Transcript Why Smart Energy?

SMART ENERGY
&
MACHINE LEARNING
Benjamin Manning
[email protected]
706-344-2878
2
Who am I?
• BS, MS University of Southern Miss
• Machine Learning Thesis – 1999
• Software Engineering Company – 2001
• GT PhD Program – 2001
• CSU Systems Engineering Program – 2013
• UGA PhD Program – 2016
• Currently Teach Engineering Informatics
• Data Analytics PD Program at UTA and Rutgers
• Google Internet of Things (IoT) Technology Research Award
Recipient (2016)
• Research Interests:
• IoT, Smart Energy, Machine Learning, Virtual Reality
3
Smart Energy
• A smart energy system is a cost-
effective, sustainable and secure
energy system in which
renewable energy production,
infrastructures and consumption
are integrated and coordinated
through energy services, active
users and enabling technologies.
Getting Smart About Smart Energy Robert D. Cormia Foothill College
Smart Energy Defined
• Integrating key technologies
• Power grid / distribution
• Power generation (RE)
• Power systems & AMI
• Transportation systems
• Telecommunications (HAN)
• Information Technology (IT)
• A Smart Grid transforms the way power is delivered, consumed and
accounted for. Adding intelligence throughout the newly
networked grid increases reliability and power quality; improves
responsiveness; increases efficiency; handles current and future
demand; potentially reduces costs for the provider and consumer;
and provides the communication platform for new applications (The
Smart Grid in 2010 – Green Tech Media Research)
Smart Energy Paradigm
Generation
Demand
Distribution
Storage
The ‘Right Energy at the Right Time’ – ‘Right Sourcing’ Generation, Distribution, and End Use of Energy
Five Key ‘Quadrants’
• Electrical Generation
• Transmission and Distribution – T&D
• Energy Storage (battery, hydro, fuel cell)
• Electrical Load (demand / management)
• Management (sensors and analytics)
Generation Types
• Fossil – coal and natural gas
• Nuclear – base load
• Hydro – clean and affordable
• Wind – clean but intermittent
• Solar – peak shaving
• Geothermal – steam energy
Grid Overview (T&D)
• Generation
• Transmission
• Distribution
• High voltage / AC
• Substations
Grid Definitions
The electric grid delivers electricity from points of
generation to consumers, and the electricity delivery
network functions via two primary systems: the
transmission system and the distribution system. The
transmission system delivers electricity from power plants
to distribution substations, while the distribution
system delivers electricity from distribution substations to
consumers. The grid also encompasses myriads of local
area networks that use distributed energy resources to
serve local loads and/or to meet specific application
requirements for remote power, village or district power,
premium power, and critical loads protection.
http://www.oe.energy.gov/smartgrid.htm
Traditional Power Grid
Electrical Grid Networks
An electric grid is a network of synchronized power
providers and consumers that are connected by
transmission and distribution lines and operated by one
or more control centers. When most people talk about
the power "grid," they're referring to the transmission
system for electricity.
The continental United States does not have a national
grid. Instead, there are three grids: the Eastern
Interconnect, the Western Interconnect and the Texas
Interconnect. In Alaska and Hawaii, several smaller
systems interconnect parts of each state.
http://whatis.techtarget.com/definition/electric-grid.html
Demand (Load)
• Voltage drops / current
•
•
•
•
•
demand
Distribution networks
Appliances plug loads
EV charging
ESP – Energy Service
Providers
Storage (systems)
Power Management
• Management is everything!
• Monitoring loads over time
• Using predictive analytics
• Storing energy for timely distribution
• Holding load back (load shifting)
• Coordinating load uptake (renewables)
Why Smart Energy?
• System is ‘load driven’
• One way communication
• Load isn’t ‘managed’
• ‘Right sourcing’ energy
• Power grid needs to act like an
intelligent system
MY WORK
SOLAR RADIATION
PREDICTION
WHAT IS THE PROBLEM?
Problem Statement
• As the need for alternative energy resources begins to
increase the adoption of using solar energy has become a
popular alternative for many families
• This is causing a fluctuation in the estimated energy
needs forecasted by energy companies
• Can machine learning be used to build a predictive model
to help predict the amount of available solar radiation at
any given time?
Proposal and Goals
• A new ‘general’ model that will predict solar radiation at
any time; this is based on the reliability of a given
combination of predicted weather features and does not
rely on a given times series.
• The main goal was to find a combination of the lowest
number of features that could be used to build a predictive
model that would meet or exceed a 90% R2 success value
• By also reducing the number of weather related features
needed the risk created from fluctuating weather
variability was also reduced.
Data
• Georgia Automated Environmental
Monitoring Network (AEMN) - solar
radiation measurements taken over a
twelve-year period.
• Training data was created from one
year (2003) of data consisting of
observations taken every fifteen
minutes.
• Data from 2014 was selected for
testing because it was the furthest year
away from 2003 available at the time of
testing.
• This was done to reduce the likelihood
of presenting any biased testing data
to the model that may have been
present due to climatological or
seasonal similarities present in years
closer to 2003.
Pre-processing and Feature Selection
• A dataset containing forty-
three attributes from AEMN
was analyzed using R
• Low ranking attributes and
constants were removed
from the original dataset
• Data used for the study was
normalized to ensure that
any feature scaling did not
influence or disadvantage
any of the machine learning
algorithms used in the study.
Pre-processing and Feature Selection
• Three feature selection
methods used were:
• Filtering
• Embedded
• Wrappers
• Recursive Feature
Elimination and correlation
was used to achieve the
lowest number of needed
features needed by the
model
• Ten-fold cross validation was
used for all training and the
data was randomly stratified
before training began.
Models
• Ten-fold cross validation
was used for all training
and the data was
randomly stratified
before training began.
• After removing the
identified features seven
predictive models were
created for this
comparison study.
Wrapper - RFE
Bagged Tree
Random Forest
Embedded Model
Stepwise Linear Regression
Non-RFE Ensemble Models
Bagged Tree
Random Forest
Cubist Regression Tree
Gradient Boosting
Models with Feature Selection
1
0.9
0.8
0.7
0.6
RMSE
0.5
R2
0.4
0.3
0.2
0.1
0
Random Forest
Bagged Tree
Stepwise Linear Regression
Models without Feature Selection
1
0.9
0.8
0.7
0.6
0.5
RMSE
R2
0.4
0.3
0.2
0.1
0
Bagged Tree
Ensemble
Random Forest
Ensemble
Cubist Model Tree
Ensemble
Gradient Boosting
Ensemble
Generalization
Using a naïve estimate of 75% being the current prediction accuracy electric
companies can ascertain:
For example: understanding that 30% of the potential electrical energy
consumed in one residential location in Georgia can come from solar energy and
the average home in this location uses 450-kilowatt hours monthly, one can
generalize that an electric company could possibly overproduce up to 112.5kilowatt hours monthly just for this one home or 1350-kilowatt-hours annually if
their estimates have a 75% probability of always being correct.
Given this newly proposed generalized predictive model with an estimated
variance understanding of 91.0% one can ascertain this overproduction could be
reduced from 112.5 kilowatt-hours monthly down to ~10.0 kilowatt-hours monthly
or from 1350 kilowatt-hours annually to ~120 kilowatt-hours annually for a single
household.
Even with more cautious estimates based on the new generalized model the
savings could be sizeable when scaled across a region.
Generalization
• It is important to note that future analysis is needed to support the use of
automated feature selection methods when used in this manner
• In addition this general predictive model does contain an error rate, as do
most regression type problems of this sort. Since the data was normalized the
RSME values are reflective of this method and are larger in reality.
• It is important to note that the predicted values were less that the actual
values in almost all cases; this is important because it provides the energy
companies with floor estimations when using the model.
Future Work
• Supporting external features will be investigated as possibly being added to
the data to improve the accuracy of the model(s).
• Usage of genetic algorithms for feature selection
• The conversion of this regression type problem to a classification type
problem will be part of the next study
• Learning Vector Quantization (LVQ) model will also be used to rank feature
importance
Questions?