user habits so that
appliance management decisions do not conflict with them,
causing a disaffection that may lead the user to turn off the
system [
Currently, most of the literature considers user comfort as
a set of hard constraints on appliance usage, which are a
priori set considering general statistics [
In this paper, user preferences and habits about appliance
usage are continuously monitored, recognised and predicted,
by means of a BECM system based on sensors deployed
inside the reference building. The system merges two
previous studies about activity recognition [
The main contributions provided by this paper can be
summarised as follows:
an activity recognition algorithm used to model user
profiles, which was first proposed in [
To the best of the authors’ knowledge, this is the first comprehensive system to use sensor-based activity prediction and occupants’ preference inference, and integrate them into a BECM. Accordingly, based on simulations of the system on a real dataset, this paper further analyses how the proposed system affects energy-related costs.
The remainder of the paper is organised as follows. Section II presents past works and the required background. In Section III an overview of the proposed system model is provided.
Section IV describes the reference use case considered to test the performance of the system. Finally, in Section V a performance analysis is provided. Conclusions and final remarks are drawn in Section VI.
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
II. RELATED WORKS
Smart technologies can be used in all kinds of different buildings (i.e., residential, office, and retail sectors) to improve the comfort and the safety of people in their home, concerning various topics, from healthcare and providing living assistance, to environmental monitoring and ensuring energy saving.
Accordingly, BECM systems have the objective of combining
power consumption minimisation while preserving user
comfort [
It is evident that user preferences and habits severely affect results of BECM systems. For this reason, in recent years researchers have started to observe users’ behaviour, in order to infer their habits and preferences. The monitoring of activities of people in their home can be done by analysing data that can be gathered with different technologies. Different studies proposed solutions based on using cameras and wearable sensors or gathering data provided by phone accelerometer and gyroscope [15][16]. These solutions are not very practical in home scenarios where people are often not inclined to accept those devices. To monitor what activities people are performing in their house, non-intrusive sensors are often preferred: typical devices that are installed in the environment are motion sensors, door sensors or temperature and pressure sensors [17]. The data collected from sensors inside resident houses are analysed using data mining and machine learning techniques to build activity models that are used as the basis of behavioural activity recognition.
With reference to modelling and classification methods, researchers have investigated the recognition of resident activities using a variety of mechanisms, such as Na¨ıve Bayes classifiers, Markov models, and dynamic Bayes networks. In multiple cases, in spite of its simple design and simplified assumptions, Na¨ıve Bayes classifiers often work much better than expected, especially when a specific group of sensors can easily be identified as characteristic of a certain activity [18].
Sensors s3 s1 s2 s4
Event Detection Home
Profile User Interface
Activity Recognition Activity Prediction Appliance Scheduling
User Profile
BECM System
Actuators
In this paper, the considered scenario is that of a BECM system, based on distributed smart home sensor networks.
An overview of this system is represented in Figure 1. More
specifically, sensors are used to make observations on users
and their interactions with the surrounding environment; the
combinations of these interactions, which are detected by
the Event Detection module as events, provide meaningful
information on users’ activities. As described in more details
in [
In the following, more details will be provided about the core modules of the proposed BECM system, i.e. the Activity Recognition module, the Activity Prediction Module and the Appliance Scheduling module.
The activity recognition approach used in this paper was
earlier proposed in [
The main task of the the Activity Prediction module is to provide, for the next Appliance Scheduling module, a possible scenario in time t ahead in the future, that explains the probabilities of every activity that can begin in t, calculated thanks to information about all the activities happened and recognised before the current time t0. Starting from the assumption that activities are linked between one another, so that when an activity Ai occurred at time ti it is possible to evaluate the probability of another activity Aj to be performed by the user in a different time tj , the module has to evaluate the probability in t for every single activity Aj .
During the training phase, two different kind of probabilities have been evaluated for every activity under consideration: p(Ai(ti)) indicate the prior probability for activity Ai of starting at time ti; p((Aj jAi)(k t)) indicate the conditional probability of activity Aj of being carried out since activity Ai has started (k t) before.
The period between the current time t0 and the time t in which the prediction is needed, is split in different k time intervals of duration ( t).
All the activities that have been recognised before the current time t0 are stored with their respective starting time ti, so that it is known how many time intervals outdistance every ti up to t, and it is possible to calculate the conditional probability of activity Aj in t considering all the activities Ai that occurred in ti. Then, for every (k t) between t0 and t, the probability of Aj to be happening in t is calculated with respect to the fact that Ai could be happening in (k t). These two contributions are added together according to the equation below: p(Aj (t)) =
X p((Aj jAi)( i t t t i ))+ X X p((Aj jAi)(k t)) p(Ai(t0 + k t))
i k with k 2 0; (t t0)= t . (1)
Every activity coincides with one of the appliances in the house, so that the probability for each activity in t, calculated as explained, is then translated in the probability of one appliance to be used at time t. Therefore, the output from this module is going to enable the scheduling algorithm to make the validation necessary for the scheduling of controllable appliances and for evaluating energy consumption. The algorithm decides to schedule at time t only those appliances corresponding to activities that have their value of probability higher than a certain threshold.
The appliance scheduling algorithm is based on the smart
home energy management system proposed in [
RiCi ln
Tiexp + RiPiheat
(2) where Tiout(t) and Tiin(t) Piheat are the initial outside and inside temperature respectively, and Piheat, Ri and Ci are characteristic parameters for the appliance. More specifically, Piheat is the heat rate (in Watt), Ri is the equivalent thermal resistance ( C/Watt) and Ci is the equivalent heat capacity (Joule/ C). If the appliance is off, Piheat = 0.
The appliance scheduling algorithm then schedules
appliances according to their related cost contribution value, which
includes both the energy consumption- and user
annoyancerelated costs. User annoyance is computed according to the
results of a survey, completed by 427 people, as reported in [
For G2 appliances, the cost to start at time tiST and end at time tiEND = tiST + tiexec is defined as
CiG2(tiST ) =
where (t) is the electricity tariff at time t, and tiST is the relative satisfaction level for a time interval tiST = tiST tiP T , which is in inverse proportion with the user annoyance of shifting the appliance starting time. If tiST = 0, the cost value CG2(tiST ) ! 1.
i
For G3 appliances, the cost to start at time tiST and end at time tiEND = tiST + tiexec (Tiexp) is defined as
+ ( Tiexp)
The algorithm for modelling the activities and then discovering what the resident is doing is implemented and tested using the Aruba real-word dataset from the CASAS smart environment project of the Washington State University [19]. The data were collected from one smart apartment provided with motion sensors, contact sensors in the doors or cabinets and temperature sensors. The events decoded by these sensors are significant for recording elementary actions that people are performing, while the aggregation of these elementary actions defines one activity of interest. To correctly evaluate the correlation between the sets of events and the observed user’s activities, without interference from other people, a dataset with only one resident living in the home was considered.
To evaluate the proposed system, in addition to the activities of the Aruba real-word dataset, some other activities have been simulated as performed by the same user inside this home scenario, using the same kind of sensors already installed in the house. The simulated activities are the following three activities not reported in the real dataset: using the washing machine, using the dish washer, taking a shower, which, along with the activity of washing dishes by hand, causes the water heater to turn on. Taking a shower is supposed to be carried out by the user in the bathroom, therefore involving the motion sensors already installed close to this room and assuming that hot water is used, thus causing the water heater to switch on. The use of the dish washer is supposed to be performed in the kitchen, involving the sensors in that area and simulating the presence of a specific cabinet containing the appropriate detergent and with a magnet sensor to understand its opening or closing, so that the activity of loading the dish washer could be recognised concluded only when this cabinet had been closed. The same thing was done for the activity of using the washing machine, by setting up another specific cabinet with its magnetic sensor, and placing it in a room of the house where there are not other linked activities.
The system needs a correspondence between some of the activities and the use of certain household appliances, in order to predict energy consumption based on the probability of the activities to occur. Table I shows the considered activities along with their corresponding appliance owned by the user.
With respect to only the four activities corresponding to
controllable appliances, i.e. appliances belonging either to
G2 or G3, the recognition algorithm presented in [
As for the prediction of future activities, the algorithm has an overall accuracy of 67%. The activities more accurately predicted are those with many samples and recurrent starting time, like the activity “Taking a Shower”, because statistics about them are quite significant. For other less frequent activities, i.e “Wash Dishes”, the prediction is instead less reliable.
The algorithm has been compared with two different situations with respect to the case where no scheduling is involved.
There are then three possible scenarios: the first one is the classic situation where appliances are normally used by the resident and the scheduling is never programmed (Without Scheduling Algorithm-WSA); the second one is based on a perfect knowledge of the time in which the user wants to use some of the appliances in the house (Scheduling Based on Perfect Time-SBPT). This case coincides with the possible scenario in which the user instructs the system about the exact moment they want the appliance to start, but it has the disadvantage of requiring continuous interactions between users and system; the last one bases its scheduling evaluations on the probability of using any of the appliance at time t, calculated as explained in equation 1 (Scheduling Based on Probability-SBP). This solution allows to avoid interactions between users and the system, considering only the system’s previous knowledge about user habits. The training phase to obtain all the information about user’s behaviours and preferences, and that allowed the system to perform the calculations on the probabilities indicated in subsection III-B, took into consideration two months of data about performed activities. Due to the fact that the shortest duration for the activities in exam is around 15 minutes, while the longest activities can elapse for several hours, conditional probabilities between activities was valuated choosing duration of intervals ( t) equal to 15 minutes. The scheduling algorithm was instead tested on one week of data. For every interval of time t0 during the testing week, the algorithm schedules appliances that are going to be used every k t time intervals after t0, with k 2 N, trying to improve energy savings and user’s comfort. Simulations were done considering time intervals of 30 minutes and predicting future activities in t up to 9 hours forward in the future, so that every half an hour the scheduling algorithm could re-evaluate its scheduling based on the new information about previously user performed activities and with new calculations of probabilities p(Aj (t)).
The obtained results were considered with respect to energy
consumption in one week, comparing the case with scheduling
in relation to the case of normal use of household appliances,
and evaluating if the scheduling could generate some kind of
annoyance for the user. The evaluation of the energy costs has
been made using two different tariffs listed in Table II, based
on some typical Italian tariffs. The annoyance rate is defined
as in [
Table III shows the results about energy saving compar- in Fig.2, where the average annoyance rate is presented, it is
ing the two cases with scheduling against the case without possible to observe how for every appliance the annoyance
scheduling. These results are obtained taking into account rate is always close to the lowest value of 1. The knowledge
the fact that cost savings are coming from a scheduling of of user behaviours has therefore guaranteed the scheduling of
switching controlled high loads to hours where the energy has the appliance with the best trade-off between energy costs and
lower prices and considering that there is a reduction in energy user preferences.
consumption due to a better optimisation in the usage of the A slightly different discussion has to be done with reference
water heater, which is switched on only at times of interest to the scenario in which the scheduling is evaluated based
for the user and not every time the temperature drops below on the probabilities of activities and, accordingly, on the
probabilities of using a certain appliance. Even in this case
there is an evident reduction in energy consumption during the
week, as shown in table III. Most of the saving come again
from the wiser use of the water heater thanks to the scheduling
only at appropriate time. Looking at Fig.2 it is possible to see,
however, how the annoyance rate reaches a higher level. This
is due to the fact that the use of the water heater is linked to
two different activities, as shown in Table I. While activity
6 is always easily recognised and predicted, and therefore
scheduled, activity 4 gives some problems because it is often
confused with other activities [
This paper focuses on a solution for energy and comfort management inside buildings, with the purpose of reducing energy waste thanks to a proper control over appliances, while on the same time ensure the well-being of users. To this aim, a BECM system is proposed that integrates a solution for two different problem: the first one concerns the needs for such a system to be able to know users behaviour and preferences and to predict usual activities; the second is about the necessity to manage appliances with respect of that behaviours and preferences and with respect of energy consumption.
The system has been tested in a real scenario, evaluating if the predictions were correct and proposing a coherent scheduling that could guarantee energy savings. The obtaining results show that, as expected, the scheduling of the appliances can guaranteed energy savings, reducing consumption over a week of at least 49:2% in comparison with classic use of energy and appliances. The prediction module permitted a quite accurate scheduling basing on probabilities, even if some of the activities has given some problem due to the fact that the statistic data about them were based of few instances. Furthermore, it was possible to guarantee that the annoyance rate was never too high, thus respecting user comfort.
Future works will investigate the adaptability of the proposed system to different real-case scenario, trying to improve the prediction module considering a larger training phase and more instances of the activities and corresponding use of appliances. Furthermore, it will be evaluated how the presence of Renewable Energy Sources could affect appliance scheduling and improve energy savings.
This work was partially funded by the INSIEME MiSE project, HORIZON 2020, PON 2014/2020 POS. 395, and by the LEAPH project, POR FESR Sardegna 2014-2020, Asse 1, Azione 1.1.3. [12] ——, “Intelligent multi-objective optimization for building energy and comfort management,” Journal of King Saud University-Engineering Sciences, vol. 30, no. 2, pp. 195–204, 2018. [13] M. Fayaz and D. Kim, “Energy consumption optimization and user comfort management in residential buildings using a bat algorithm and fuzzy logic,” Energies, vol. 11, no. 1, p. 161, 2018. [14] P. Du and N. Lu, “Appliance commitment for household load scheduling,” in PES T&D 2012. IEEE, 2012, pp. 1–1. [15] G. Bhat, R. Deb et al., “Online human activity recognition using lowpower wearable devices,” in 2018 IEEE/ACM International Conference on Computer-Aided Design (ICCAD). IEEE, 2018, pp. 1–8. [16] M. M. Hassan, M. Z. Uddin et al., “A robust human activity recognition system using smartphone sensors and deep learning,” Future Generation Computer Systems, vol. 81, pp. 307–313, 2018. [17] Y. Liu, L. Nie, L. Liu, and D. S. Rosenblum, “From action to activity: sensor-based activity recognition,” Neurocomputing, vol. 181, pp. 108– 115, 2016. [18] N. C. Krishnan and D. J. Cook, “Activity recognition on streaming sensor data,” Pervasive and mobile computing, vol. 10, pp. 138–154, 2014. [19] D. J. Cook, “Learning setting-generalized activity models for smart spaces,” IEEE intelligent systems, vol. 2010, no. 99, p. 1, 2010.