Historically, electricity has been available whenever it was required.
This was achieved by varying the amount of generation to match the
demand at all times. This requires very expensive over-provisioning of
generation capacity. More importantly, as society makes the transition
to sustainable energy sources such as wind and solar, generation can
only occur when the resource is available.
Demand response (DR) refers to any system in which the supply-demand balance
is achied by deferring certain demands. Typically these demands are
from things such as charging batteries or pumping water through swimming
pools, which can be done at any time.
This project will investigate important open questions in demand
response systems, such as:
How do we provide suitable incentives to get people to take part
in DR systems?
What effect do the short-term incentives to be
part of a DR system have on long-term behaviour, such as the choice of
which appliances to purchase.
How much is total demand reduced rather than simply rescheduled by
DR? Can DR be seen as a queueing system for demand?
How does DR interact with the distribution network?
Energy storage management
As noted above, there is a growing need for technology to match
electricity consumption to non-controllable generation. Energy storage
is one such technology.
Storage technology such as batteries and pumped hydro can be used on
very short timescales in frequency regulation markets or to provide
spinning reserve, or on longer timescales to store solar energy for
night time, or store wind energy for cloudy days.
Optimal management of storage is made difficult by imprecise forecasts
of future load and future generation. For this reasons, techniques such
as Lyapunov optimization have been applied for storage management.
This poject will addess stoage management of ealistic stoage
devices, with impefectiosn such as inefficiency and self-dischage.
In paticular, it will look at management of a potfolio of devices with
different impefections, and investigate how much benefit comes from
having diversity among storage types.
The power grid used to be centrally managed infrastructure, with a small
number of generators whose states could be observed, but a sparsely
monitored transmission and distribution network. With the advent of
smart meters and distributed generation, we have many more sources of
information but also many more states that need to be observed.
This project will explore the use smart meter data for tasks such as:
Estimating the size and orientation of domestic solar panels,
based on daily variation in energy consumption.
Estimating the true electricity consumption of houses that have
In order to save electricity, it is important to know where the electricity
is going. Putting an electricity meter on every device in a house is
expensive, and so there has been considerable interest in estimating the
consumption of individual devices based on the measurement of a house's
total energy consumption.
Current techniques for this either require a large amount of manual
intervention to set up, or are unreliable -- or both. This project
Develop methods for "learning" device characteristics with minimal
Investigate new characteristics by which devices can be
identified, such as the spectrum of the noise in their steady-state
Energy management of IT systems
Dynamic capacity provisioning
Data centres are typically designed to process their peak workload, but
workload varies substantially throughout the day. As a result, many
people have proposed that some servers be turned off during periods of
low load. However, there is a cost for turning servers on and off, and
it is not known in advance how long the load will be low for.
This poject will investigate popeties of LCP, an on-line algoithm
that was ecently poposed to solve this poblem and othe "smoothed
online convex optimization" poblems. Specific questions are:
For what classes of cost functions is LCP better than
How can partial knowledge of future loads best be used by LCP,
what what performance guarantees can it provide?
Can LCP incorporate randomization to improve its average
Geographical load balancing
Many systems, such as Microsoft's Azure, spread work among many
geographically dispersed data centres. The ease with which data can be
transported makes this an opportunity to use data centres as
demand-response participants. Load can be shifted to data centres where
the availability of renewable energy temporarily exceeds the demand from
However, rerouting work often requires large databases to be migrated,
and so there can be a substantial cost to changing the allocation of
work. This makes it hard to determine the optimal routing without
knowing future conditions.
This project will explore the use of "metrical task system" algorithms
to distribute work between data centres without requiring any future
I am looking for students who
Want to improve the sustainability of global energy use
Have an excellent mathematical background, or a strong background
in both mathematics and programming