Home Energy Saving through a User Profiling System based onWireless Sensors ∗

Antimo Barbato, Luca Borsani, Antonio Capone, Stefano MelziDipartimento di Elettronica e Informazione, Politecnico di Milano, Italy

{barbato, borsani, capone, melzi}@elet.polimi.it

AbstractThe high energy required by home appliances (like white

goods, audio/video devices and communication equipments)and air conditioning systems (heating and cooling), makesour homes one of the most critical areas for the impact ofenergy consumption on natural environment. In this paperwe present a work in progress within the European projectAIM for the design of a system that can minimize energywaste in home environments efficiently managing devicesoperation modes. In our architecture we use a wireless sen-sor network to monitor physical parameters (like light andtemperature) as well as the presence of users at home andin each of its rooms. With gathered data our system createsprofiles of the behavior of house inhabitants and through aprediction algorithm is able to automatically set system pa-rameters in order to optimize energy consumption and costwhile guaranteeing the required comfort level. When userschange their habits due to unpredictable events, the systemis able to detect wrong predictions analyzing in real time in-formation from sensors and to modify system behavior ac-cordingly. By the automatic control of energy managementsystem it is possible to avoid complex manual settings of sys-tem parameters that would prevent the introduction of homeautomation systems for energy saving into the mass market.

Categories and Subject DescriptorsC.2 [Computer-Communications Networks]: Network

Architecture and Design; H.3.4 [Information Storage andRetrieval]: Systems and Software—user profiles

General TermsAlgorithms, Design, Experimentation, Management∗This work has been supported by the European project AIM (A

novel architecture for modelling, virtualising and managing the en-ergy consumption of household appliances), European Commission7th Framework Programme, Grant Agreement Number 224621.

Permission to make digital or hard copies of all or part of this work for personalorclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. To copy otherwise, to republish, to post on servers or to redistributeto lists, requires prior specific permission and/or a fee.BuildSys’09,November 3, 2009, Berkeley, CA, USA.Copyright 2009 ACM 978-1-60558-824-7 ...$5.00

KeywordsEnergy saving, Home automation, Smart monitoring,

User profiling, Wireless sensor networks

1 IntroductionAccording to recent studies [1] energy consumptions are

increasing year after year, and if effective energy saving poli-cies will not be adopted, in 2030 they will double with re-spect to 1980 level and will increase by 28% on 2006 level.The residential sector accounts for an increasing percentageof the total consumption which is now above 27.5% (sourceEarthtrends). Indeed, in other sectors like the industrialonethe introduction of strategies for the reduction of energy con-sumption have been stimulated by the urgent need to im-prove production efficiency, while residential users have alow awareness of the problem and usually lack of tools formeasuring and optimizing the energy consumption of theirdaily activities. The top four residential end-uses of en-ergy (as a percentage of primary energy) are: space heating(26.4% of total primary energy end use), space cooling (13%of total primary energy end use), water heating (12.5% of to-tal primary energy end use), lighting (11.6% of total primaryenergy end use).

These predictions have recently increased the interest ofthe research community as well as of the industry world inthe use of new generation home automation systems for en-ergy saving. The general goal is to use monitoring and con-trol devices to measure in real time the energy consumptionof home appliances and to set them to low power modeswhen possible in order to save energy. Moreover, the infor-mation exchange between the home automation system andthe energy utility through a data communication network al-lows to improve the efficiency in energy production and tostimulate a wise energy use with differentiated tariffs pertime period.

In this paper, we present an integrated system for intel-ligent energy management at home currently under devel-opment within the European project AIM. In particular wefocus on the role played by wireless sensor networks to auto-matically control home appliances (mainly devices used forthe space heating/cooling, lighting) according to the user’shabits. The main function enabled by the sensor network isuser profiling. User profiling process includes basically twoprocedures: a mechanism for recording some events that cancharacterize the way in which users interact with the home

environment and the available appliances, and a predictionalgorithm that allows extracting from all these data somereasonable settings of the energy management system that isexpected to be the most appropriate meet user requirements.The final goal of user profiling is that of replacing some ofthe required system settings based on a manual interactionwith a user interface with an automatic configuration pro-cedure that can be performed on request. For this purpose,this function must be able to provide inputs to the energymanagement system exactly in the same way a user could dothrough the user interface and it can be considered a plug-inof the system that can be enabled or disabled by the user.

The paper is organized as follows: In Section 2 we reviewprevious work on home automation focusing on energy sav-ing systems and present the basic characteristics of the sys-tem under development within the AIM project. In Section 3we present the wireless sensor network architecture adoptedand the middleware MobiWSN that we designed and devel-oped to manage data gathering and processing in a flexibleand efficient way. In Section 4 we present the user profilingmechanisms focusing in particular on presence sensors. InSection 5 some numerical results are presented to provide ex-amples of possible advantages achievable with the proposedsystem in terms of energy consumption.

2 Home Automation for Energy ManagementIn recent years, several research efforts have been carried

out to design the so calledsmart home[4]. One of the mostattractive potentiality of this kind of environment is the pos-sibility to reduce the energy consumption managing intelli-gently the devices into the house.

2.1 Related workA common way to model an intelligent home environ-

ment control system is to divide it in four cooperating lay-ers [5]. Thephysicallayer contains the hardware within thehouse including individual devices, transducers, and networkhardware and is in charge to collect data from the environ-ment and to perform actions. Thecommunicationlayer al-lows the exchange of information within the network of de-vices. Theinformation layer gathers, stores, and generatesknowledge useful for thedecisionlayer.

Previous works focus on one or more of these layers and,using different approaches, aim at creating specific function-alities to increase comfort for the home habitants while re-ducing energy consumption. Exploiting algorithms based onfuzzy logic, in [7] a system able to learn users preferences,to predict users needs (e.g. light intensity, temperature ), andto self adjust system behavior when users change their habitsis proposed. In [9] a different approach based on neural net-works is proposed for the decision layer with the goal of sim-plifying the learning process of user preferences without atraining phase. Neural networks are adopted also in [11] tocreate a system able to control temperature, light, ventilationand water heating.

Light automatic management is an interesting applicationon which research has particularly focused because of its rel-evance for the total energy consumption. Some systems thathave been proposed try to exploit external light sources inorder to increase energy saving with what is usually called

daylight harvesting[12]. Context awarenessis one of thekey elements to building smart environments and obviouslyuser presence is part of it. One of the simplest ways of usinginformation on user presence is to control dynamically light-ing, but the raw data of sensor presence (e.g.infrared sensor)can not be enough to optimize energy consumption [6].

Smart home can be very helpful to electric utilities to es-timate and even control customer power demand in order tooptimize energy production and transmission and to avoidcostly peaks that usually requires auxiliary generators. Thegoal in this case is to equalize energy demand and to allowa more efficient use of green energy sources regulating en-ergy consumption in real time on the base of availability.Since 1991 the idea of regulating energy demand motivatethe research on home automation and communications net-work systems to control home appliances [15].2.2 The AIM project

The work presented in this paper is part of the projectAIM that aims at developing the technology for profiling andoptimizing the energy consumption patterns of home appli-ances and providing concrete examples related to three ap-plication areas: white goods, audio/video equipments andcommunication equipments [3]. To this purpose, a novel ar-chitecture allowing real-time energy consumption monitor-ing and management, as well as the virtualization of energycontrol is proposed. This architecture that consists six basiccomponents namely the home gateway, the Energy Manage-ment Device (EMD), the home network, the sensor network,and the appliances [14, 2].

The energy consumers are controlled by an EMD thatworks as the local hub of the AIM energy control. EMDcommunicates, through proper communication channelscalledInterfaceswith all the energy consumption actors us-ing one or more physical communication media and associ-ated protocols. The implicated communication technologiesare based on wireless, power-line or Ethernet connectivity.

EMD is in its turn controlled by an home Gatewaythrough an interface ensuring access to multiple EMD’s froma single access-point, either locally (’domestic’ users) or of-fers a single access-point for controlling the full system re-motely. The Gateway is capable to become the ”transfernode” between the Smart Home and the Smart Grid. Themain assets of this node from the utility point of view arethe exchange and provisioning of information between util-ity and the customer that allow implementation of servicesfor energy saving, flexible tariffs, reliable power consump-tion forecasts and the possibility to store energy if required.

To simplify the development of energy management ap-plications, AIM defines a Device Virtualization Environment(DVE) that allows to mask hardware and software imple-mentation specific attributes from the rest of the AIM archi-tecture. The DVE optimizes energy consumption and cost,scheduling tasks during the daytime for every appliances un-der user defined constrains (e.g. choosing the best time tostart washing machine program with the constraints of hav-ing clothes ready by the user defined time).

The optimization of energy consumption requires someinput information from different sources: users must providetheir preferences and needs (light level, temperature, or ac-

tivities required to appliances, etc.) and the energy utilitymust provide the cost of energy for every time of the day.To create a system where the user doesn’t need to waste alot of time in complex settings of system parameters, oneof the challenges of AIM project is to automate the set upof a part of the user preferences with a system able to pre-dict actual user preferences on the basis of previous observedbehavior. This is the main role of the sensor network thatsenses physical parameters (like temperature and light), col-lects data about the presence of users in every room of thehouse, and estimates the user presence for future periods an-alyzing historical data and adjusting prediction in real time.On the basis of this prediction it is possible to activate appli-ances at the best moment, for example heating the room atthe desired temperature before the user come in.

3 Wireless Sensors for Smart Energy Man-agement

In the AIM architecture, the wireless sensor network(WSN) provides the basic tools for gathering the informa-tion on user behavior and its interaction with appliances fromthe home environment. Moreover, the sensor network pro-vides measurements of some physical parameters like tem-perature and light that can be used by the system to performsome automatic adjustment of the energy management sys-tem. The sensor network can be implemented using severalavailable communication technologies. Generally speaking,WSNs are today considered the most promising and flexibletechnologies for creating low cost and easy to deploy sensornetworks in many scenarios including home automation.

Since in the considered scenario the area to be monitoredis usually relatively small, it would be possible to consider asingle WSN interconnecting all the sensing devices neededfor monitoring physical parameters and user activities. How-ever, in practice this option is not the most convenient one.For cost and commercial reasons heterogeneous technolo-gies (with different transmission mechanisms and/or proto-col stacks) may be adopted by appliance manufacturers andthen integrated together in the home network. Moreover, thefeatures of the sensor network can be tailored on user needsand the specific characteristics of the environment addingnew groups of sensing devices even a later stage when thesystem is already in operation.

For these reasons, we propose a hierarchical hybrid net-work architecture consisting of different islands of sensornodes (mainly homogeneous WSNs, but also wired tech-nologies can be included) interconnected through gateways.These are higher layer devices able to communicate via het-erogeneous links that can be either wireless (WiFi, ZigBee,etc.) and wired (PLC, ethernet, etc.) and to perform somedata aggregation and processing tasks. The network topol-ogy created among the gateways is possibly meshed to en-sure reliability and resilience to failure.

To hide the complexity of the underlaying heterogeneousnetwork to application developers and users, and to providesome tools to manage different sensor networks we proposea middleware [8], calledmobiWSN, that we specifically de-signed and implemented for hybrid hierarchical WSN archi-tectures. MobiWSN allows to ease up the interaction of the

user client applications (like home automation applications)with the network, the network configuration and reconfigu-ration, as well as the use of computation and communica-tion resources. We have implemented and tested the pro-posed middleware using MICAz and Intel Mote2 devices(two motes architectures based on CC2420 communicationchipset with IEEE 802.15.4 technology) and Linux-PC basedgateways with IEEE 802.11 wireless interfaces [10].

In MobiWSN architecture gateways are interconnectedand can communicate with an additional node, called man-ager, that is in charge of managing network creation and re-configuration. The coordinator of each WSN island is inte-grated in or directly connected to a gateway. All the sensorsbelonging to the WSN are connected to the gateway throughmulti-hop paths which form a routing tree rooted at the gate-way. The manager dynamically controls network topologyand ensures gateways can directly communicate. The list ofactive gateways and of the sensors associated with them ismanaged by an activeRemote Method Invocation(RMI) in-terface. When a new gateway is switched on, it registers onthe manager through this interface to become reachable fromremote client applications that want to interact with the con-nected WSN.

MobiWSN consists of two main software components:the first one is (written in Java in our prototype implemen-tation) is installed on gateways, and the second one is a motemiddleware layer (written in nesC withTinyOSoperatingsystem [13] in our prototype implementation).

The middleware component running on the gateways im-plements three main functionalities: i) Sensor and Mote Ob-ject Management, that is used to create new objects (sensorsnodes or group of sensor nodes) when new devices are con-nected to network and to associate specific functions to spe-cific objects defining the library of operation of the system;ii) WSN Interrogation Libraries, which define the informa-tion exchange with the WSN and specify the signalling mes-sages; and iii) Class Loader, that allows an external clientoranother gateway to load on a specific gateway new classesof objects and the set of functions related to them. This lastfunctionality is particularly important and is one of the novelissues of MobiWSN since it allows application client to re-motely load on the gateway new objects and functions man-aging the whole system like a flexible programmable plat-form that allows to adapt the system to the specific charac-teristics of the application scenario and the installed devicesand to update it with new applications and services writtenby third parties.

MobiWSN also defines a stateless protocol, calledInfor-mation Exchange Protocol(IEP), to allow information ex-change between the sensor networks and their respectivegateways. This protocol requires to the network layer ofthe WSN the support of broadcast and unicast addressingand the control of duplicate packets. IEP includes four mes-sage types:request, response, commandand information.These messages can be addressed to a single mote, a groupof nodes or to all motes of the WSN. In this architecture agroup can include motes of different WSNs, and the man-ager has the task of its management. In the proposed archi-tecture for home energy management we associate a group

to each room of the house so as to simplify the informationexchange among nodes that monitor the same room and re-duce signaling overhead. Moreover, it is possible to createagroup of sensors (e.g. temperature) to deliver a read to thewhole group with a single IEP message.

MobiWSN allows not only a flexible management ofgateway functionalities and software components but also adynamic control of tasks performed by motes. Every nodecan have severalfunctionalities(a piece of code that imple-ments a specific task) which can be started and stopped bythe gateway. The goal of these functionalities is to provideto the application client some high level data processing pro-cedures that are performed locally on the WSN. These canbe used as building blocks for creating complex applications,reducing the number of exchanged messages, and exploitingthe distributed processing capabilities of the system.

In the proposed architecture for home energy manage-ment, besides the temperature and lighting information forevery room, we need to create a user profile through pres-ence detection. This has been obtained using a low-cost tech-nology based on multiple infrared sensors deployed in allrooms. To filter the motion events captured by the infraredsensors and to process this raw information within the motesof a group, we implement a high level functionality whichhas the task to report to the gateway only information on thepresence users in the room, reducing the number of long pathmessages. By using these announcement messages, a homeautomation client application installed on the energy man-agement platform can create presence profiles without con-sidering the specific presence detection technology adoptedbut focusing only on the problem of extracting from data in-formation on user habits. Obviously, this has the advantageof simplifying the task of application software developmentand of increasing software reusability.

As mentioned above, the proposed middleware archi-tecture requires some basic services to the network layer,namely: best-effort delivery, discarding duplicate packets,unicast and broadcast addressing. The routing protocolwe adopted for WSNs, calledHierarchical Addressing Tree(HAT), allows the creation of a network with a tree topol-ogy and a hierarchical addressing. The network address isdivided into levels according to the number of hops from theroot of the tree, each of which has the same number of bitsthat represents the maximum number of sons for every nodein the hierarchical tree topology. The routing tree is createdwith a node association procedure and maintained throughthe exchange of periodic beacon messages.

4 User profilingData collected by sensors located in the house are used

for monitoring the environmental parameters, such as userpresence, temperature and light, in two different ways:Off-line modeandReal-time mode.

In the off-line mode information collected by the WSNsis aggregated providing inputs for user profiling. The basicfunction of the user profile is the characterization of usersbehavior so that some settings of the energy managementsystem can be made automatically. Different types of pro-file can be created such as user presence profile, temperature

profile, light profile. In the user presence profiling the sen-sor network collects 24 hour information (here called“dailyprofile” ) about users presence/absence in each room of thehouse in a given monitoring period (i.e. week, month). Atthe end of the monitoring time the cross-correlation betweeneach couple of 24 hour data presence is computed for eachroom of the house in order to cluster similar daily profiles.The daily profilesy(t) andx(t) are considered similar if:

r(x,y) >1−A

2[r(x,x)+ r(y,y)]

Wherer(x,y) is the mean value of the cross-correlationbetween signalsx(t) and y(t) calculated with an accepteddelay±B (in minutes),A andB are constants (respectivelyequal to 0.12 and 10 in our numerical results).

0 2 4 6 8 10 12 14 16 18 20 22 240

0.2

0.4

0.6

0.8

1Final Presence Profile

Day Time [hours]U

ser

Pre

senc

e P

roba

bilit

y

Figure 1. Example of the daily presence profile.

For each cluster the average of the daily profiles identifiesa user presence profile that provides the 24 hour probabilitydistribution of the user presence in the room the cluster isassociated with (Figure 1). At the end of calculation a matrixis generated where each room is associated with a columnthat represents the sequence of presence profiles identifiedinthe monitoring period. Each matrix column is statisticallyelaborated in order to predict the presence profile in a givenday, for each room, on the basis of the observed profiles inthe past days. For roomi the prediction algorithm performs:

1) the column of profiles associated with roomi is selectedand the auto-correlation is computed; the distance be-tween the first two highest positive peaks of the auto-correlation function identifies the repetition periodTi ofthat room profiles sequence;

2) for each presence profilej in the selected column, theprobability that it occurs after the sequence of profilesof the pastM days in roomi (with M = 1) is calculated;

3) if a profile j exists with such a probability higher thana threshold (experimentally set to 0.75), the algorithmstops andj will be the predicted profile; otherwiseM isincreased by 1 and the algorithm goes back to step 2 (ifM = Ti −1 the procedure stops anyway and returns themost probable profile found in the last iteration).

In the temperature and light profiling the sensor networkcollects the daily temperature and light level in each room ofthe house during the same monitoring period used for thepresence. For each room and user presence profilei, theassociated temperature and light profiles are calculated asthe average of the past values collected in the days wherethe profilei was experimented. The assumption here is that

0 2 4 6 8 10 12 14 16 18 20 22 24

1

2

3

4

5

6

7Used Profile Dynamic Updating

Day Time [hours]

Use

d P

rofil

e Id

entif

ier

Figure 2. Single room presence profile dynamic updating

during the monitoring period the user can manually regu-late preferred light and temperature levels so that they canberecorded by the system and then used during the operationperiods, however a separation between the monitoring andthe operation periods is not necessary since users can alwaysmanually adjust parameter settings that are automaticallyse-lected by the system. At the end of the off-line mode threetypes of profile are available with each temperature and lightprofile associated with a presence profile.

In the real-time mode the sensor network collects realtime measurements of physical parameters that can be usedby the system to perform some automatic adjustment of theenergy management system (like e.g. regulating lighting sys-tem according to the level of natural light from windows,control the heating/conditioning system to set temperaturein the rooms according to the user profile, etc.). The realtime user presence information is also used to dynamicallyupdate the predicted profile during the day, through theLo-cal Updating Algorithm(LUA), tracking the user behavior(Figure 2): for each room the cross-correlation between thestored presence profiles and the real time detected ones(t)is calculated dynamically during the day. A switch from apredicted profilex(t), to another oney(t), is performed if:

r(y,s)− r(x,s) > C(r(y,s)+ r(x,s))

Wherer(x,y) is the mean value of the cross-correlation be-tween signalsx(t) andy(t) calculated with an accepted de-lay of ±D (in minutes),C andD are constants (respectivelyequal to 0.1 and 10 in our numerical results).

When the LUA, changes the current used profile of theroom i in a new one, also the profiles of the other rooms canbe updated through theGlobal Updating Algorithm(GUA):the profile matrix is used to verify if the new profile of theroom i is frequently associated with particular profiles of theother rooms. If an association is found the updating is alsoperformed for the other rooms.

Data collected in the off-line/real-time mode can be usedfor performing several tasks, like for example:

- House temperature management: if the system esti-mates through the presence probability profile that thereis a high probability that at 7:00 PM the user is usuallyback home from work and from the temperature pro-file it is known that the preferred temperature level inthe living room in the evening is 22◦ C, the system willautomatically optimize the use of the heating system inorder to minimizing energy consumption and reachingthe desired temperature level at that hour of the day;

- Lighting system management: if the system detects thatat 7:00 PM the user is in the living room and it is knownthrough the temperature profile that when he is in thatroom at that hour the desired light level is “high”, thesystem will automatically manage the light adding tothe sun light the gap to obtain the desired level;

- Devices working mode (on, off, stand-by) management:if the system predicts, for example, through the pres-ence probability profile, that at 7:00 PM the user is backhome from work and connects to internet, it will turn onthe modem at that hour of the day.

5 Numerical resultsAs mentioned previously, we implemented a prototype

version of the proposed sensor network architecture for en-ergy management. However, to evaluate the performance ofthe user-behavior prediction algorithms we have been forcedto rely on simulation mainly because of the long period oftime required for testing them in a real environment and thedifficulties to create a realistic setting. The system has beensimulated referring to a five room house with a simulatingperiod of 300 days. For each room of the house a regularsequence of realistic daily presence, light and temperatureprofiles has been created. From that sequence the simulateddaily profiles have been generated introducing random vari-ations on the original profile (the user will not wake up everyday exactly at 7:00 A.M.). Moreover we introduced someexceptions in the users periodic behavior in order to simu-late a real use case: for example, user could not go to workin a working day because he is sick.

In the simulation performed a monitoring period of 30days has been used; in the first monitoring period the systemjust collects information on user behavior. After the first 30days the system begins to work using always the last 30 dayslike monitoring period to generate, for example, the presenceprofiles. The presence prediction algorithm has been sim-ulated in three user behavior exceptions cases: exceptionsspike (there are 20 isolated exceptions in the users behavior),exceptions burst (there are 4 sequences of 4 contiguous ex-ceptions in the users behavior) and changing behavior (userchanges his behavior two times during the year, for exam-ple as we move from winter to spring/summer the user goesto sleep later). The results of the 300 days simulation arepresented in Table 1.

Room1 2 3 4 5

Exceptions 88.00% 90.33% 87.67% 88.67% 87.00%SpikeExceptions 94.00% 92.00% 93.67% 91.67% 93.33%BurstBehavior 90.00% 91.00% 90.00% 91.67% 90.67%Variation

Table 1. Percentage of correctly predicted profiles foreach room of the house.

Room1 2 3 4 5

LUA 2.05% 2.55% 1.78% 0.58% 0.79%

GUA 1.38% 1.67% 1.54% 0.36% 0.66%

Table 2. Perc. of time with a wrong user presence predic-tion when the predicted profile of the day is not correct.

The dynamic profile updating has been tested in all threecases and the average values have been computed for theLUA and GUA Table 2. Considering all the simulating pe-riod (300 days) the percentage of time with a wrong userpresence prediction is very small.

The presence, temperature and light profiles can be usedto optimize the using time of home appliances and to mini-mize the home energy consumption. In the simulation per-formed the home temperature management has been consid-ered, using the presence and temperature profiles to automat-ically manage the temperature in each room of the house.In Figure 3 and 4 we present an example of the differencebetween the working time of the cooling system with andwithout the automatic temperature management. The man-agement system allows some energy savings turning off thecooling system of the rooms that are not required to beair conditioned because the user will not enter those roomswith high probably and turning it off in the whole house ifthe user is not present and probably will not return for along time. In contrast, in the “classical scenario” the cool-ing system is supposed to be On in all rooms and to bepreprogrammed from the user to approximately follow hisdaily/weekly schedule.

In the simulation performed, the home temperature man-agement has reduced the working time of the cooling systemby nearly 28 percent. Obviously, the working time reductionand the resulting energy saving are strictly related to the spe-cific scenario considered since they depend, for example, onthe presence/absence periods in each room of the house.

0 2 4 6 8 10 12 14 16 18 20 22 24Cooling System is Off

User is Present

Cooling System is On

User is Absent/

Daily Cooling System Working Mode

Day Time [hours]

User PresenceCooling systemworking mode

253.15 minutes

Cooling system workingtime = 316.17 minutes

63.02 minutes

Figure 3. Room 3 daily cooling system working modewithout home automation

0 2 4 6 8 10 12 14 16 18 20 22 24Cooling System is Off

User is Present

Cooling System is On

User is Absent/

Daily Cooling System Workig Mode

Day Time [hours]

User PresenceCooling systemworking mode

63.15 minutes

73.13 minutes

Cooling system workingtime = 199.45 minutes

63.21 minutes

Figure 4. Room 3 daily cooling system working modewith home automation

6 ConclusionIn this paper we presented a home energy management

system under development within the European project AIM.We proposed a heterogeneous hierarchical sensor networkarchitecture to gather physical parameters and to monitoruser behavior. We designed and implemented a middle-

ware able to deal with network heterogeneity and dynam-ics, as well as to greatly simplify application development.Data collected by the sensors are used to create user pro-files. Based on user profiles and real-time information pro-vided by the system, we can predict user behavior and op-timize the energy consumption controlling in an automaticway home appliances. We proposed a new approach to im-plement a self adaptive prediction algorithms to set severalparameters (light intensity, temperature, etc.) according touser estimated preferences. The presented solution is simplerthan other profiling systems, mentioned in 2.1, which rely oncomplex learning techniques: just replicating a previouslyobserved set up that satisfied the user in a similar context pro-vides good results and requires shorter training periods. Weimplemented a prototype version of the propose sensor net-work architecture based on off-the-shelf devices. Moreover,we simulated our prediction algorithms over long time peri-ods and showed their effectiveness in estimating user pres-ence in normal conditions and their ability to quickly detectanomalous conditions and to correct estimations.

7 References[1] US Dep. of Energy, ”Buildings Energy Data Book”, 2008.[2] H. Hrasnica, S. Tombros, A. Capone, M. Barros. A New

Architecture for Reduction of Energy Consumption of HomeAppliances. IneENVI 2009, Prague, March 25-27, 2009.

[3] AIM Project. http://www.ict-aim.eu/.[4] D. Cook and S. Das. How smart are our environments? An

updated look at the state of the art.Pervasive and MobileComputing, 3(2):53–73, 2007.

[5] D. Cook, M. Huber, K. Gopalratnam, and M. Youngblood.Learning to control a smart home environment. InInnovativeAppl. of Artificial Intelligence, 2003.

[6] V. Garg and N. Bansal. Smart occupancy sensors to reduceenergy consumption.Energy & Buildings, 32(1):81–87, 2000.

[7] H. Hagras et al. Creating an ambient-intelligence environmentusing embedded agents.IEEE Intelligent Systems, pages 12–20, 2004.

[8] K. Henricksen and R. Robinson. A survey of middleware forsensor networks: state-of-the-art and future directions. InInt.Workshop on Middleware for Sensor Networks, 2006.

[9] K. Hunt, D. Sbarbaro, R.̇Zbikowski, and P. Gawthrop. Neuralnetworks for control systems: a survey.Automatica (Journalof IFAC), 28(6):1083–1112, 1992.

[10] A. Laurucci, S. Melzi, and M. Cesana. A Reconfigurable Mid-dleware for Dynamic Management of Heterogeneous Appli-cations in Multi-Gateway Mobile Sensor Networks.Demopresented at IEEE SECON, Rome (Italy), June 22–26, 2009.

[11] M. Mozer. The neural network house: An environment thatadapts to its inhabitants. InAAAI Spring Symp. on IntelligentEnvironments, 1998.

[12] V. Singhvi, A. Krause, C. Guestrin, J. Garrett Jr, andH. Matthews. Intelligent light control using sensor networks.In International conference on Embedded networked sensorsystems, 2005.

[13] TinyOS. http://www.tinyos.net/.[14] S. Tombros, N. Mouratidis, M. Draaijer, A. Foglar, and

H. Hrasnica. Enabling applicability of energy saving appli-cations on the appliances of the home environment.IEEENetwork, to appear, 2009.

[15] K. Wacks, H. Ind, and M. Stoneham. Utility load manage-ment using home automation.IEEE Transactions on Con-sumer Electronics, 37(2):168–174, 1991.