Situation-Aware Energy Control by Combining Simple Sensors and Complex Event Processing Leonard Renners and Ralf Bruns and Jürgen Dunkel 1 Abstract. In recent years, multiple efforts for reducing energy us- To achieve these goals, our approach uses two different models: age have been proposed. Especially buildings offer high potentials for energy savings. In this paper, we present a novel approach for in- 1. A Semantic Domain Model describes the expert knowledge about telligent energy control that combines a simple infrastructure using the domain, i.e. in our case the structure of the building and its low cost sensors with the reasoning capabilities of Complex Event expected usage. Processing. The key issues of the approach are a sophisticated se- 2. An Event Model defines all those events occurring in the building mantic domain model and a multi-staged event processing architec- that are relevant for energy control. Different event sources can be ture leading to an intelligent, situation-aware energy management distinguished: low cost sensors yield information about the cur- system. rent incidents in a building. Furthermore, domain-specific events are created by various knowledge sources, like heating plans or lectures schedules. 1 Introduction In recent years, global warming, greenhouse effect, as well as escalat- To react on relevant situations in real-time we apply Complex ing energy prices have lead to enhanced efforts for reducing energy Event Processing (CEP) that has been proposed as a new archi- usage and carbon emission. Especially, buildings offer high poten- tectural paradigm for in-memory processing continuous streams of tials in reducing energy consumption: electric lighting, heating and events. CEP is based on declarative rules to describe event patterns, air conditioning are highly energy-intensive and not well-adjusted to which are applied to the event stream to identify relevant situations actual usage necessities. in a certain domain. In many buildings, there are already first approaches of more ef- The integration of the Semantic Domain Model with the Event fective control mechanisms to improve energy consumption. Typi- Model allows the enrichment of event processing rules with domain cally, light in public areas like corridors is controlled by using motion knowledge. It is a key issue of our approach and provides an intelli- sensors instead of classic switches, preventing unnecessarily turned gent sensor data processing, leading to a reasonable, situation-aware on lamps. On the contrary, heating is mostly controlled on base of heating and energy management. fixed heating plans that determine the heating period by a predefined The remainder of the paper is organized as follows. The next sec- timetable. Often the heating schedule is not related to single rooms tion 2 describes an application scenario for our approach. In section but to the entire building or larger parts of it (e.g. total floors). Over- 3 the related work is discussed. In the subsequent section 4 we intro- all, this leads to the situation that many rooms are heated although duce our approach using CEP techniques to implement an intelligent they are not in use, thus causing a huge waste of energy. In summary, energy management. Section 5 shows an example set-up and yields we can conclude that in general energy control of buildings is not an evaluation. The final section 6 contains some concluding remarks situation-aware. and provides an outlook to future directions of research. In the following we present an approach for intelligent energy control that combines a simple infrastructure using cheap sensors 2 Scenario: Energy Control in Buildings with event stream processing of the plain sensor data. Our approach should reach the following goals: We selected the energy management of buildings as our application scenario. In particular, the scenario helps to discuss in some more de- (a) Individual energy control for every single room (instead of con- tail the benefits that intelligent energy control provides. As already sidering the entire building). mentioned in the introduction, buildings have high potentials in im- (b) Low-cost solution by utilizing the existing infrastructure as well proving energy efficiency, since they have many energy consumer as cheap and simple sensors (as opposed to new, sophisticated units that are often unnecessarily turned on. In the following, we will and expensive sensors). use an university campus as a concrete example of our approach. (c) Situation awareness: the energy consumption should be con- Universities exhibit the following features that are common to most trolled according to actual usage (in contrast to predefined and public buildings and which will influence energy management sig- fixed schedules). nificantly: (d) Proactive control: the control mechanisms should exploit the knowledge about normal room occupancy, for instance based on • Various room types with different energy consumption profiles room schedules and normal user behavior patterns. can be distinguished: For instance, server rooms have to be air 1 Hannover University of Applied Sciences and Arts, Germany, email: conditioned below 20 degrees Celsius. Instead, offices and lecture forename.surname@fh-hannover.de rooms must be heated to achieve a temperature above 22 degrees, Workshop on AI Problems and Approaches for Intelligent Environments (AI@IE 2012) 29 but normal storage room must be neither cooled nor heated. In the hand first approaches report the employment of event processing following, we will focus on office and lecture rooms. technologies in energy management, but do not target the same field • Non-uniform usage profiles: Each room exhibits its individual oc- and way of application. cupancy depending on the room type and the specific behavior However, none of them is presenting a comprehensive approach of of different user groups. For instance, lecture rooms are occupied real-time situational awareness for the whole energy management, according to a prefixed schedule that might be changed only ex- based on the integration of an extensive semantic model of the ap- ceptionally. Instead, offices or cube farms are used according to plication domain in the event processing reasoning process. In par- the personal behavior of their occupants including absences due ticular, the investigated control of the heating process requires more to vacation times or illness. sophisticated semantic models and more advanced event processing • Spontaneous occupancies: Furthermore, rooms are spontaneously rules. and individually used, e.g. due to ad-hoc meetings, rescheduled lectures or unplanned project work. Note that these individual oc- 4 Intelligent Energy Control Using CEP cupancies are inherently unpredictable. In this section we present our approach for a situation-aware en- To deal with non-uniform and spontaneous usage of rooms, which ergy control by applying intelligent sensor data processing on simple deviates significantly from predicted average occupancies, an intelli- buildings. After giving a short overview of Complex Event Process- gent and situation-aware energy control mechanism is required. ing we present our general software architecture and a Domain Event On the one hand, expert knowledge about the building and the Model that integrates knowledge about domain specific concepts and behavior of its users should be exploited for adjusting the energy events. Finally, we illustrate the intelligent reasoning part of our ap- control to realistic usage patterns. proach by showing some event processing rules. On the other hand, actual behavior must be monitored to react ad- equately on spontaneous and individual usage actions. Especially, if 4.1 CEP Overview unexpected occupancies or periods of absence are detected in a cer- tain room, the corresponding energy consumption units (like heating Complex Event Processing (CEP) is a software architectural ap- and lighting) can be switched on or off, respectively. proach for processing continuous streams of high volumes of events To provide an individual control, we have to observe the incidents in real-time [10]. Everything that happens can be considered as an and states in every single room. To achieve this goal we will incor- event. A corresponding event object carries general metadata (event porate already installed sensors in the building as well as we will ID, timestamp) and event-specific information, e.g. a sensor ID and equip the rooms with simple and cheap sensors: motion sensors for the measured temperature. Note that single events have no special detecting movements, temperature sensors to measure the heating, meaning, but must be correlated. CEP analyses continuous streams and contact switches that register when a door or window is opened of incoming events in order to identify the presence of complex se- or closed, respectively. quences of events, so called event patterns. A pattern match signifies a meaningful state of the environment and causes either creating a new complex event or triggering an ap- 3 Related Work propriate action. Fundamental concepts of CEP are an event processing language Our approach is based on the exploitation of fine-grained sensor data (EPL), to express event processing rules consisting of event patterns emitted by networks of simple sensors in buildings. Sensor networks and actions, as well as an event processing engine that continuously possess intrinsic problem properties that are perfectly addressed by analyses the event stream and executes the matching rules.2 Com- complex event processing. Several published approaches prove the plex event processing and event-driven systems generally have the suitability of event processing for sensor networks (e.g. [8, 11]). following basic characteristics: Determining the current status of usage is an important topic in smart homes and intelligent facility management systems. Several • Continuous in-memory processing: CEP is designed to handle a approaches have shown how occupancy detection can be used to consecutive input stream of events and in-memory processing en- implement more effective and powerful behaviour [1, 2]. Other ap- ables real-time operations. proaches put the main focus on occupancy and movement prediction • Correlating Data: It enables the combination of different events algorithms instead of real-time reactions [6, 5, 9]. from distinct sources including additional domain knowledge. There is also increasing interest using CEP technologies for en- Event processing rules transform fine-grained simple events into ergy management. Holland-Moritz and Vandenhouten examined dif- complex (business) events that represent a significant meaning for ferent solutions for an intelligent management system (in general) the application domain. and identified CEP as one very suitable and suggestive concept [7]. • Temporal Operators: Within event stream processing, timer func- Xu et al. introduced a CEP approach with ontology and semantics tionalities as well as sliding time windows can be used to define supporting occupancy detection for an intelligent light management event patterns representing temporal relationships. system [12, 14]. • Distributed Event Processing: Event processing can be distributed Another approach in a similar direction was made by Wen et al. on several rule engines (physically or logically). Thereby scalabil- within their industrial experience report about using CEP for energy ity and the separation of different functionalities can be realized. and operation management, while focussing on predictive elements and adaptable behavior [13]. 4.2 Event Processing Architecture In summary, on the one hand some approaches address energy Luckham introduced the concept of event processing agents control adapted to the current occupancy/usage, but do not rely on (EPA) [10]. An EPA is a software component specialized on event event processing techniques, thus not taking advantage of features like in-memory processing and real-time capabilities. On the other 2 Sample open source CEP engines are Esper and Drools Fusion. 30 Workshop on AI Problems and Approaches for Intelligent Environments (AI@IE 2012) stream processing with its own rule engine and rule base. An event • Domain Agent: The cleaned sensor events contain only low-level processing network (EPN) connects several EPAs to constitute a soft- technical information, e.g. sensor IDs that have no specific mean- ware architecture for event processing. Event processing agents com- ing in the application domain, and are often incomplete for fur- municate with each other by exchanging events. ther processing. Therefore, they should be transformed to domain EPAs provide an approach for modularizing and structuring rules: events by mapping plain sensor event data to domain concepts. For Light-weighted agents with few rules fulfill a coherent domain- instance, a measured temperature should be related to a certain specific task and improve comprehensibility and maintainability. room and to the desired temperature of the corresponding room Furthermore, distributing the EPAs on different computing nodes en- type. The information necessary for this content enrichment step hances system performance and scalability [3]. is retrieved from the backend systems. The Domain Agent trans- Thus, the event-driven architecture of our energy control system is forms cleaned sensor events into enriched domain events and for- based on a multi-staged EPN for structuring and organizing the event wards them to the Situation Agent. processing rules. Figure 1 depicts the different EPAs and illustrates • Situation Agent: In a diagnosis step various domain events are syn- the flow of events: thesized to a new (complex) situation event that characterizes a particular state of the building. For example, contactSensor events and movement events are correlated to a new roomOccupied event signalizing that somebody is staying in a certain room. In sum- mary, the Situation Agent processes a correlation step to create new types of complex events that are propagated to the Energy Control Agent. • Energy Control Agent: Finally, the situation diagnosed from the stream of sensor events must be correlated with the information received from the general knowledge sources. The Energy Control Agent emits an action event to trigger a certain control action that reacts appropriately on the actual state of the building. For instance, lectureStart events emitted by a calendar are com- bined with roomOccupied events generated by the the Situation Agent to trigger an appropriate control actions by creating an ac- Figure 1. Event Processing Network (EPN) for Energy Management tion event of type increaseTemperature. Event Sinks: The backend systems of the building management Event Sources: We can distinguish different types of event serve as event sinks of the events produced by the Energy Control sources that correspond to the information that is used by our energy Agent. Figure 1 shows two examples: an increaseTemperature event control system (as already mentioned in section 2). could be sent to the energy control system to change directly the heat- ing of a certain room. As another example, we can consider a reserve- • General knowledge sources: There are general knowledge sources Room event that could be forwarded to the building management sys- that can emit application-specific events relevant for the build- tem for generating automatically an entry into the occupancy plan of ings energy management. For instance, a calendar containing the the corresponding room. lecture schedules might create lectureStart events that signal the scheduled starting time of a teaching session. • Sensors: Low cost sensors as described in section 2 are used to 4.3 Domain Event Model monitor the incidents in the building. For instance, motion sensors A main contribution of our approach is integrating general domain and temperature sensors emit movement events as well as temper- knowledge with sensor events in a Domain Event Model. Figure 2 ature events. The contact switches produce contactSensor events shows the general structure of the Domain Event Model that distin- that signal if doors or windows are opened or closed, respectively. guishes two dimensions, and thus yielding four different quadrants: Event Processing Network (EPN): Event processing is nothing else than event transformation: the simple events emitted by the event sources are transformed into more abstract application-specific events for inferring appropriate control steps. The event transforma- tions are processed by the EPAs depicted in figure 1. • Cleaning/Filtering Agent: Due to technical problems, sensor data Figure 2. Structure of the Domain Event Model is often inconsistent: e.g. duplicated readings or outliers must be compensated. Therefore, in a cleaning step all sensor events have • The World Model describes the structural or static concepts re- to be pre-processed to overcome inconsistencies [3]. garded in the system: First, it defines the domain concepts like Furthermore, not all events are required in subsequent processing buildings, rooms or class schedules. Secondly, it defines the sen- stages. For instance, motion sensors may emit many movement sor infrastructure the building is equipped with. For instance, what events within a small time interval, which are related to the same different kind of sensors are used and where they are installed. incident. Therefore irrelevant events are filtered out to reduce the • The Event Model defines the dynamic aspects of the system, i.e. total number of events. Using various processing rules, the Clean- all types of events that are considered in the system [4]. First, all ing/Filtering Agent forwards cleaned sensor events events to the sensor events emitted by the different sensor types are described. Domain Agent. Secondly, it considers all domain events in the system: On the one Workshop on AI Problems and Approaches for Intelligent Environments (AI@IE 2012) 31 hand, these are the application events that are generated by CEP Operators rules and have a certain meaning in the application domain. For AND Combination of events or constraints instance, that a room is occupied for a certain time. Furthermore, NOT Negation of a constraint it defines context events, that are produced by general knowledge -> Followed-by operator. Sequence of conditions. sources, like calendar applications containing room schedules. Timer Timer(time) defines a time to wait Timer.at(daytime) is a specific (optionally peri- Of course, there are interrelations between the concepts of the dif- odic) point of time. ferent model parts. For instance, the sensor concept ’contact sensor’ .within defines a time window for an event in which the is related to the domain concept ’room’ specifying the concrete posi- event has to happen to be considered. tion of a certain sensor. Figure 3 shows an excerpt of the Domain Event Model for our The following two rules are part of the rule base of the Situation university building scenario. Note that for simplicity and clarity, no Agent (see Figure 1) and produce a situation event. Note that these attributes are depicted in the diagram. rules detect a certain situation in the building that can be exploited in different application domains. Here, we show how the Energy Con- The Model of the Domain Concepts (upper left quadrant) describes trol Agent can use identified situations to derive energy control ac- the hierarchy of different types of rooms in an university, such as tions. But also other kind or agents, for instance Security Agents can course rooms, offices, and server rooms. Course rooms can be fur- make use of situation events for detecting security risk or incidents. ther refined into lecture rooms, labs, etc. Other concepts modeled The first rule produces a situation event of type RoomOccupiedE- in figure 3 are heating plans that are related to each room and class vent indicating that a certain room is currently occupied. schedules for each course room. Neighbouring rooms are specified by the adjacent relationship. Furthermore, the model defines room rule: "room occupied" equipment as windows and doors.3 CONDITION DoorOpenEvent AS d -> The Model of the Sensors (lower left quadrant) specifies the differ- Timer(5 minutes) -> ent types of sensors installed in the building. Note that in figure 3 MovementEvent AS m sensor characteristics like measured variables, and quality of mea- AND (d.room = m.room) sures, like availability or accuracy, are omitted for clarity. However, ACTION new RoomOccupiedEvent(d.room) the model represents the location of the sensors by relating them to a physical item of the domain model. A room is assumed to be occupied if the door is opened and five The Model of the Sensor Events (lower right quadrant) defines the minutes later a movement is still observed. The delay will prevent types of sensor events. Depending on the specific type of a sensor false positives, like only cleaning the room or just quickly picking up different data can be produced. For instance, a contact sensor might some things. Note that the rule correlates two sensor events to derive produce data for signalizing that a door has been opened. Further- a new complex event of type situation event with a new application- more, each sensor event is related by the sensed-by relationship to specific meaning. a corresponding sensor, and thereby to a certain position. Note that The next rule considers the opposite situation: a room is not oc- this relation shows the connection between the world and the event cupied if the door is closed and there is no movement within the model. following 10 minutes. The Model of the Domain Events (upper right quadrant) presents all rule: "room not occupied" application events considered in the system. On the one hand, there CONDITION DoorCloseEvent AS d AND are the context events that are produced directly by software compo- NOT MovementEvent. nents. Figure 3 defines heatingStart events and lectureStart events as within(10 minutes) AS m specific examples of context events, which might be produced by a AND (d.room = m.room) heating plan and a class schedule, respectively. On the other hand, ACTION new RoomNotOccupiedEvent(d.room) situation events signal that a certain situation has occurred in the building, e.g. a room has been occupied or freed. The following rules reside in the rule base of the Energy Control Furthermore, action events are considered like increaseTempera- Agent (see Figure 1) and correlate situation events to derive some ture or reserveRoom events. These events trigger some actions in the action events triggering some reactions in the backend system. backend system, e.g. turning up the heating. They are produced by The first occupancy per day of an office is of special importance, CEP rules that might correlate situation events and context events. since from then on the office is in use and needs to reach its operat- ing temperature. Before that, room temperature could be a bit lower 4.4 Event Processing Rules yielding a reduction of the heating costs. In the following, we present some exemplary rules in a pseudo rule: "first usage" language to provide a better understanding of the intelligent CONDITION Timer.at(06:00 AM) -> reasoning capabilities our approach. An event processing rule NOT RoomOccupiedEvent AS n -> contains of two parts: a condition part describing the require- RoomOccupiedEvent AS r ments for the rule to fire and an action part to be performed AND (r.room.type = office) if the requirements are matched. The condition is defined by AND (n.room = r.room) an event pattern using several operators and further constraints. ACTION IncreaseTemp(r.room) 3 Note that this is not a very sophisticated model: many aspects are not shown, The rule considers the situation in a certain room after 6:00 AM. e.g. different user types and their behavior defined by working times and the If then a RoomOccupied event r occurs and there was no other usually used rooms. RoomOccupied event n (between 6:00 and the occurrence of event 32 Workshop on AI Problems and Approaches for Intelligent Environments (AI@IE 2012) Figure 3. Domain Event Model r) then the room is used for the first time that day and the tempera- rule: "course room not used for more ture should be increased. than 1 hour" Another situation of interest is the ’final’ absence of an employee. CONDITION NOT RoomOccupiedEvent. A shorter break, e.g. having lunch or a meeting, should not have the within(60 minutes) as r effect of cooling down the employees office. But after the typical end AND (r.room.type = course) of the workday the probability that the room will be in use again is ACTION lowerTemp(r.room) very low. Therefore, the heating can now be lowered to reduce the energy consumption. If a course room is not used for at least 60 minutes, then the temper- ature of the room can be decreased. This rule will match for a lecture rule: "after hour" room, but not for other rooms as offices. CONDITION Timer.at(06:00 PM) -> Note that we will investigate the modeling of much more semantic RoomNotOccupiedEvent AS r relationships using an appropriate formalism in further researches. AND (r.room.type = office) For instance with OWL, relationships between concepts can be de- ACTION LowerTemp(r.room) scribed much more precisely. For OWL object properties ranges and domains can be specified as well as further property characteristics If after 6:00 PM a RoomNotOccupied event is captured in an office (as transitivity, symmetry or reflexivity). This information can be ex- room the temperature will be reduced. ploited by more sophisticated reasoning, for instance by using C- The next rule illustrates how context events from general knowl- SPARQL query language. edge sources and sensor events are correlated to derive an action event. In particular, the rule describes the situation that though a lec- ture is scheduled, the lecture room is not occupied. 5 Case Study / Evaluation rule: "planned, but not used" We have equipped one room of our university building with a sam- CONDITION LectureStartEvent AS l -> ple setup of different physical sensors and implemented a prototype NOT RoomOccupiedEvent. of our event-driven energy control system. As sensor hardware we within(15 minutes) AS r used a Phidget4 Interface-Kit and corresponding motion sensor and AND (l.room = r.room) contact switches. ACTION lowerTemp(r.room) The event processing is implemented with the open source CEP engine Esper 5 . Esper provides the essential features of typical CEP The rule will fire if a LectureStart event is captured for a certain systems like time windows, external method calls, and event pattern room, but within the following 15 minutes no RoomOccupied event operators. The event processing rules are defined in a SQL-like rule occurs. This leads to the assumption that the lecture will not take language , the so-called Continuous Query Language (CQL). In con- place and accordingly the room will not be in use and the temperature trast to SQL the CQL queries are not executed on a Database, but can be lowered to the idle state. directly in-memory on the continuously arriving event stream. Finally, we present a simple rule that exploits further domain- Our experimental evaluation has proven the capabilities of CEP specific knowledge in the event reasoning. Several semantic rela- to correlate the sensor data and achieve a real-time analysis and tionships are represented in the Domain Event Model, which can be reaction based on the sensor data stream and event patterns. exploited to enhance the reasoning capabilities of event processing. For example, lecture rooms, seminar rooms and laboratories are all Since we could only realize an example installation for one dis- of type course room as specified by a ’is-a’ relationship in the Do- tinct room and only including sensors and no actuators, the useful- main Event Model. The semantical meaning of the ’is-a’ relationship can be used in event processing: A rule for course rooms is implicitly 4 http://www.phidgets.com valid for all subtypes as well. 5 http://esper.codehaus.org Workshop on AI Problems and Approaches for Intelligent Environments (AI@IE 2012) 33 ness of our approach is evaluated on the basis of assumptions about Directions of future research are, among others, the further en- the real usage. We assume a typical heating behavior in a static man- hancement of the semantic Domain Event Model as well as the de- ner starting heating at 6 AM until 9 PM. The typical workday of an velopment of advanced concepts for the incorporation of the seman- university lecturer (as an example of a non-uniform user type) may tic knowledge in event processing languages. Furthermore, develop- be structured as followed: start of work at 8 AM, lecture between 10 ment towards an automated rule creation, by the means of removing and 11:30 AM, lunch break between 11:30 AM and 12 PM, exercise the necessities to hand-code the scenarios, could be considered. lesson between 12 and 1:30 PM, and end of work at 5:30 PM. Figure 4 visualizes the different heating behaviors by example of a 7 Acknowledgement lecturer’s office room: (a) dynamic heating with our approach based on situational awareness compared to (b) the static solution with a This work was supported in part by the European Community (Eu- fixed heating plan. The Human Presence depicts the occupancy of ropäischer Fonds für regionale Entwicklung – EFRE) under Research Grant EFRE Nr.W2-80115112. REFERENCES [1] Y. Agarwal, B. Balaji, R. Gupta, J. Lyles, M. Wei, and T. Weng, ‘Occupancy-driven energy management for smart building automa- tion’, in Proceedings of the 2nd ACM Workshop on Embedded Sensing Systems for Energy-Efficiency in Building (BuildSys). ACM, (2010). [2] J. C. Augusto and C. D. Nugent, ‘The use of temporal reasoning and management of complex events in smart homes’, in Proceedings of the 16th Eureopean Conference on Artificial Intelligence (ECAI), pp. 778– 782. IOS Press, (2004). [3] R. Bruns and J. Dunkel, Event-Driven Architecture: Softwarearchitek- tur für ereignisgesteuerte Geschäftsprozesse, Springer-Verlag, Berlin Heidelberg, 2010. [4] J. Dunkel, A. Fernández, R. Ortiz, and S. Ossowski, ‘Injecting seman- tics into event-driven architectures’, in Proceedings of the 11th Inter- Figure 4. Static versus dynamic heating national Conference on Enterprise Information Systems (ICEIS), pp. 70–75. Springer, (2009). the room according to the typical workday defined above. The two [5] Varick L. Erickson, Yiqing Lin, Ankur Kamthe, Rohini Brahme, Amit Surana, Alberto E. Cerpa, Michael D. Sohn, and Satish Narayanan, ‘En- curves describe the heating level on the y-axes with respect to the ergy efficient building environment control strategies using real-time time. occupancy measurements’, in Proceedings of the First ACM Workshop As can be seen, the biggest differences, and therefore energy sav- on Embedded Sensing Systems for Energy-Efficiency in Buildings, pp. ings, appear during the time before and after the workday. Notice that 19–24. ACM, (2009). [6] Ge Gao and Kamin Whitehouse, ‘The self-programming thermostat: the usual hours of work may differ from lecturer to lecturer and thus optimizing setback schedules based on home occupancy patterns’, in the static schedule can not be fitted to the typical behavior of one Proceedings of the First ACM Workshop on Embedded Sensing Systems lecturer. for Energy-Efficiency in Buildings, pp. 67–72. ACM, (2009). In contrast, our approach provides a room-specific and situation- [7] R. Holland-Moritz and R. Vandenhouten, ‘A flexible architecture for in- aware control mechanism enabling a precise energy management that telligent management systems’, in Proceedings of the 3rd International Symposium on Logistics and Industrial Informatics (LINDI), pp. 83–86. additionally enables the heater to turn lower during temporary ab- IEEE Computer Society, (2011). sence. In order to keep the room in a comfortable state, if the user [8] S. R. Jeffery, G. Alonso, M. J. Franklin, and J. Widom W. Hong, ‘A returns, the heaters level is only decreased and not completely turned pipelined framework for online cleaning of sensor data streams’, in Pro- off for temporarily unoccupied rooms. ceedings of the International Conference on Data Engineering (ICDE), p. 140, (2006). Based on these assumptions a calculation comparing the two dif- [9] J. Landay, Y. Shi, D. J. Patterson, Y. Rogers, X. Xie, J. Scott, A.J. Bern- ferent heating behaviors (static versus dynamic) results in a heating heim Brush, J. Krumm, B. Meyers, M. Hazas, S. Hodges, and N. Vil- reduction up to 30% and, accordingly, lower energy consumption and lar, ‘Preheat: Controlling home heating using occupancy prediction’, in carbon emission. Proceedings of the 13th International Conference on Ubiquitous Com- puting (UbiComp), pp. 281–290. ACM, (2011). [10] D. C. Luckham, The Power of Events: An Introduction to Complex 6 Conclusion Event Processing in Distributed Enterprise Systems, Addison-Wesley, 2002. In this paper, a novel approach for intelligent energy management [11] S. Selvakennedy, U. Rohm, and B. Scholz, ‘Event processing middle- ware for wireless sensor networks’, in Proceedings of the International by means of complex event processing and simple sensors has been Conference on Parallel Processing Workshops (ICPPW), p. 65, (2007). presented. The approach is different from other approaches in that [12] N. Stojanovic, D. Milenovic, Y. Xu, L. Stojanovic, D. Anicic, and is based on a sophisticated representation of domain as well as sen- R. Studer, ‘An intelligent event-driven approach for efficient energy sor knowledge and a multi-staged event processing architecture. By consumption in commercial buildings: Smart office use case’, in Pro- the integration of domain knowledge and semantic information into ceedings of the International Conference on Distributed Event-based System, pp. 303–312. ACM, (2011). to the reasoning process we achieve an intelligent, situation-aware [13] J. Y. C. Wen, G. Y. Lin, T. Sung, M. Liang, G. Tsai, and M. W. Feng, ‘A behavior. complex event processing architecture for energy and operation man- The approach allows an individualized, situation-aware energy agement’, in Proceedings of the 5th International Conference on Dis- management of buildings according to the current occupancy status tributed Event-Based Systems (DEBS), pp. 313–316. ACM, (2011). [14] Y. Xu, N. Stojanovic, L. Stojanovic, D. Anicic, and R. Studer, ‘An ap- of the separate rooms. By means of complex event processing an ex- proach for more efficient energy consumption based on real-time situa- isting infrastructure with everyday sensors can be expanded into an tional awareness’, in The Semanic Web: Research and Applications, pp. intelligent environment. 270–284. Springer, (2011). 34 Workshop on AI Problems and Approaches for Intelligent Environments (AI@IE 2012)