A Novel Smart Home Energy Management System: Cooperative Neighbourhood and Adaptive Renewable Energy Usage

1 A Novel Smart Home Energy Management System: Cooperative Neigbourood and Adaptive Renewable Energy Usage Matteo Cabras, Virginia Pilloni, Luigi Atzori DIEE, University of Cagliari, Italy {virginia.pilloni,l.atzori}@diee.unica.it Abstract Energy usage optimization in Smart Homes is a critical problem: over 30% of te energy consumption of te world resides in te residential sector. Usage awareness and manual appliance control alone are able to reduce consumption by 15%. Tis result could be improved if appliance control is automatic, especially if renewable sources are present locally. In tis paper, a Smart Home Energy Management system tat aims at automatically controlling appliances in groups of smart omes belonging to te same neigborood is proposed. Not only is electric power distribution considered, but also renewable energy sources suc as wind micro-turbines and solar panels. Te proposed strategy relies on two algoritms. Te Cost Saving Task Sceduling algoritm is aimed at sceduling ig-power controllable loads during off-peak ours, taking into account te expected usage of te non-controllable appliances suc as fridge, oven, etc. Tis algoritm is run wenever a new need of energy from a controllable load is detected. Te Renewable Source Power Allocation algoritm re-allocated te starting time of controllable loads wenever surplus of renewable source power is detected making use of a distributed max-consensus negotiation. Performance evaluation of te algoritms tested proves tat te proposed approac provides an energy cost saving tat goes between 35% and 65% wit reference to te case were no automatic control is used. Index Terms Renewable sources; Smart Home; Energy Management Systems I. INTRODUCTION Te last few years ave been caracterized by te tecnological revolution of te Internet of Tings (IoT) [1]. Te aim of tis paradigm is to enable te network objects to dynamically cooperate and make teir resources available, in order to reac a common goal, i.e. te reduction of energetic consumption in a building. One of te main application areas related to IoT is represented by Smart Homes, and particularly Smart Home Energy Management (SHEM) systems [2]. Smart Homes are residential buildings equipped wit devices wic cooperate in order to acieve a common set of goals. Some key features caracterize many Smart Home environments: i) available node energy is often limited. Tis is te case, for example, of battery powered nodes, wic ave limited energy amounts. ii) Smart devices, wic give te opportunity to monitor and to remotely control key equipment witin omes. iii) Decisionsupport tools aimed to aid users in making more intelligent Tis work as been partially funded by te project Artemis JU Demanes, Design, Monitoring and Operation of Adaptive Networked Embedded Systems, grant agreement no. 295372. decisions and based on maximizing te benefits gained by te end users wen tey utilize energy services. Te importance of energy usage optimization in Smart Homes is proven by statistics, wic indicates tat te electricity consumption in te residential sector represents over 30% of te energy consumption of te world [3]. As demonstrated by te literature, usage awareness alone as te potential to reduce consumption by 15% in private ouseolds [4]. In particular, te effects of te Italian Time-of-Use (TOU) tariffsbased Demand Side Management (DSM) program on demand and load sifting were examined in [5]. Comparing results wit flat tariffs, it was observed tat TOU tariffs lead to iger electricity demand and lower prices values. Te problem of SHEM is treated in many different studies: [4] and [6] propose a middleware for energy awareness integration into Smart Homes; [7] studies an automatic costeffective ligt adjustment system; [8] and [9] introduce SHEM systems tat take into account Renewable Energy Sources (RES). None of te analysed studies focus on automatic sceduling and control mecanisms of controllable appliances based on: TOU tariffs, RES power, and a User Profile inferred troug a predictive model by appliance usage. In tis work we consider a Smart Home scenario were te aim is to reduce te electricity cost by monitoring energy consumption abits and RES production, and dynamically sifting tasks of controllable appliances. Te approac is twofold: te Cost Saving Task Sceduling (CSTS) algoritm scedules tasks caracterised by ig power load in off-peak times, considering te User Profile. Te task starting time is postponed as muc as possible in order for appliances to wait for available RES power, and consequently cut electrical costs; te Renewable Source Power Allocation (RSPA) algoritm uses a max-consensus negotiation among appliances to dynamically coose wic tasks sould start immediately in order to maximise te use of RES power tat is made available by neigbours. Te remainder of te paper is organised as follows. In Section II some past studies and ow tey approaced te SHEM problem are analysed. In Section III, te reference arcitecture is introduced. Section IV describes te task sceduling model and te algoritms used, specifically CSTS and RSPA. Finally,

2 in Section V a performance analysis of te proposed algoritms is provided, and in Section VI conclusions are drawn. II. PRELIMINARIES SHEM systems are based on monitoring and controlling ouseold appliances so tat teir usage is adjusted in a costeffective way. Many of te studies proposed in te literature focus on non-automated SHEM systems, in wic suitable advices are provided to te users. Jan et al. [4] propose a middleware for energy awareness integration into Smart Homes. By using tis middleware, users can observe teir appliances energy consumption and make decisions to reduce energy costs. In [6] an optimization framework were te sceduling time of appliances is suggested is described. Te aim of tis work is to sift appliances load to off-peak time, in order to reduce electricity costs. In te last years SHEM systems ave evolved into more dynamic systems, were controllable appliances are monitored and managed by a central controller. A cost-effective control system for te reduction of ligting energy consumption is provided in [7]. Te autors of tis paper studied ow ome energy consumption can be improved by automatically adjusting ligts based on room occupancy, dayligt and time of te day. In [8] a SHEM system were appliances are controlled taking into account renewable energy gatered from te ouses witin te same neigbourood is described. RES are considered also in [9], wic furter makes use of predictive models for sort term power forecasts of te RES. Accordingly, te autors address te problem of preventing te callenges due to te sporadic nature of wind and solar power generation in designing sceduling tecniques. Altoug many studies on appliances usage profiles and prediction ave been accomplised in [10][11][12][13], tere are still no studies were predictive models on appliance usage and RES are used in automatised SHEM systems. III. REFERENCE SCENARIO In tis work we consider a Smart Home scenario were te aim is to dynamically postpone or bring forward te execution of tasks of controllable appliances so tat te electricity costs are reduced. We refer to controllable appliances as to tose wose start can be delayed, provided tat tey are executed before a given deadline. Our reference scenario is tat of a group of ouses suc as a block or a condominium, wic we call Cooperative Neigbourood. Inside eac ouse tere are appliances (e.g. electric oven, fridge, boiler, battery carger, ligt bulbs) tat consume energy. On te oter and, power supplies suc as electric grid, solar panels, and micro wind turbine provide energy tat can be used to run appliances. Smart Meters and actuators are associated to tese appliances to monitor teir energy consumption/production and control teir activation/deactivation. Te appliances are divided into 4 groups, based on teir caracteristics and requirements: Group 1: small loads suc as ligts, battery cargers; Group 2: not controllable ig loads suc as ovens, eaters; Group 3: controllable loads suc as wasing macines, dryers, electric cars; Group 4: supplies suc as solar panels, micro wind turbines. One or more tasks are associated to eac consuming appliance. We refer to te set k = {1,..., K} of tasks associated to an appliance as to te functions tat it is able to carry out. For instance, te set of tasks of a wasing macine consists of all its wasing cycles. A particular case is represented by appliances tat are able to perform only one task, suc as te TV, wic can only be turned on. For tese cases, te set of tasks is made of one single element. A power consumption amount is associated to eac task. At first, information related to involved appliances caracteristics, and tasks tat tey are able to perform, will be detected by Smart Meters and sent to a Central Unit. Users abits, i.e. ow family members usually use appliances, are monitored and sent to te Central Unit as well. Based on tis information, a profile of teir energy consumption abits, namely User Profile, will be associated to users. If, for example, te ouse is empty during working ours, it is unlely tat appliances suc as TV or ligts are turned on during tis span of time. At a later stage, information acquired and processed by te Central Unit is delivered to te appropriate Virtual Objects (VOs). As depicted in Figure 1, eac VO is responsible for managing te communication of all te appliances inside a ouse. More precisely, eac VO acts as an interface between te appliances of a single ouse and te central unit. Te role of te VO can be taken by any Smart Meter tat monitors te ouse appliances. Fig. 1. Home 1 Reference scenario Home 2 Home 3 VO Central Unit We index te appliances wit i {1,..., I} and te omes wit {1,..., H}. Eac ouse s VO, namely V O, stores te following information about appliance i, depending on wic Group it belongs to: Group 1: condition i G 1, were G1 is te set of appliances of Group 1 inside ome ; state (on/off) x (t) for appliance i related to task k, at time t (often tey are able to perform only one task); power P cons consumed by appliance i to carry on task k; probability P r (t) tat appliance i performs task k at time t, as indicated by te User Profile. Since power consumption for appliances of tis Group is negligible, we suppose a fixed energy consumption wen tey are on. Terefore, information about

3 power consumption is delivered by te Smart Meter only te first time; Group 2: condition i G 2, were G2 is te set of appliances of Group 2 inside ome ; state (on/off) x (t) for appliance i related to task k, at time t; power P cons consumed by appliance i to carry on task k; probability P r (t) tat appliance i performs task k at time t, as indicated by te User Profile or based on user needs (e.g. if te video recorder is set to turn on at time t ST and turn off at time tdl, P r (t ST t tdl ) = 1); Group 3: condition i G 3, were G3 is te set of appliances of Group 3 inside ome ; state (on/off) x (t) for appliance i related to task k, at time t; power P cons consumed by appliance i to execute task k; time t exec needed by appliance i to perform task k; deadline t DL before wic appliance i needs to perform task k; time t ST wen appliance i started to perform task k, if task k is running (i.e. x (t) = ON); Group 4: condition i G 4, were G4 is te set of appliances of Group 4 inside ome ; state (on/off) x (t) for appliance i at time t; power P prod (t) produced by appliance i at time t. Note tat te supplier can perform only one task, so k is only equal to 1. Neverteless, we keep subscript k for notation convenience. IV. TASK SCHEDULING MODEL Te proposed SHEM system is designed to perform tree basic functions: It monitors and analyses users abits wit reference to appliance usage. Based on tis information, a User Profile is created. It detects power surplus due to RES production and distributes tis power to te ouses of te same neigbourood, wit te aim of maximising its consumption. It sets te most convenient starting time of controllable appliances so tat teir tasks are executed wen it is more convenient, according to TOU tariffs and RES energy production. In order to accomplis tis function, two algoritms are developed: Te CSTS, wic scedules tasks caracterised by ig power load in off-peak times, considering te User Profile; Te RSPA, wic dynamically sifts tasks in order to maximise te use of renewable energy tat is made available by neigbours. Te sequence of steps to be performed is sown in Figure 2. As soon as appliance i placed in ome needs to start task k, it sends an activation request to V O. If appliance i is not controllable or it is not a supplier (i.e. it belongs to G 1 or G 2 ) it just needs to notify to V O tat it is canging state (x (t) = ON) for te wole duration of te task. V O sets its probability to be on to 1 accordingly. Wen task k stops, appliance i informs V O, wic sets P r (t) to its probability to turn on again, according to te User Profile. Its power Fig. 2. Task assignment steps consumption and duration values are monitored and sent to te Central Unit, wic analyses tem and updates te User Profile accordingly. If appliance i is a controllable consumer, i.e. it belongs to G 3, CSTS is started. CSTS is a centralized algoritm tat is performed by te VO to postpone te starting time t ST of G3 appliances, so tat teir tasks are executed during off-peak ours, wen electricity carge is lower. Te user can set te minimum starting time t minst and te deadline t DL wen te task needs to be carried out. Terefore, te starting time t ST is computed by te CSTS according to te user preferences, provided tat te available power P max is not exceeded by te simultaneous usage of te appliances tat made an activation request. If appliance i is a supplier (i.e. it belongs to G 4 ), or a power surplus coming from neigbouring ouses is detected by te V O, it computes te P surplus (t) value of te power surplus related to ouse at time t. P surplus (t) takes into account all te power surplus contributions tat are made available by te neigbour ouses along wit te power supplied by G 4 appliances, and it is decreased by te power consumed by te appliances inside ome if tey are on P surplus (t) = P surplus (t) i G 3,k P cons P cons i {G 1,G2 },k x (t) + P prod i G 4,k P r (t) Wenever P surplus (t) > 0 is verified, V O broadcasts tis information to te appliances it controls. If tere is any G 3 appliance tat is waiting to turn on and its power consumption is lower tan te available surplus power, (t) (1)

4 RSPA is started. RSPA is a distributed consensus algoritm were appliances compete for te same resource, negotiating among eac oter. After te algoritm as converged, tose appliances tat ave won te negotiation immediately turn on. If tere is any surplus power still available, it is sent to te closest VO. A. Cost Saving Task Sceduling algoritm Te CSTS is a centralized algoritm based on te concept tat, wenever possible, tasks tat can be postponed sould be performed during off-peak ours, wen electricity carge is lower. Wen appliance i G 3 sends to V O an activation request, it sends its deadline value t DL and its minimum starting time t minst. Consequently, V O starts CSTS to assign/reassign to all G 3 appliances te most convenient starting time according to TOU tariffs. Hence, a suitable starting time t ST in te range [t minst, t DL t exec ] is computed, provided tat te available power P max is not exceeded by te simultaneous usage of several appliances. Te optimization only takes into account consumer appliances and teir probability to be turned on. It neglects suppliers, wose power is negotiated among appliances during RSPA. Note tat it is preferable tat appliances wait for available RES power as long as it is possible, so tat electrical costs are cut. For tis reason, CSTS assigns te most convenient t ST tat is closest to tdl. Finding an optimal sceduling assignment is an NP-ard problem [14], wic complexity scales exponentially wit te problem size. In order to reduce te complexity of te algoritm, and tus its convergence time and energy needed to be run, we propose a greedy approac, wic is caracterised by a linear complexity. As described in details in Algoritm 1, te concept on te basis of CSTS is tat tose tasks tat consume more energy, i.e. tose tat present iger values of energy consumption E cons = P cons t exec, are tose tat generate more energy cost saving wen tey are sifted to off-peak ours. Terefore, tose tasks ave te priority to be sceduled for tose ours were TOU tariffs are lower, provided tat P max is not exceeded. Let Λ be te array of appliances i G 3 tat made an activation request, and E cons = (E cons ). We define a tuple Γ = (Λ, E cons ) of all te appliances tat made a request to V O and teir related energy consumption. We also define P tot (t) as te expected instant total power tat is lely to be consumed at time t by all non-controllable appliances managed by V O as P tot (t) = (t) P r (t) (2) i {G 1,G2 },k P cons P tot (t) is updated wenever te probability P r (t) canges. Te sequence of steps of CSTS is described as follows. P tot (t) is initialised wit te value of P tot (t). Te tasks of te controllable appliances tat made an activation request are ten sorted in descending order wit respect to teir energy consumption value. Starting from te task wit te igest E cons, te starting time to wic corresponds te lowest electrical energy cost C min is found. If tere is more tan one t ST tat corresponds to te C min value, te algoritm assigns te igest one. In tis way, if some P surplus (t) is available, te task as more probability to be able to negotiate to start before te assigned t ST. Te total power consumption is ten updated for te time wen te task is expected to be in execution. Algoritm 1 CSTS 1: Let P tot (t) = P tot (t) t 2: Sort in descending order Γ by its second element E cons 3: for all i Γ, k do 4: Let C min = and t ST = 5: for all t [t minst, t DL 6: if P T OT (t ) + P cons texec ten 7: Let C = t+t exec t =t C(t ) 8: if C min C ten 9: C min = C and t ST 10: end if 11: end if 12: end for 13: Let P tot (t) = P tot (t) + P cons 14: end for ] do P max t [t, t + t exec = t B. Renewable Source Power Allocation algoritm ] t [t ST, tst + texec ] Wenever V O detects some surplus power, weter it is caused by RES belonging to ome or it comes from neigbouring VOs, RSPA is started to distribute tis power to te appliances tat V O manages. In particular, since G 1 and G 2 appliances are turned on independently from te VO decisions, RSPA is run to control Γ appliances (recall from Section IV-A tat Γ is te array of controllable appliances tat made an activation request to te VO). Since surplus power value continuously cange, te algoritm needs to be as ligtweigt as possible to quickly adapt to canges. Furtermore, communication wit appliances tat are not visible from te VO needs to be quick. For tese reasons, RSPA is cosen to be a distributed algoritm, were appliances negotiate in order to reac a consensus on wic one sould turn on first. As described for Algoritm 1, wen referring to energy cost saving, tasks tat consume more power P cons ave te priority to be sceduled wen it is more convenient, i.e. wen surplus power is available. Furtermore, te priority needs to be given to tasks caracterised by closer deadlines. Calling t te current time, tasks wit closer deadlines ave iger values of te ratio t texec. t DL Summarising, if te available surplus power is sufficient, RSPA assigns it to te appliances caracterised by iger benefit values, defined as b (t) = P cons t DL t texec In order for appliances to reac a consensus on te appliance wit te igest b (t) value, a max consensus algoritm is (3)

5 used. Specifically, a Random-Broadcast-Max consensus algoritm as been cosen for its fast convergence to te solution in wireless cannels [15]. Let b max be te consensus variable and b max be te local consensus variable. Te steps of RSPA are described as follows. If some surplus power is detected, its value is broadcast by te VO to its controlled appliances tat made an activation request. Te algoritm is started by te VO sending te initial benefit value equal to 0. Wile tere is some surplus power available and tere are appliances tat can use tis power, te negotiation runs. Wenever an appliance receives a message wit surplus and benefit values, it evaluates its power consumption and benefit. If its power consumption is iger tan te available power, and if its benefit is lower tan te maximum benefit value, te appliance updates te local value of te maximum benefit and forwards te received message. Oterwise, it sets te maximum benefit value to its benefit value and broadcast te new b max value, along wit te P surplus (t) value. Wen te timeout is reaced, every appliance cecks if its local maximum benefit value corresponds to its benefit value. If it is, tis means tat it task represents te igest benefit. Te available surplus power is updated by subtracting te task power consumption, te maximum benefit value it initialised to 0 again so tat a new negotiation can start, and te task is started on te appliance tat won te negotiation. Algoritm 2 RSPA (t) > 0 is detected by V O 2: Te P surplus (t) value is broadcast by V O, along wit a benefit value b max = 0 3: wile P surplus (t) > 0 or at least one appliance can participate to te negotiation do 4: for all i Γ, k do 5: wile t < timeout do 6: Let b = b (t) 7: if i receives a message ten 1: P surplus 8: if P cons ten 9: Let b max P surplus (t) and b > b max = b 10: else 11: Let b max = b max 12: end if 13: Broadcast P surplus (t) and b max values 14: end if 15: end wile 16: if b max = b ten 17: Update P surplus (t) = P surplus (t) P cons and 18: b max = 0 values and broadcast tem 19: Start task k 20: end if 21: end for 22: end wile V. PERFORMANCE ANALYSIS Te SHEM system described in tis paper as been tested supposing to ave ouses wit random user profiles. Wit reference to TOU rates, it as been supposed to use te pricing set by te Italian electricity utility company, ENEL. Appliance operation as been simulated and controlled using Arduino Mega 2560 boards [16] wit a XBee DigiMes 2.4 radio module [17]. Power consumption values of real appliances and power production values of real RES power were considered [10]. Data traffic among appliances as been monitored using te X-CTU software [18]. Tests were run supposing to ave up to 3 appliances of group G 3 per ome, wit deadlines and minimum starting times set randomly. Results sow te energy cost savings obtained wen using te proposed SHEM system, wit respect to te case were no SHEM system is used. Since CSTS is run for all te times between t minst and t DL texec, results accuracy and computational complexity depend on ow te time is discretised, i.e. tey depend on te widt of te time slot between one time and te next one. Results are sown in Figure 3 for time slots of 10, 20, 30 and 60 minutes, wit different numbers of controlled appliances. Solid lines are used to sow results for te case were no RES are installed in te ouses (i.e. only CSTS is run). On te oter and, dased lines correspond to te case were RES are present (i.e. bot CSTS and RSPA are run). In particular, te power production of a potovoltaic system as been simulated. Te produced power as been varied randomly, up to a igest value tat is consistent wit te one of a commercial potovoltaic ome system. Fig. 3. Energy cost savings for different numbers of controllable appliances and time slot widts. Solid lines correspond to te case wit no RES. Dased line correspond to te case wit RES It is evident tat cost saving is muc iger wen RES are present: it is from 30% to 48% iger wit respect to te case witout RES. Note tat cost savings increase wit te number of controlled appliances. However, te slope is steeper wen te appliances are fewer, particularly for results wit iger accuracy, i.e. 10 minutes time slots. Tis is because te iger te number of controlled appliances, te lower te power still available, and tus te more difficult te sceduling of all te tasks in off-peak ours. Time slot widts are critical wit reference to results accuracy: a time slot of 10 minutes results in an energy cost saving of about 10% more tan tat corresponding to a time slot of 60 minutes. Tis would lead to te conclusion tat narrower time slots are

6 preferable. However, narrower time slots correspond to iger computational complexity. As demonstrated by Figure 4, CSTS complexity srinks exponentially wit te increment of time slot widts, and complexity for a time slot of less tan 10 minutes migt become proibitive for large numbers of controlled appliances. Note tat, wit a time slot during 20 minutes, complexity decreases by 70%, wit a loss in cost saving of just 2 3%. For tis reason, for VOs caracterised by low computational complexity, increasing te time slot widt represents a good trade-off. Fig. 4. Complexity of CSTS considering te ATmega2560 microprocessor of te Arduino used in tese tests Wit reference to te RSPA algoritm, it is a distributed mecanism were eac node only make a few comparisons after receiving update messages. For tis reason, computational complexity is considered negligible and it is not furter analysed. VI. CONCLUSIONS In tis paper a SHEM system is proposed. A sceduling model for tasks of controllable appliances tat aims to reduce electricity costs is described. In particular, two algoritms are provided: te former, te CSTS, based on te presence of TOU tariffs, sifts te starting time of controllable appliance tasks in off-peak times, taking into account te User Profile, i.e. ow te user is expected to use te oter appliances in te ouse. Te latter, RSPA, is started wenever a RES installed in a ouse in te neigbourood make some power available. In tis case, te appliances dynamically negotiate in order to sare te available power and start teir tasks before te starting time assigned by te CSTS algoritm. Tests performed using Arduino Mega 2560 boards prove tat energy cost saving using te proposed SHEM system goes from 35% to 65%, wit reference to te case were tasks are started as soon as tey are programmed. Better results are acieved for larger numbers of controlled appliances, altoug te slope of improvement is steeper for fewer devices. It is important to note tat energy cost savings are strictly related to user abits, wit reference to teir awareness in te appliance usage. In tis sense, SHEM systems are crucial for te automation of tis process and to unburden te user from te manual control of controllable appliances. REFERENCES [1] L. Atzori, A. Iera, and G. Morabito, Te internet of tings: A survey, Computer networks, vol. 54, no. 15, pp. 2787 2805, 2010. [2] D.-M. Han and J.-H. Lim, Smart ome energy management system using ieee 802.15. 4 and zigbee, Consumer Electronics, IEEE Transactions on, vol. 56, no. 3, pp. 1403 1410, 2010. [3] N. Lior, Sustainable energy development: te present (2009) situation and possible pats to te future, Energy, vol. 35, no. 10, pp. 3976 3994, 2010. [4] M. Jan, M. Jentsc, C. R. Prause, F. Pramudianto, A. Al-Akkad, and R. Reiners, Te energy aware smart ome, in Future Information Tecnology (FutureTec), 2010 5t International Conference on. IEEE, 2010, pp. 1 8. [5] J. Torriti, Price-based demand side management: Assessing te impacts of time-of-use tariffs on residential electricity demand and peak sifting in nortern italy, Energy, vol. 44, no. 1, pp. 576 583, 2012. [6] F. Baig, A. Mamood, N. Javaid, S. Razzaq, N. Kan, and Z. Saleem, Smart ome energy management system for monitoring and sceduling of ome appliances using zigbee, J. Basic. Appl. Sci. Res, vol. 3, no. 5, pp. 880 891, 2013. [7] D.-M. Han and J.-H. Lim, Design and implementation of smart ome energy management systems based on zigbee, Consumer Electronics, IEEE Transactions on, vol. 56, no. 3, pp. 1417 1425, 2010. [8] J. A. Nazabal, F. J. Falcone, C. Fernandez-Valdivielso, and I. R. Matias, Energy management system proposal for efficient smart omes, in New Concepts in Smart Cities: Fostering Public and Private Alliances (SmartMILE), 2013 International Conference on. IEEE, 2013, pp. 1 5. [9] A. Tascaraoglu, A. Boynuegri, and M. Uzunoglu, A demand side management strategy based on forecasting of residential renewable sources: A smart ome system in turkey, Energy and Buildings, 2014. [10] N. C. Truong, J. McInerney, L. Tran-Tan, E. Costanza, and S. D. Ramcurn, Forecasting multi-appliance usage for smart ome energy management, in Proceedings of te Twenty-Tird international joint conference on Artificial Intelligence, 2013, pp. 2908 2914. [11] J. Gil-Quijano and N. Sabouret, Prediction of umans activity for learning te beaviors of electrical appliances in an intelligent ambient environment, in Web Intelligence and Intelligent Agent Tecnology (WI- IAT), 2010 IEEE/WIC/ACM International Conference on, vol. 2. IEEE, 2010, pp. 283 286. [12] A. Kasif, X. H. B. Le, J. Dugdale, and S. Ploix, Agent based framework to simulate inabitants beaviour in domestic settings for energy management. in ICAART (2), 2011, pp. 190 199. [13] J. Z. Kolter and J. Ferreira, A large-scale study on predicting and contextualizing building energy usage. in AAAI, 2011. [14] V. Pilloni and L. Atzori, Deployment of distributed applications in wireless sensor networks, Sensors, vol. 11, no. 8, pp. 7395 7419, 2011. [15] F. Iutzeler, P. Ciblat, J. Jakubowicz et al., Analysis of max-consensus algoritms in wireless cannels, IEEE Transactions on Signal Processing, vol. 60, no. 11, pp. 6103 6107, 2012. [16] Arduino mega 2560, ttp://arduino.cc/en/main/ ArduinoBoardMega2560. [17] Digi, Xbee digimes 2.4, www.digi.com/products/ wireless-wired-embedded-solutions/zigbee-rf-modules/ zigbee-mes-module/xbee-digimes-2-4. [18] Digi, X-ctu software, ttp://www.digi.com/support/kbase/kbaseresultdetl?id=2125.