Over the decades, energy issues have been getting more important.
In the U.S., 48% of energy consumption is from buildings, of
which 25% is associated with heating and cooling at an annual
cost of $40 billion. Furthermore, on an annual basis, buildings
in the United States consume 73% of its electricity.
With rising energy costs, the need to design and integrate scalable
energy consumption reduction strategies in buildings calls
for novel approaches. There are numerous challenges associated
with energy resources such as supply and depletion of energy resources
and heavy environmental impacts. The rise in energy
consumption in buildings can be attributed to several factors such
as enhancement of building services and comfort levels, through heating, cooling and lighting needs and increased time
Motivated by this, we focus on the study of sustainability in order to conserve energy. The goal is not only to construct
the correct model for capturing important phenomena in such domains, but to generate optimal strategies to effectively achieve given goals.
Specifically, this work includes the following projects:
TESLA (Transformative Energy-saving Schedule-Leveraging Agent)
One of Research Testbed Buildings: Leavey Library at USC
Energy-oriented Scheduling Agent
Reducing energy consumption is an important goal for sustainability. Thus, conserving
energy in commercial buildings is important as it is responsible for significant
This energy consumption is significantly affected by a
large number of meetings or events in those buildings. Furthermore, a recent study
shows that meeting frequency in commercial buildings is significant and continues
to grow. In 2001, U.S. Fortune 500 companies are estimated to have held 11
million formal meetings daily and 3 billion meetings annually.
Energy-oriented scheduling can assist in reducing such energy consumption.
Although conventional scheduling techniques compute the optimal schedule
for many meetings or events while satisfying their given requirements (i.e., computing
a valid schedule), they have not typically considered energy consumption
explicitly. More recently, there have been some trials to conserve energy
by consolidating meetings in fewer buildings.
To that end, we present TESLA (Transformative Energy-saving
Schedule-Leveraging Agent), an agent for optimizing energy use in commercial
buildings. TESLA is a goal-seeking (to save energy), continuously running autonomous
agent. TESLA's key insight is that adding flexibility to meeting schedules
can lead to significant energy savings. Users in a commercial building continuously
submit meeting requests to TESLA while indicating their flexible meeting preferences.
TESLA schedules these meetings in the most energy efficient manner while
ensuring user comfort; but in cases where shifting meeting times can lead to significant
savings, TESLA interacts with users to request such a shift.
This work particularly provides four key contributions. First, it provides
online scheduling algorithms, which are at the heart of TESLA, to solve a
two-stage stochastic mixed integer linear program (SMILP). This SMILP considers
the flexibility of people's preferences for energy-efficient scheduling of incrementally/
dynamically arriving meetings and events. Second, TESLA also includes an algorithm to effectively
identify key meetings that could lead to significant energy savings by adjusting their
flexibility. Third, this work provides an extensive analysis of the energy saving results
achieved by TESLA. Lastly, surveys of real users are provided indicating that
TESLA's savings can be realized in practice by effectively leading people to change
their schedule flexibility. To validate our work, we used a public domain simulation
testbed, which is described below, fitted it with details of our testbed building, and compared the simulation
results against real-world energy usage data. Our results show that, in a validated
simulation using our testbed building, TESLA is projected to save about 94,000 kWh
of energy (roughly $18K) annually. Thus, TESLA can potentially offer energy saving
benefits to all commercial buildings where meetings affect energy usage.
TESLA architecture: TESLA is a continuously running agent that supports four key features: (i)
energy-efficient scheduling; (ii) identifying key meetings; (iii) learning user preferences; and (iv) communicating
Research Testbed Buildings & Real Data
In this work, we consider two sets of research testbed buildings: i) Leavey library, which is one of main libraries at the University of Southern California and
ii) 7 buildings and 1 open space at the Singapore Management University.
In these buildings, meetings are requested by users by a centralized online room
reservation system. In the current reservation system, no underlying
intelligent system is used; instead, users reactively make a request based on the availability of room and time when they access the system.
In collaboration with building system managers, we have been collecting data specifying
the past usage of group study rooms, which are collected for 8 months (January through August in 2012) at USC
and for 3 months (August
through October) in 2011 at SMU. The data for each meeting request includes the time
of request, starting time, time duration, specified room, and group size. The data set
contains over 110,000 unique meetings, which are used for evaluating TESLA.
SAVES (Sustainable multiAgent system for optimizing Various objectives including Energy and Satisfaction)
Multiagent Systems to Conserve Energy
Recent developments in multiagent systems are opening up
the possibility of deploying multiagent teams to achieve complex goals in such energy domains that inherently have uncertain and
dynamic environments with limited resources. To model and optimize buildings' energy consumption, building
agents, facility managers and human occupants are demanding robust, intelligent and
adaptable ambient planning techniques. To realize both tangible benefits
such as energy and operation savings, value property, reduction
in occupant complaints as well as the intangible benefits such as
occupant comfort and satisfaction, designers must develop energy
adaptive capabilities within the building environmental control systems.
This project focuses on a novel application to be deployed
at Ralph & Goldy Lewis Hall (RGL) at the University of Southern
California as a practical research testbed to optimize multiple
competing objectives: i) amount of energy used in the buildings; ii)
occupants' comfort level; and iii) practical usage considerations. This work
provides three key contributions. First, we explicitly consider uncertainty while reasoning
about coordination in a distributed manner. In particular, we uses a novel algorithm for generating optimal MDP (Markov Decision Problem)
policies that explicitly consider multiple criteria optimization (energy
and personal comfort) as well as uncertainty over occupant
preferences when negotiating energy reduction
behaviors and their occupancy preferences are incorporated into
planning and modeled as part of the system. As as result, our system
is capable of generating an optimal plan not only for building usage
but also for occupants. Third, the influence of various control
strategies for multiagent teams is evaluated on an existing university
building as the practical research testbed with actual energy
consumption data in the validated simulation testbed. Since the simulation environment
is based on actual data, this result can be easily deployed into
the real-world. For future
work, we consider opportunities for direct occupant participation
and incentivization via handheld devices and deploy our system to
Research Testbed Building: RGL at USC
Validated Simulation Testbed
Human Subject Study on Energy Conservation
We also design and conduct a validation experiment on a group of human occupants in commercial buildings via a set of agents in our system: room and proxy agents.
There is a dedicated room agent per office
and conference room, in charge of reducing energy consumption
in that room. It can access sensors to retrieve room information
and energy use and impact the operation of actuators. A proxy
agent is on an individual occupant's hand-held device and it
has the corresponding occupant's models. Proxy agents communicate
on behalf of an occupant to the room agent based on their
adjustable autonomy - when to interrupt a user and when to act
autonomously. Room agents may directly communicate with occupants
without proxy agents, and different room agents coordinate
among themselves as well as with proxy agents.
We conduct this investigation: i) to verify if our system can lead to
changes in occupants' behaviors and to reduce energy consumption in commercial buildings,
ii) to validate the parameter values used during the negotiation process such as the acceptance/compliance rate for the suggestion
and iii) to understand what types of feedback are most effective to affect occupants' energy-related decisions.
THINC (Energy Saving Incentivization )
Abating energy consumption is an important goal to secure a sustainable future.
Specifically, commercial buildings are a key contributor to United States energy
use, representing roughly 25% of total energy expenditures according to the United
Large number of meetings or events significantly impacts such expenditures. U.S.
Fortune 500 companies annually hold more than 11 million meetings daily. Using
algorithms and optimization technique, known as THINC, energy minimization is
possible by scheduling meetings back to back; reducing reheating or re-cooling
In order to gain a greater global maximize savings we ask participants to specify a
single time slot and location for their meeting to be held in a campus library which
schedules over 300 meetings per day. When indicating a time slot you may offer
a wider window around your preferred time slot and location by being flexible;
offering different times and locations.
Assume that you are looking to schedule a meeting using a meeting reservation system,
where you can specify a single option in terms of time and the location for the meeting.
When scheduling this meeting you may offer a wider window around the previous
preferred slot by being flexible; offering different times and locations. This will allow
energy savings as an energy efficient algorithm generates a schedule; reducing re-cooling
costs of vacant rooms.
Shapley allocation divides the combined energy savings of the group amongst its
participants. Incentivizing participants through monetary compensation may yield
towards future offerings of flexibility. Shapley value is taken for each participant and is
an average over all possible scenarios or combinations of the groups flexibility related to
the scheduling effect of a single participant. This effect is the difference of going from a
state of no flexibility to their offered flexibility related to the current state of the system.
Thus the value is the individual's contributes to the whole as it relates to the contribution
of all other participant in the system.
Although the shapely value is algorithmically fair in dividing cooperative gains an
individual participant's perception of fairness need to be evaluated. It of utmost
importance from a policy and behavioral perspective that people believe that they have
been treated fairly as it is advantages for them to participate again. By finding ways to
positively promote energy saving behavior by offering flexibility better schedules and
greater achievement of energy savings are possible.
Jun-young Kwak, Pradeep Varakantham, Rajiv Maheswaran, Yu-Han Chang, Milind Tambe, Burcin Becerik-Gerber, and Wendy Wood,
"TESLA: An Energy-saving Agent that Leverages Schedule Flexibility," in Proceedings of the 12th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), May, 2013 (Innovative Applications Track)
Jun-young Kwak, Pradeep Varakantham, Rajiv Maheswaran, Milind Tambe, Farrokh Jazizadeh, Geoffrey Kavulya, Laura Klein, Burcin Becerik-Gerber, Timothy Hayes, and Wendy Wood,
"Sustainable Multiagent Application to Conserve Energy," in Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), June, 2012 (Demonstration Track)