Adaptive Knowledge Representation for a
Self-Managing Home Energy Usage System
Martijn Warnier
1
, Marten van Sinderen
2
and Frances M.T. Brazier
1
1
Faculty Technology, Policy and Management
Delft University of Technology, The Netherlands
2
Faculty of Electrical Engineering, Mathematics and Computer Science
University of Twente, The Netherlands
M.E.Warnier@tudelft.nl,
m.j.vansinderen@ewi.utwente.nl,F.M.Brazier@tudelft.nl
Abstract. Automated and efficient energy management has many potential ben-
efits for producers and consumers of energy, and the environment. Focusing on
energy management on the consumer side, this paper considers two forms of en-
ergy management: minimizing energy usage in single households and avoiding
peaks in energy consumption in a larger area.
A combination of context aware and autonomic computing is used to describe
an automated and self-managing system that, by analyzing context information
and adapting to its environment, can learn the behavior of household occupants.
Based on this information, together with user defined policies, energy usage is
lowered by selectively powering down devices. By powering specific thermostat-
ically controlled devices on or off energy can also be redistributed over time. This
is utilized to avoid global peaks in energy usage.
The self-managing system reasons about context and other information and acts
when required. This information is the knowledge with which it can adaptively
reason, about to take to ensure efficient energy usage. This paper explores the
requirements that hold for representing this knowledge and how the knowledge
base can continuously and adaptively be updated: to be self-managing.
1 Introduction
Efficient energy management forms an important challenge in today’s society as con-
ventional energy sources become more and more scarce and more eco-friendly alterna-
tives are not yet evaluated at a large scale. Automated approaches to efficient energy
management are currently still limited and mostly used by large power consumers such
as factories [10, 16, 17], but with the advancement of the Smart Grid [2] and other re-
cent advancements in sensor networks [1, 19], such as cognitive networks [18], this
changes rapidly.
This paper addresses this challenge at two different, and possibly conflicting, levels:
(i) at the level of a single household, where the goal is to lower energy consumption and
(ii) at the level of group of households (a neighbourhood), where the goal is to lower
peak usage in energy consumption. In [15] a new approach based on autonomic and
context aware computing is introduced. This paper extends that approach. It explores
Warnier M., van Sinderen M. and van Sinderen M.
Adaptive Knowledge Representation for a Self-Managing Home Energy Usage System.
DOI: 10.5220/0004465901310140
In Proceedings of the 4th International Workshop on Enterprise Systems and Technology (I-WEST 2010), pages 131-140
ISBN: 978-989-8425-44-7
Copyright
c
2010 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
the requirements for an adaptive knowledge representation in the context of a home
energy management system. The system exploits sensor information to monitor electric
appliances and their surroundings. Based on this environmental information the status
of appliances is updated, influencing energy consumption.
The core of the home energy management system is formed by a rule based sys-
tem [5]. Such systems are typically very deterministic in nature, always producing
the same outcomes for the same input: a more dynamic approach is required for self-
management. Therefore, the main challenge addressed in this paper is the question how,
in the context of a home energy management system, a knowledge base can continu-
ously and adaptively be updated: to become self-managing.
The remainder of this paper is organized as follows: in the next section the ap-
proach introduced in [15] is summarized. Section 3 discusses requirements for the self-
managing home energy system and Section 4 outlines an approach to adaptive and
self-managing knowledge representation. The paper ends with a brief discussion and
conclusions.
2 An Autonomic and Context Aware Home Energy Management
System
A new green computing approach on the consumer side (demand side management) is
proposed in [15]. The proposed system considers two forms of energy management:
minimizing energy usage in single household and avoiding peaks in energy usage for a
larger residential area. A basic architecture is proposed to achieve this goal. A service
oriented framework, based on the complementary approaches of autonomic comput-
ing and context aware computing, is introduced. Context information is continuously
used to analyze the energy requirements of a household. Based on this information, the
home energy management system determines whether and how to influence the energy
consumption of individual devices. By selectively powering thermostatically controlled
devices –such as fridges, air conditioning units, electrical heating– on or off, energy
consumption is redistributed over time [12, 15] avoiding peaks in energy usage. Such
devices make up around 25% of the total energy consumption in the USA [9].
Residences of the household can add their own preferred energy usage profile. The
resulting service oriented framework reduces energy usage in households in an intel-
ligent and user friendly manner. Actual consumption can be analyzed with respect to
consumption constraints set by the system. If either consumption constraints are (close
to being) violated or user needs are unnecessarily high, then energy consumption should
be decreased and corresponding control actions need to be exercised on the appliances
of interest.
Fig. 1 illustrates the proposed architecture of the home energy management system.
The MAPE (monitor, analyze, plan, and execute) control loop from autonomic com-
puting [6] is adapted to be used to analyzing energy consumption of electrical appli-
ances. Considering a pool of electrical appliances that are instrumented to allow mon-
itoring and control. The monitoring consists of measuring the energy consumption of
these appliances. The measurements are fed to a control process (the Appliances Man-
agement Process or AMP in Fig. 1), which interprets these as the actual consumption,
133
Focus of
this paper
Sensors
Controlled domain
(e.g. household)
Appliances
Rules
CP
CMP
AMP
MAPE
ECA
User Input
App mgr
Decision Unit
Ctx mgr
Action performer
App mgr
App mgr
Ctx mgr
Ctx mgr
Fig. 1. The basic architecture of the home energy management system. Adapted from [15].
and compares the consumption with the consumption constraints (in the Decision Unit
in Fig. 1). If, as a result of this analysis, it is decided that control actions are needed,
an action plan is produced. The action plan is derived with an algorithm that considers
time-shifting of the active state of appliances. Subsequently, the plan is executed (in
the Action Performer in Fig. 1) by performing the indicated control actions on the se-
lected appliances. Figure 1 illustrates the application of the MAPE control loop in the
top of the figure. The Decision Unit, which forms the focus of this paper (see Fig. 1) is
discussed in detail in Section 4.
With regard to analyzing consumer needs, the event-control-action (ECA) pattern
from context-aware computing [4] is used. Figure 1 illustrates the application of the
ECA pattern at the bottom of the figure. The environment of the appliances is con-
sidered. It is assumed that this environment is instrumented with sensors that are able
to measure relevant conditions. For example, measurements may be used to determine
context changes or situations, such as the presence of one or more persons in the house
or in a particular room, the activity mode (sitting, walking, sleeping) of a person, or a
person entering or leaving the house. Context situations and changes can generally not
be directly or reliably measured by a single sensor. A context management process (the
CMP in Fig. 1) is responsible for producing events that indicate the occurrence of a
context situation or change, based on reasoning which potentially involves sensor data
134
from several sources. Events are fed to a control process, which applies them in rules
to determine actions related to perceived needs. For example, if nobody is in the house,
a rule may establish the action to set the preferred value of the heating at 15 degrees
Celsius. Whether the actions are really required depends on the supported needs. For
example, if the preferred value of the heating is already set to 15 degrees Celsius, no ac-
tion is needed. The comparison of the perceived and supported needs leads to an action
plan, which, if not empty, is subsequently executed by performing the indicated control
actions on the selected appliances.
A service-orientated architecture (SOA) is used to implement the approaches out-
lined above, see [15] for more details.
3 Requirements for Knowledge Representation
The high level goal of the energy management system is twofold: On the one hand, on
the household level, the system should minimize the total energy consumption of the
household, within fixed boundaries set by the household owner. On the other hand, on
the neighborhood level, the system should minimize fluctuations in energy demand, i.e.,
keep the energy consumption of the whole neighborhood as constant as possible. Note
that these goals can be conflicting. Energy providers can prioritise one over the other
by providing (monetary) incentives. The energy management system’s high level goals
provide the first system’s requirement:
1. the system should be flexible enough to optimize either of the two high-level goals
of the system: minimize local energy consumption or minimize global fluctuations
in energy consumption
Moreover, the system needs to be highly adaptive, in particular:
2. the system should be able to adapt its behavior at runtime and change the high-level
goal, depending on input from the environment
Finally, users should be able to customize the systems to their specific needs, setting
limits to the adaptive behavior of the system:
3. users, i.e., home owners, should be able to customize the system to their specific
needs.
Note that the last requirement is the most important one. This can potentially limit the
adaptive behavior of the systems. However, it is crucial that users should be able to
override the energy management system, even if this means that, for example, the air
condition is set to maximum in each room. No users will allow a fully autonomous
system to manage their energy usage. This issue is discussed further in Section 5.
135
Rules
User Input
Other CPs
Action Performer
Decision Unit
APMCPM
Sensor infoRule Adapt
Rule Select Rule Engine
Fig. 2. The decision unit from Fig. 1.
4 Towards Adaptive and Self-Managing Knowledge
Representation
The main self-managing component of the energy management system is formed by the
Decision unit shown in Fig. 1. This unit has to (autonomously) decide how appliances
are adjusted to meet the goals of the system. Fig. 2 shows the decision unit in detail.
The decision unit as a whole takes input from four different sources, namely (i)
sensor input, (ii) user input, (iii) other control process units (CPs) at other households
and (iv) rules. The sensor input comes from the Context Management Process (CMP)
and the Appliances Management Process (AMP). These units process the sensor in-
formation and provide (aggregated) context and appliances status information which
forms the basis for the adaption process. The user input gives the home owner the op-
portunity to override the (autonomous) decision unit, for example, indicating that the
air conditioning unit in a room cannot be put off.
To meet the goal of the system, global (at a neighborhood level) minimization of the
fluctuations in energy consumption, the control processes (CPs) in individual house-
holds need to communicate with each other. The CPs can be organized in a virtual
tree overlay [13] and work together to meet this goal, for example using the approach
outlined in [12, 14].
The heart of the Decision Unit is formed by a rule based system which consist of
Rule Select, Rule Adapt and Rule Engine components, see Fig. 2. It is assumed, since
136
the application domain is known and unlikely to change rapidly, that the knowledge
base of the home energy system uses a fixed ontology, i.e., all rules are formulated
in the same (fixed) language. Rule based systems are traditionally very deterministic
systems. They consist of rules of the form shown in Example 1 below:
Example 1 (Rule base system)
matching condition 1 effect 1
matching condition 2 effect 2
. . . . . .
matching condition n effect n
4
By default the ordering of the rules defines the (operational) semantics of the system.
Rules are evaluated in order, and the first rule with a matching condition is executed, i.e.,
later rules that might match are discarded. Execution of a rule leads to an effect. In this
case something like altering the status of an appliance, i.e., lowering the temperature
of a fridge. Since the environment will change continuously different rules will be ex-
ecuted over time. However, given the same conditions the same rules will be executed,
making the system completely deterministic (static). To make the system more adap-
tive to its environment it should evolve over time, to meet the demands of a specific
household. There are several options that can be used to make the rule based system
adaptive, these include: (i) load specific set of rules based on the environment, for
example in the rule select unit a specific set of rules can be loaded that has as goal to
minimize global fluctuations instead of minimizing the local energy use of a household.
(ii) evolving rules, the rules can be changed based on genetic algorithms [3] or a neural
network [7]. (iii)weighted rules, in this approach all rules are weighted. Instead of ex-
ecuting the first matching rule all rules are selected and the rule with the highest weight
is executed. Fuzzy logic [8] like approaches can be used for this. (iv) hybrid approach,
combine some of the options above.
Evolving rules (item 2 above) effectively is typically difficult. Neural networks and
genetic algorithms try to merge and combine existing rules to produce new (better)
ones. However, a fitness function is required to determine if a newly generated rule
is better then existing ones. Finding a suitable fitness function is typically very hard.
Therefore this alternative is not further studied in detail here. Instead, a hybrid approach
that combines item 1 and 3 above is explored. The proposed adaptive rule based system
has the following properties:
rules are weighted
rules are bundled in a set, called the device set, per device
rules can be added and removed per set
sets can be added and removed to the rule base, the active rule sets that are used by
the rule engine
137
For each device there is an associated device set, consisting of weighted rules, that
determines the (adaptive) behavior of the device. Consider, for example, the following
three rules, shown in Example 2 below, that are part of the device set for controlling the
ac-unit in the master bedroom:
Example 2 (Weighted Rules controlling an AC unit)
. . . . . .
(0.3) #people in room 1 & t > t
max
& AC = off AC = on & cooling=10
(0.7) #people in room 1 & t > t
max
& AC = off AC = on & cooling=2
. . . . . .
(0.5) #people in room = 0 & AC = on AC = off
. . . . . .
4
All three rules have a weight (0.3, 0.7 and 0.5 respectively). The first two rules share the
same pre-conditioning, that evaluates to the value true if there is more then one person
in the room, the ac-unit is turned off and the current temperature (t) is higher then some
predetermined temperature (t
max
, for example 25 degrees Celsius). However, the result
of executing the rules is different: executing the first rule results in turning the AC unit at
setting 10, in the second case the AC unit is turned on at setting 2. A higher setting leads
to a faster cooling of the room, but also a higher (at least initial) energy consumption.
Since both these rules share the same precondition, the one with the highest weight (the
second rule with weight 0.7) is executed.
By adapting the weights of rules, different energy consumption patterns emerge.
Again looking at the two rules mentioned above, in the current situation energy con-
sumption is low over a longer period. If the weights of the two rules are swapped, the
result would be a higher energy consumption, but for a longer period. In effect, the first
mechanism that is used to adapt the system is this changing (adapting) of the weights
of the rules. This makes it is possible to adapt the reaction of the system to a specific
situation.
Note that the third rule (with weight 0.5) has a different pre-condition. It evaluates
to true if there is no one in the room and the AC unit is on. The effect of executing the
rule is that the AC unit is turned off.
The Rule Select unit (from Fig. 2) loads selected device sets into the rule engine. By
periodically loading different rules into the rule engine, i.e., by changing the rule base,
the system adapts to its environment. Based on input from the environment different
(possibly conflicting) high level goals can be met by different rule-bases. This loading
of different device sets provides the second adaptation mechanism of the system.
User input, i.e., from home owners, can be mapped easily to rules that deal with a
specific appliance, for example, the air conditioning unit in the bedroom. The rule is
seen in the Example 3:
138
Example 3 (User generated Rule controlling an AC unit)
(1.0) true AC = on
4
User generated rules should always evaluate to true (hence the pre-condition in
the rule above). Also note that such rules should always be loaded (unless specifically
cancelled by the user) and should have high weights (in this example, the maximum
value of 1.0) to ensure that user generated rules are executed. Finally, note that the
weight of such rules should, in principal, not be adapted by the system.
The system provides two adaptation mechanisms: weight adaptation which is han-
dled by the adapt rule component and selective device set loading (and unloading),
which is handled by the rule select unit. Separating these mechanisms has the advan-
tage that its easier to reason about adaptation policies at a higher (strategic) level. This
is left for future work, as are specific rule adaptation policies.
In summary, the Decision Unit takes input from the user and other CPs. Based on
input from the environment (sensor info), weighted rules are adapted and selected. The
rule engine selects all matching rules and chooses the ones that have the highest weight
per appliance. These are then send to the action performer which adapts the status of
the appliances. This whole process is repeated periodically.
5 Discussion and Conclusions
This paper discusses an approach and architecture for a home energy system based
on an adaptive and self-managing knowledge representation. The system is based on
a weighted rule based system that adapts continuously to its environment. One of the
main challenges of this system is to meet its different, possibly conflicting, goals. And
while the current architecture should make this possible it remains to be seen if these
goals are not too conflicting to be unifiable in practice. It might be necessary to drop
the goal of lowering the global fluctuations in energy consumption to meet the user’s
preferences and minimize the local energy consumption of the household. Simulations
and/or experiments should provide more insight on this issue. This is left for future
work.
Another issue is how to scale this up to collections of households. A hierarchical
structure could be used in which the architecture can be repeated at different levels of
granularity. For example, a household has appliances as units that are being controlled;
an apartment has living units as the units that are being controlled; a city block has apart-
ments as units being controlled; etc. Finding the correct clustering of households [11]
that are controlled by one processing unit forms another challenge.
From a technical perspective it is not very difficult, with the proposed architec-
ture, to force control processes in different households to cooperate to reduce peaks in
(global) energy usage. However, this might lead to some considerable discomfort with
home owners, for example if they can use their air conditioning unit at the maximum
139
setting, because a global reduction in energy consumption is required. Monetary incen-
tives, provided by energy producers who benefit from reduced peak usage, might help
lessen the discomfort of the home owner, as would specific policies set by local gov-
ernments. However, if this will be an acceptable solution remains to be seen. This issue
is further outside the scope of this paper.
A related issue is if an (semi) autonomous home energy management system will
be accepted by users. However, since there is both a monetary incentive (energy usage
is lowered which in turn leads to a lower energy bill) and since users can override the
behavior of the system this is probably less of an issue then the one discussed above.
Acknowledgements
The authors like to thank Boris Shishkov who was jointly (together with the first two
authors) responsible for the main idea behind the proposed energy management system.
The research presented in this paper has partially been funded by the Next Generation
Infrastructures project ‘Self-Managed Dynamic Institutions in Power Grids: Sharing the
Cost of Reliability’.
References
1. I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks: a
survey. Computer networks, 38(4):393–422, 2002.
2. P. Clark W. Gellings. The Smart Grid, enabling energy efficiency and demand response. The
Fairmont press, 2009.
3. A. Corcoran and S. Sen. Using real-valued genetic algorithms to evolve rule sets for classi-
fication. ICEC, 94(2405):120–124.
4. P. Dockhorn Costa, L. Ferreira Pires, and M. van Sinderen. Concepts and architectures for
mobile context-aware applications. In Handbook of research on mobile multimedia, Infor-
mation Science Reference. Hershey, New York, 2008.
5. F. Hayes-Roth. Rule-based systems. Communications of the ACM, 28(9):932, 1985.
6. IBM Corporation. An architectural blueprint for Autonomic Computing. White paper, 2005.
7. E. Keedwell, A. Narayanan, and D. Savic. Evolving rules from neural networks trained
on continuous data. In The 2000 Congress on Evolutionary Computation CEC 00, pages
639–645, 2000.
8. G. Klir and B. Yuan. Fuzzy sets and fuzzy logic: theory and applications. Prentice Hall
Upper Saddle River, NJ, 1995.
9. P. Mazza. The Smart Energy Network: Electrical Power for the 21st Century. Climate
Solutions, 2002.
10. A. Middelberg, J. Zhang, and X. Xia. An optimal control model for load shifting - With
application in the energy management of a colliery. Applied Energy, 86(7-8):1266 1273,
2009.
11. E. Ogston and F. M. T. Brazier. Apportionment of control in virtual power stations. In In
the proceedings of the international conference on infrastructure systems and services 2009:
Developing 21st Century Infrastructure Networks, 2009.
12. E. Pournaras, M. Warnier, and F. M. Brazier. Local agent-based self-stabilisation in global
resource utilisation. International Journal of Autonomic Computing (IJAC), 2010.
140
13. E. Pournaras, M. Warnier, and F. M. T. Brazier. Adaptive Agent-based Self-organization for
Robust Hierarchical Topologies. In ICAIS ’09: Proceedings of the International Conference
on Adaptive and Intelligent Systems. IEEE, September 2009. (to appear).
14. E. Pournaras, M. Warnier, and F. M. T. Brazier. A distributed agent-based approach to sta-
bilization of global resource utilization. In Complex, Intelligent and Software Intensive Sys-
tems, International Conference, pages 185–192, Los Alamitos, CA, USA, 2009. IEEE Com-
puter Society.
15. B. Shishkov, M. Warnier, and M. van Sinderen. On the Application of Autonomic and
Context-Aware Computing to Support Home Energy Management. In The proceedings of
the 12th International Conference on Enterprise Information Systems (ICEIS 2010), 2010.
16. M. Stadler, W. Krause, M. Sonnenschein, and U. Vogel. Modelling and evaluation of control
schemes for enhancing load shift of electricity demand for cooling devices. Environmental
Modelling & Software, 24(2):285 – 295, 2009.
17. G. Strbac. Demand side management: Benefits and challenges. Energy Policy, 36(12):4419
4426, 2008. Foresight Sustainable Energy Management and the Built Environment Project.
18. R. Thomas, D. Friend, L. DaSilva, and A. MacKenzie. Cognitive networks: adaptation and
learning to achieve end-to-end performance objectives. IEEE Communications Magazine,
44(12):51–57, 2006.
19. F. Zhao and L. Guibas. Wireless sensor networks. Communications Engineering Desk Ref-
erence, page 247, 2009.
141