Advanced Metering Infrastructure: Challenges in Machine Learning and Wireless Networking Domains

cobblerbeggarΤεχνίτη Νοημοσύνη και Ρομποτική

15 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

52 εμφανίσεις

dvanced Metering Infrastructure
: Challenges

in Machine Learning and
Wireless Networking Domains

M Goyal,
J Bockhorst,
H Hosseini, D Yu, A Nasiri

EECS Department, University of Wisconsin Milwaukee, Milwaukee, WI, USA

These are challenging times. The world

depends on fossil fuels to run its cars and trucks, heat homes and generate
electricity. Exorbitant use of such fuels has brought us on verge of an environmental disaster. Fluctuations in the
price of oil routinely play
havoc with national economies.

nations, including United States, have recognized
their dependence on imported oil as a significant threat to their national security. The current economic crisis has
precipitated things and today, in the United States, we are witnessing a strong push tow
ards developing
hybrid/electric vehicles,

deploying wind and solar energy at a large scale and transitioning to a


The research challenges are numerous. The plug
in electric vehicles (PEVs) may worsen the peak demand levels
eyond what the exis
ting grid could support
. Integration of the wind and solar energy farms as local sources of
electricity requires design of community
micro grids
and calls for intelligent methods for automatically
forecasting the power output of th
ese intermittent power sources. The structure of the main grid and its control
protocols need to be redesigned to avoid cascaded failures and survive malicious attacks. A key component of the
new smart grid will be an elaborate

ructure that provides periodic (e.g. every 15
minutes) updates on energy consumption by individual households as well as aggregated demands at different
geographical scales. This, so called
Advanced Metering Infrastructure

(AMI), will allow real
time track
ing of energy
demand, thereby enabling frequent optimization of load on the grid’s distribution network and adjustment of
production schedules. The interaction between the real
time energy demand, production schedules and
constraints of the distribution ne
twork will allow real
time energy prices to feed back to individual households,
buildings and factories, enabling them to

their energy consumption to reduce energy costs. Thus, the
smart grid

will encourage consumers to save energy. Ener
gy cyber
physical systems will be integral to
bringing about these changes. However, there are significant shortcomings limiting current energy CPS design. This
position paper focuses on the advances needed in machine learning


wireless network
ing domai

Predicting Load and Generation Patterns

Continuous learning of load and generation patterns at multiple spatial and temporal scales is needed to support
manual and automated decision making at all levels. Load forecasting is presently used for unit c
reserve scheduling, and other activities. Load is typically forecast on fairly large geographical areas on hourly time
scales. Since it is mostly dispatchable, generation forecasting is not widely used. However, in order to support key
of energy CPS

for example, demand response, increased penetration of renewable power sources,
efficient charging of PEVs, distributed generation and storage

accurate predictions of load and generation at fine
granularities in space and time are necessa
ry. Load prediction will be more challenging in the future grid. Fine
grained predictions are more difficult as they lose the smoothing effect of coarse predictions, and load patterns
may be significantly changed by introduction of PEVs. Generation foreca
sting of wind and solar power will become
essential to the efficient use of these resources. In particular methods for incorporating the outputs of multiple
generating sites in the predictive models are needed.

As the characteristics of these renewable pow
er sources vary from site to site, prediction models will need to be
learned for each installation. However, since models at different sites, especially physically close sites such as PV
arrays in the same city, have much in common there is an opportunity

for sharing knowledge between sites.
Transfer learning is the branch of machine learning that deals with methods for sharing information between
related learning tasks. Transfer learning will be integral to an efficient energy CPS, not only for generation

forecasting, but for many other tasks such as learning energy use models of devices, rooms and buildings, learning
user energy consumption behaviors, identifying faulty and inefficient equipment etc. Over all, it seems that the
home/building needs to hav
e the smartness to adjust its energy consumption in face of high prices or grid
overloads based on its understanding of energy consumption profile of different devices and how they impact user
comfort levels. This would require
smart devices

with multiple
operation modes allowing wireless control of their
operation in addition to the traditional manual control.

Advanced Metering Infrastructure: Wired or Wireless?

The smart meters in homes/buildings could communicate with utility controllers via traditional
methods to
connect to Internet, e.g. DSL or Cable. Another option is to use
power line communication

(PLC) technologies
especially for electricity usage information. A third option is for the utilities to deploy
data collectors

in the
neighborhoods that a
ct as an interface between the smart meters and the main controllers in the utility premises.
The communication between the data collectors and the main controllers could be based on either wired (e.g. PLC)
or metropolita
n wireless (e.g. IEEE 802.16

) technology. However, the communication between the data
collectors and the smart meters is expected to be wireless, either using IEEE 802.11 (
) or IEEE 802.15.4

protocols. IEEE 802.11 technology allows very high data rates to be achieved, e.g. depl
oyed IEEE 802.11g networks
can easily support several 10s of Mbps and the proposed IEEE 802.11n networks can theoretically support
250Mbps or more. However, if the smart meters or the data collectors are constrained to be battery
IEEE 802.11 opera
tion would result in battery getting drained within few days. In such cases, use of IEEE 802.15.4
seems more appropriate. IEEE 802.15.4 networks can support a maximum data rate of 250 Kbps. However, IEEE
802.15.4 operation allows a device to have very smal
l duty cycles, i.e. a device could

for most of the time,
and thus survive on the batteries for several years. Significant technological problems need to be solved before
IEEE 802.15.4
based communication infrastructure between smart meters and data c
ollectors could be deployed
at a large scale. In the following paragraphs, we describe two of the essential issues



Routing in Low Power Wireless Networks

IEEE 802.15.4 specifies a
medium access control

(MAC) layer protocol based
carrier sense multiple access

(CSMA), where a node competes with other nodes in its radio range to transmit packets, and a

protocol that determines how an individual packet is encoded before transmission. Thus, IEEE 802.15.4 allows a

to communicate with another node within its radio range. Communication with a node not in the radio range
would require the nodes to participate in a routing protocol that allows a node to send packets to an out
destination via other nodes in the

network. Since the smart meters are unlikely to always have a data collector in
the radio range, such
hop routing

is a key requirement. Current routing protocols used in deployed sensor
networks have scalability issues. For example
, a popular protoc
ol, Zigbee
, does route discovery by doing a
wide broadcast of messages. Packet transmission is an expensive operation in IEEE 802.15.4 networks
because low data rates supported by IEEE 802.15.4 and also because packet transmission and reception con
energy, a precious resource in battery
powered nodes. A network
wide broadcast involves every node in the
network re
broadcasting a received message and hence is an overly expensive operation. Thus, frequent route
discovery, in face of dynamic networ
k conditions, is difficult with Zigbee protocol. Significant efforts are being
made at
Internet Engineering Task Force

(IETF) to develop scalable routing solutions for wireless
low power

. An acceptable routing protocol must have: 1) low
control packet overhead, i.e. the routing
protocol must not require too many control packets to be generated for its operation; 2) low computational
overhead since the individual nodes will have very limited computing capacity (e.g. 8
bit CPU); 3) low mem
overhead since the individual nodes will have very limited memory (only few hundred KB of RAM). Further, the
memory, computational and control traffic overhead of the routing protocol should scale well with the size of the
network. If a suitable routin
g solution is available, the number of data collectors required to communicate with
smart meters in a neighborhood could be significantly reduced.

The IETF is also engaged in developing solutions to integrate IEEE 802.15.4
based networks with Internet
. Th
technology would allow IPv6 addresses to be assigned to individual smart meters and appliances in the homes and
buildings. This technology can significantly expedite the large
scale deployment of IEEE 802.15.4
based AMI and
smart homes/buildings
ously monitoring and controlling the operation of
smart devices

over an IEEE
802.15.4 network to save energy. While allowing significant flexibility in terms of monitoring and control, having
individual meters and appliances accessible over Internet also
leaves the grid susceptible to malicious Internet
based attacks.

Reliability in Low Power Wireless Networks

Wireless communication is not as reliable as wired communication.

propagation effects and PHY noise
can result in a large fraction of pack
ets to be lost. Two additional problems are: 1) packet collisions due to

nodes and 2) poor spectrum
sharing between wireless protocols operating in the same radio range. Unlike IEEE
802.11, IEEE 802.15.4 does not have a built
in protection against t
ransmissions from

nodes, i.e. nodes
de the radio range of a node
. Consider the following scenario: node A sending packets to node B. Suppose, a
third node C is hidden from A but is within the radio range of B. Since nodes A and C are hidden fro
m each other,
IEEE 802.15.4 protocol cannot prevent concurrent transmissions from these two nodes. If node A is sending a
packet to node B, any concurrent transmission by node C would render A’s transmission unintelligible to B. IEEE
802.11 protocol solves

the hidden node problem by requiring the source and destination nodes to perform a
handshake by exchanging small packets before the main (large) packet transmission. The small packets intimate

nodes about the imminent transmission thereby p
reventing concurrent transmissions from hidden
nodes. IEEE 802.15.4 cannot benefit from a similar strategy since packet sizes are small, hence protecting a small
packet’s transmission by exchanging a handshake involving additional packet transmissions does

not make sense.
Other strategies, such as the use of multiple radio channels, need to be investigated to solve hidden node problem
in IEEE 802.15.4 networks. The second problem is that of poor coexistence between different wireless technologies
in the same frequency range. An overlap between the channels used by IEEE 802.11 and IEEE 802.15.4
networks can significantly
deteriorate their operation
. This is especially true for low power IEEE 802.15.4 networks.
Both IEEE 802.11 and IEEE 802.15.4 MAC
protocols are based on CSMA, although CSMA versions used by the two
protocols are quite different. There is a need to investigate changes in the protocol operation that would allow
these protocols to coexist with each other and with other protocols in the
same frequency range.