VERGE 23: A Practical Guide to Harnessing AI for Decarbonization

Tutorial
A Practical Guide to Harnessing
AI for Decarbonization
#VERGE23

Climate Change and AI: Opportunities,
Challenges and Considerations
Utkarsha Agwan
Climate Change AI
Verge 2023
Session: A Practical Guide to Harnessing AI for Decarbonization
Based on the ICML 2022 tutorial “Climate Change and ML: Opportunities, Challenges, and Considerations” by Priya Donti, David Rolnick, and Lynn
Kaack

*OECD AI Principles.
What is artificial intelligence?
Artificial intelligence (AI): Any computer algorithm that makes predictions,
recommendations, or decisions on the basis of a defined set of objectives.*
Machine learning (ML): AI algorithms that infer patterns from data.
Especially popular / eﬀective recently, thanks to deep learning / neural networks.
Weaknesses
● Sensitive to bad or biased data
● Donʼt have ʻcommon senseʼ
● Often cannot explain why an
answer is true
Strengths
● Performing simple tasks quickly and
automatically
● Finding subtle patterns in large datasets
● Optimizing complex systems
2

Climate change warrants rapid action
Impacts felt globally
Disproportionate impacts on most
disadvantaged populations
Filippo Monteforte | AFP | Getty Images David Mcnew | Getty Images
NASA Piyaset | Shutterstock.com
Need net-zero greenhouse gas
emissions by 2050 (IPCC 2018)
▸ Across energy, transport, buildings,
industry, agriculture, forestry, etc.
How does ML fit into this picture?
3

Electricity systems Buildings Transportation
Climate prediction Industry Societal adaptation

Land use

Distilling raw data into actionable information
Optimizing complex systems
Improving predictions
Accelerating scientific discovery
Approximating time-intensive simulations
Roles for ML in mitigation, adaptation, & climate science
See also: https://www.climatechange.ai/summaries

1. Distilling raw data
Role: Distilling raw data into actionable information
Some relevant ML areas: Computer vision, natural language processing
7
▸ Gathering data on building footprints/heights [M]
▸ Evaluating coastal flood risk [A]
▸ Parsing corporate disclosures for climate-relevant info [A]
Examples (M: Mitigation, A: Adaptation)
▸ Mapping deforestation and carbon stock [M]

2. Optimizing complex systems
Role: Improving efficient operation of complex, automated systems
Some relevant ML areas: Optimization, control,
reinforcement learning
8
▸ Optimizing rail and multimodal transport [M]
▸ Demand response in electrical grids [M]
Note: Beware of misaligned objectives and rebound effects
Examples
▸ Controlling heating/cooling systems efficiently [M]

3. Improving predictions
Role: Forecasts and time series predictions
Some relevant ML areas: Time series analysis,
computer vision, Bayesian methods
9
▸ Forecasting electricity demand [M]
▸ Predicting crop yield from remote sensing data [A]
Examples
▸ “Nowcasting” for solar/wind power [M]

4. Accelerating scientific discovery
Role: Suggesting experiments in order to speed up the design process
Some relevant ML areas: Generative models,
active learning, reinforcement learning,
graph neural networks
10
▸ Algorithms for controlling fusion reactors [M]
Examples
▸ Identifying candidate materials for batteries, photovoltaics,
and energy-related catalysts [M]

5. Approximating simulations
Role: Accelerating time-intensive, often physics-based, simulations
Some relevant ML areas:
Physics-informed ML, computer vision,
interpretable ML, causal ML
11
▸ Simulating portions of car aerodynamics [M]
▸ Speeding up planning models for electrical grids [M]
Examples
▸ Superresolution of predictions from climate models [A]

Roles for ML in mitigation, adaptation, & climate science
Distilling raw data into actionable information
Optimizing complex systems
Improving predictions
Accelerating scientific discovery
Approximating time-intensive simulations
See also: https://www.climatechange.ai/summaries

Questions that we asked in identifying priorities
▸ Is ML needed to address the problem?
▸ What is the scope of the impact? (in rough terms)
▸ What is the time horizon of the impact?
▸ What is the likelihood that a solution can be found?
▸ Can a solution feasibly be deployed?
▸ What are the potential side eﬀects of deploying the candidate solution?
▸ Who are the relevant stakeholders who are involved in or aﬀected by
the application?
13

Key considerations
ML is not a silver bullet and is only relevant sometimes
High-impact applications are not always flashy
Sophisticated algorithms can be required, but aren't always
Interdisciplinary collaboration
▸ Scoping the right problems
▸ Incorporating relevant domain information
▸ Shaping pathways to impact
Equity considerations
▸ Empowering diverse stakeholders
▸ Selecting and prioritizing problems
▸ Ensuring data is representative
14

ML applications
in climate change
mitigation
ML applications
that increase
emissions
MLʼs carbon footprint
15
MLʼs system-level
impacts
Emissions from
ML computation
& hardware
Kaack, L. H., Donti, P. L., Strubell, E., Kamiya, G., Creutzig, F., & Rolnick, D. (2022). Aligning artificial intelligence with climate change mitigation. Nature Climate Change,
1-10. 15

ML applications
in climate change
mitigation
ML applications
that increase
emissions
MLʼs carbon footprint
16
MLʼs system-level
impacts
Emissions from
ML computation
& hardware
1-10. 16

Immediate application impacts
1-10. 17

System-level impacts of ML
Rebound eﬀects
Reducing energy consumption reduces costs → money saved
may be used and cause more emissions
Example: ML for optimizing systems
Lock-in and path
dependency
Technologies compete and dominate → lock-in to suboptimal
technologies hampering decarbonization
Example: Autonomous driving and car use
Consumer behavior
Trends and advertising may change consumption patterns →
embodied emissions in those products
Example: ML in advertising and social media
Communication and
education
Societal support for climate action essential
Example: ML on social media
18

Reports with opportunities for
researchers, practitioners, and
policymakers
New community-driven Wiki w/
datasets & additional resources
Digital resources
Climate Change AI
Catalyzing impactful work at the intersection of climate change & ML
Webinars & happy hours
Newsletter, blog, & community
Calls for Submissions
Funding
Projects & Courses
Readings
Jobs
Learn more & join in:
www.climatechange.ai
@ClimateChangeAI
Webinar series (monthly)
Virtual happy hours (biweekly)
Global research funding
for impactful projects
Funding programs
19
Workshop series
▸ Next workshop @ NeurIPS ʼ23
▸ Browse past accepted papers:
www.climatechange.ai/papers
Summer school (multiple tracks)
Conferences & events

Harnessing AI for Decarbonization
Trends, Applications and Opportunities
David Groarke | VERGE | October 2023

The
Role
of
Technology
in
Energy
Time
1970s-
1990s
2000 -
2020
Current and rapid
inflection point
2020 -
2040
We are entering a Third Epoch in the Power Sector…
1970s – 1990s
Restructuring
Characterized by deregulation of generation market,
increased transmission infrastructure buildout, and a drop in
R&D spending.
2000s - 2020
Digitizing
Deployment of smart grid technology, high volume of asset
and grid data capture, pilots and tests of new business
models.
Automating
2020 - On
An acceleration of industrial AI, novel applications, new
markets and benefit areas, and monetization of energy data.
© Indigo Advisory Group 2023 2

…where multiple trends are converging…
Electrification* 25% of energy for electricity 30% of energy for electricity 50% of energy for electricity Growing utility reliance / importance
1970s – 1990s 2000s - 2020 2020 - On
INSIGHT
Renewable
Economics*
Regulation
Decarb.
Narrative
Energy Tech
R&D Spend**
Industrial
Age
30% drop in installed cost of solar 70% drop in installed cost of solar 13% utility scale solar price decline Renewable cost declines
Utility restructuring, IPPs emerge Renewables regulation, tax credits New business models, markets Evolving regulatory structures
Renewable Energy Energy Transition Net Zero Increasing importance
EMS digitizes, microprocessors Digitization & Smart Grid New data, control applications Waves of digitization
74% decline 1993 - 2000 Vendor spend increases 40% Growth in energy tech VC Supply side innovation
ML Focus, expert system design Robotics, computer vision, NLP Significant progress, novel ideas AI for societal challenges
Industry 3.0 (Automation,
Computers, Microprocessors)
Industry 4.0 (IoT Cyber Systems,
Networks)
Industry 5.0 (AI Management, Self
Optimization)
Macro Industrialization trends
impact utilities
AI Trends
Power Trends
Tech Trends
Restructuring Digitization Automating
*IEA **IEEE
3

• The energy landscape is
becoming increasingly
complex. While the depth and
breadth of changes depend on
the region and jurisdiction,
there is a convergence of
industry transformation afoot
• Companies are having to
change how they manage new
assets and attract new
skillsets
• AI can help to alleviate and
accelerate some of these Third
Epoch Power Trends
…various dimensions define this new era…
4

© Indigo Advisory Group 5
…with a complex data journey unfolding…
System Data
Traditional EE Analytics Predictive AI & Automation
Existing Static Data Emerging Dynamic Data
Reliability
Data
New Epoch
Increasing Data Value
New
Forms
of
Data
(TB),
Decisions,
AI
Maturity
Net Zero Journey (Time)
Asset & Load Data Markets & Regulatory Data
Capacity Data
HVAC
Data
Lighting
Data
Asset Information
Data
Boiler
Load Profiles
Comfort Baseline
Data
O&M Data
Asset Manual Data Hosting Capacity
Data
Transition Plan Data
Fleet Data
Solar + Storage
Data
DER Data
Grid Interactive
Asset Data
Heat Pumps
Data
Fleet Data
Grid Value Data
Resiliency Data
Geothermal Data
Microgrid Data
EV Station Data
VPP Data
BMS
Meter Data Control Sys. Data
Data Repositories
DERMS Data
Visualization & GIS
Data
Building Automation
REC Data
EO22/EO88 Data
Compliance Reporting
Data
FERC 2222
Data
New Regulation Data
Rate Redesign
Data
Tax Incentive Data (IRA)
Repository of
assets
Modelling
(ASHRAE, etc.)
New Tech, Alerts
& Triggers
ILLUSTRATIVE TYPES OF DATA
New Paradigms
Data Produced
Advisory
Journey

…where AI will lead with intelligent decision making…
Incremental
Strategic
Innovative
Static Dynamic Predictive
AI & Data Dynamics
Opportunity
Complexity
Traditional
Energy Mgmt.
New Technology
& Markets
Compliance
Regulatory
Programs
EVs, Solar +
Storage
DER Deployments
Microgrids
Fuel Cells
Energy
Efficiency
Loads
HVAC
Syst. Controls
Grid Interactive Assets
New Markets (FERC 2222)
Leading Edge Technologies
Price Signals
Carbon Registry
Carbon Signals
Monitoring & Diagnostics
Geothermal
BESS
Capture Repository of
Energy Consuming Assets
Modelling for Price,
Sequencing
Data Alerts and the
‘Next Best Thing’
Asset Inventory
Plug & Process
Disruptive Technology
Industry Benchmarking
Data & Relationship Maturity
Using data to capture
appropriate decision-
making elements
while identifying
required solutions
AI in Commercial Energy involves the use of sophisticated computer systems and algorithms to perform tasks typically
requiring human intelligence. These systems often mimic or even surpass human capabilities in learning, predicting, analyzing,
and automating crucial functionalities for efficiency and reliability.

Deployed Emerging
Machine Learning Computer Vision Natural L. Processing Robotics Predictive Analytics
The use of algorithms to learn
patterns in data in order to make
insightful predictions or
decisions.
Data Types: Historical data,
real-time data, sensor data,
weather data
A type of AI that allows
machines to interpret and
understand the visual world.
Data Types: Image and video
data, sensor data
A branch of AI that enables
machines to understand,
interpret, and respond to human
language.
Data Types: Text data,
customer interaction data,
maintenance reports
The use of robots and
machines to perform tasks that
are typically done by humans.
Data Types: Sensor data,
image and video data
The use of statistical models
and machine learning
algorithms to analyze data and
make predictions about future
events.
real-time data, weather data
Distributed AI
AI systems that are distributed
across multiple devices and
locations, enabling decentralized
decision making.
Data Types: Real-time data,
sensor data, historical data
Explainable AI
AI systems that can explain
how they arrived at a decision,
making their decision-making
process interpretable.
Data Types: Text data,
customer interaction data,
transactional data
Reinforcement Learning
A type of machine learning that
involves training agents to make
decisions in an environment
based on trial and error.
Data Types: Real-time data,
simulation data
Knowledge Reasoning
The ability of machines to
reason and draw conclusions
based on knowledge and rules.
real-time data, sensor data
Digital Twins
A virtual replica of a physical
system that can be used to
simulate, monitor, and optimize
its performance.
Data Types: Sensor data,
historical data, simulation data
AI capabilities can leverage these vast C&I datasets…

OPERATIONS
Data Center Specific
• Power & Service Outage
Mgmt.
• Cooling Motor
Amperage Mgmt.
• Thermal Balance of
Data Processing Units
• Redundancy Mgmt.
Demand Management
• VPP/DERMS
• Demand Response &
Load Forecasting
• Dynamic Tariffs
• Remote Grid Edge
Control
Energy Efficiency
• Real-time Monitoring &
Optimization
• Energy Consumption
Monitoring
Asset Management
• Predictive Maintenance
• Digital Twins
• Energy Meter
Connectivity
• Outage Management
• RT Electrical Dist.
System Monitoring
Fleet Electrification Mgmt.
• Fleet Electrification
• Electric Vehicle
Charging Infrastructure
Management
• EV Active Managed
Charging
• Vehicle-to-Grid (V2G)
DER Mgmt.
• DER/Renewable
Investment Planning
• Site Selection
• Pre-Construction
Design
• Connection Request
Mgmt.
Energy Proc.
• Energy Cost
Forecasting
• PPA Mgmt.
• RFP Mgmt. &
Response
Automation
Carbon Accounting
• Emissions
Accounting
• Report Generation
• Carbon Simulations
• Decarbonization
Recommendations
Reg. Compliance
• Emergency Power
Compliance Testing
• Regulatory Change
Monitoring & Impact
Analysis
• Compliance Audits
• Tax Analysis
Platforms
• Building Mgmt.
Systems (BMS)
• VPP/DERMS
• Asset
Performance
Mgmt. (APM)
Platform
PLANNING
Operation Specific
Enterprise Wide
Operation Specific
Enterprise Wide
Customer Exp.
• Personalized
Customer
Journey
• Utility Bill Mgmt.
• Energy Usage
Dashboard and
KPI Mgmt.
…with multiple use cases emerging at various maturities…

Fleet Electrification
2. Top Vendors
Example 1 - AI for Fleet Electrification
• Fleet Electrification – Analyzing historical vehicle data, usage patterns & TOC to determine the best replacement
with EVs.
• Vehicle-to-Grid (V2G) - Software to discharge EV fleets batteries back into the grid to unlock storage revenue
streams.
• Charging Infrastructure Management – Optimizing charging station investment planning (location, capacity
planning) based on utilization rates and leveraging real-time data to predict maintenance for EV stations.
• EV Active Managed Charging – TOU performance for EV charging.
1. Use Cases
3. AI & Data Types
• Machine Learning
• Historical load data
• Hosting capacity
• Total Cost of Ownership (TOC)

Asset Management
2. Top Vendors
Example 2 - AI for Asset Managemnt
• Predictive Maintenance – Learning normal system operating parameters in order to predict next system failure such
as boilers to heat pump, HVAC, etc.
• Digital Twins - Building digital representation of physical assets using AI modeling techniques to simulate various
configuration scenarios.
• Real-time Monitoring – Remotely monitoring health of electrical equipment and low voltage distribution assets such
as switchgear.
• Outage Management – Monitor asset and equipment performance interruptions and outages across facility.
1. Use Cases
3. AI & Data Types
• Digital Twins
• System sensor data
• Asset health data
• Predictive Analytics

Demand Management
2. Top Vendors
Example 3 - AI for AI Demand Management
• VPP/DERMS - Aggregating Distributed Energy Resources (DERs) to coordinate their capabilities in unison, bidding
them into wholesale markets.
• Demand Response and Load Forecasting - Analyzing weather patterns, facility load data, and historical price data
to determine demand curves.
• Dynamic Tariffs – Assessing dynamic time-of-use tariffs and identifying optimal times to adjust load to maximize
energy savings.
• Remote Grid Edge Control - Connect to remote sites and buildings for remote monitoring, energy alerts, HVAC
setbacks, etc.
1. Use Cases
3. AI & Data Types
• Predictive Analytics
• Historical load data
• Hosting capacity
• Utility tariffs

DER Planning
2. Top Vendors
Example 4 - AI for DER Planning
• DER/Renewable Investment Planning - AI algorithms and energy consumption data determine the optimal level and
type of DER investment.
• Site Selection - Analyzing large amounts of geographical and environmental data to identify optimal locations for
renewable resources.
• Connection Request Mgmt. – Assessing how proposed DER interconnections will affect power flow and
interconnection on the grid
• Pre-Construction Design – Building 4D visualizations using AI-driven designs to aid contractors in construction
plans, scheduling, and testing.
1. Use Cases
3. AI & Data Types
• Hosting Capacity
• Model Forecasting
• Granular load data
• LBMP & market
nodes

…other markets are growing, as innovation costs decline…
Demand Management
Platforms
Data Center Specific
Energy Proc.
Carbon Accounting Customer Exp.
Asset Management
DER Planning
Reg. Compliance
Energy Efficiency
Fleet Electrification Mgmt
AI Vendors for
Commercial
Energy Use
Cases
We are seeing significant high
value AI-based solutions and
vendors emerge across 11
distinct use case domains
that enable commercial
customers to better reduce
and manage their energy.

…and benefits are being realized throughout the sector…
GHG Reductions Data Center Savings HVAC Optimization V2G Benefits Renewable Bidding
The DOE & Schneider cite
that digitization of office
HVAC controls can yield
emission reductions
of 42% with 3-year
paybacks.
Google suggests that they
have already saved 40%
on power consumed for
data center cooling
purposes by implementing
AI solutions.
Startup BrainBox AI’s
HVAC optimization solution
forecasts room conditions
and can reduce energy
usage by approximately
15%-20%
Vehicle-to-grid startup
Nuvve helps 10 school
district customers receive
electric bus rebates of
$24.2 million with
bidirectional charging.
Fluence's Mosaic’s AI-
powered bidding
optimization makes energy
assets more valuable with
up to 10% increased
renewable revenue.
Customer Services
AI can improve the efficiency
of customer service
operations by 30-50% by
automating routine inquiries
using chatbots and other AI-
based interaction tools.
Microgrid Planning
XENDEE partnered with a
college campus to reduce
reliance on fossil fuel
energy through designing a
microgrid to support resiliency
and reduce GHG emissions.
EE Automation
Telco company KDDI
leveraged Nokia’s AVA
energy efficiency optimization
solution to reduce power
consumption by 50% in
low-traffic environments.
Virtual Power Plants
Stem implemented their
AthenaAI VPP solution to
improve battery storage
portfolio energy savings by
more than 30%
monthly for customers.
Peak Demand Shifting
Community Energy Lab’s AI
solution shifted about 16% of
a school’s HVAC
cooling load away from an
on-peak price period, yielding
payback period of 2 months.

As AI technologies mature and decarbonization progresses, new applications will emerge across the power
sector at various speeds, enabling ‘prices to devices’ and ‘set it and forget it’ futures.
..however, a potential AI Automated System is still distant
DER
Level
Carbon Market with
Distributed Registry
Community
P2P Markets
Transactive
Grid
Settlement
P2P Settlement at the
Transmission Level
Customer Portals
Bundled Services
Dynamic Tariffs
Direct appliance level
control
Real-time Switching
Dynamic event
notification
Dynamic electricity
consumption forecasting
Market-based demand
response
Algorithms and
analytics for market
information/ops
M&V for producers and
consumers
DER power control
Real-time load
monitoring
Power flow control
Real time/predicted probabilistic
based area substation, feeder, and
customer level reliability metrics
Automated islanding and
reconnection
Time
Grid
Technology
Evolution
16

…finally, some Commercial AI trends to watch in 2024
17
Synthetic data may level the playing field: Synthetic data creates an opportunity for companies that don't have access to large
datasets. Synthetic data generators may also enable companies to build robust AI models without need for extensive real-world data.
AI is driving down the cost of innovation: The cost of building custom applications may be falling, thanks to code-generation tools
like Copilot and the availability of robust open-source LLMs. This is enabling vendors to explore and implement AI solutions at an
The initial focus is on tailored “Narrow AI”: Near-term will focus on Narrow AI solutions specifically tailored for companies. Most
energy companies are piloting and rolling out on non-critical (CIP) components of their business.
Customizing the customer experience: AI vendors are leveraging data to better understand customer behavior and suggest “Next-
Best Actions” in their energy management journeys.
Prioritizing trust and verification: Before fully trusting AI, companies are looking for ways to double-check and verify AI solutions,
especially those from outside vendors.
Taking the “Sandbox” Approach: Many companies prefer testing out AI in controlled 'sandbox' environments to learn and adapt
before wider implementation. While some vendors provide viable AI solutions, others might not fully understand the complexity and
unique challenges of the utility sector, requiring them to trial their solutions in contained environments.
Specialized models may become a lucrative area: There is appetite from investors for specialized LLMs, specifically focusing on
finding the "right model for the right job,“ and aiming for those that offer higher accuracy, lower costs, and optimal performance.

insights@indigoadvisorygroup.com
+1 212 203 6144
www.indigoadvisorygroup.com
@indigoadvisory
linkedin.com/company/indigo-advisory-group
450 Lexington Ave, 4th FL, 10017

AI for Decarbonizing
Building Operations
Fundamentals, Preconditions, Risks
Andrew Knueppel, PE, MBA
Workplace Engineering Manager
Cushman & Wakeﬁeld @ LinkedIn

The Opportunity
● Commercial buildings generate 16% of all US CO2 emissions 1
● On average, 30% of the energy used in commercial buildings
is wasted 1
● More opportunities for carbon reduction than ever before:
○ Indoor sensor data
○ Flexibly used spaces
○ Grid-interactivity

Decarbonization Fundamentals
How We Get There
3
Maintain
5-20% energy savings 3
2
Fix
16% energy savings 2
● Retro-commission to
restore design performance
● Data infrastructure:
devices, protocols,
networks, labeling
1
Prioritize
decarbonization
4
Improve
15%+ energy savings 4
● Transition from
plan-preventative to
data-driven proactive
● Bring in an analyst to maintain
data quality & manage
changes
● Upgrade to high-performance
sequences & integrate
systems
● Implement Fault Detection &
Diagnostics to identify &
prioritize issues
● Energy/carbon manager to
optimize & identify
opportunities
Outcomes
Deep energy/carbon savings, labor time savings, resilience & ﬂexibility to the future

Fundamentals as Preconditions to AI
Buildings are not Software
3
Maintain
2
Fix
✖ Stuck valves and dampers,
uncalibrated sensors
✖ Congested networks,
isolated systems,
proprietary protocols
Which of
these can AI
ﬁx?
4
Improve
✖ Reactive maintenance
practices and service
contracts
✖ Shifting portfolios, data
gaps and new datasets
❔ What about controls
optimization?
❔ What about identifying
issues & opportunities?

Overall
● Dependency on “Fix & Maintain” to achieve and maintain any savings
Short-Term Risk: Black Box
● From standards-based and globally familiar to statistical and opaque
● Opaque diagnostics or false positives → risk of being ignored
● AI building control + occupant complaints → risk of being overridden/shut off
Long-Term Risk: Dependency
● Increasing reliance on AI provider to identify opportunities, decreasing sense of
ownership locally
● Machine ‘learnings’ that created savings are lost if service is cancelled
Outcome Risks
Shallow or degrading energy/carbon savings, overwhelm/distrust of automations, limited
vendor ﬂexibility/lock-in
Considerations & Risks for AI

What Success is Built On (AI or Not)
● Making decarbonization a priority
● Properly functioning equipment & data infrastructure
● Data-driven O&M with transparent logic & diagnostics
● Distributed data, engineering & operational expertise

References
1. energy.gov: About the Commercial Buildings Integration Program
2. LBNL: Building Commissioning
3. DOE: Operations & Maintenance Best Practices
4. LBNL: Advanced control sequences and FDD technology

AI: The State of the Art, and the Art
of the Possible
Rachana Vidhi
NextEra Energy Resources
October 24, 2023

2
U.S. electric grid contributes ~25% to the overall emissions.
Currently!
Source: Environmental Protection Agency
*Residential and Commercial are combined

3
Role of data and AI becomes more critical as we go further on the
Decarbonization Journey
Visualize Realize
Maximize
-Volume and resolution of data
-Inter-dependency of variables
-Product complexity
What does this mean
for Decarbonization?

4
Data inter-dependency and product complexity increase dramatically
Visualize Realize Maximize
Site
Load
On-site
generation
Tariff
Emissions
Building
data
Equipment
data
Location
Cost of
energy
Contract
terms
Inter-
connection
Purchased
energy
Site
Load
On-site
generation
Tariff
Reduce
Emissions
Building
data
Equipment
data
Location
Cost of
energy
Inter-
connection
Purchased
energy
Loss
Under-
perform
ance
Contract
terms
Site
Load
On-site
generation
Tariff
Emissions
Arbitrage
Building
data
Equipment
data
Location
Cost of
energy
Inter-
connection
Purchased
energy
Loss
Hedged
energy
Market
prices

5
AI and GIS based patented process is used to design the most
economically optimal solar farms given site constraints

6
Continuous Improvement
AI based trading and battery management software is used to
maximize storage revenues
Use probability distributions to
generate a likelihood of
various forecasting events
Probabilistic
Forecasting
Risk-Based Offer
Generation
Real Time Updates Performance Review and
Model Re-training
Mathematical programming
techniques generate offer
parameters based on
customer risk tolerance
Site telemetry and updated
forecasting used to provide
real-time updates to offer
parameters where available
Forecasts and models are
continuously re-trained to
improve accuracy; AI used to
monitor site equipment and
identify underperformance
Automated Offer Generation

7
Sophisticated market participation strategy can maximize revenue
from battery storage assets that are critical for grid decarbonization
0
50
100
150
200
250
300
350
400
450
500
7/18/2022 7/19/2022 7/20/2022 7/21/2022 7/22/2022 7/23/2022 7/24/2022 7/25/2022
Energy
Price
($/MWh)
Energy price at a representative node
Real time Day ahead
Split
participation in
DA and RT
Prioritize DA
participation
Prioritize RT
discharge
-50
0
50
100
150
200
250
300
5/5 5/6 5/7 5/8
RT DA
RT revenue from charging
with negative prices
RT revenue from
discharging during
peak prices

8
Reliable and cost-effective Carbon Free Energy is needed for
supporting the growth of AI
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Dec
0
200
400
0
2
4
6
8
10
12
14
16
18
20
22
• Data centers and data transmission networks each
account for ~1.5% of global electricity use(1)
• Increasing use of AI and ML is expected to increase
this even further
• Providing Carbon Free Energy to support the
corporate goals will require advanced AI
Wind Solar Storage

9
As the energy ecosystem diversifies, AI will be the differentiator
Power
Plant
revenue
Behind-
the-meter
asset
dispatch
Electri-
fication
EV
Charging
Data
centers
Controllable
Load
Asset
performance
optimization
Carbon
Emissions
V2G
Green H2
Resi solar
& battery
Water
control
VPP
Extreme
weather
Disaster
prevention
Real Zero
?
?
?
?
?
?
?
?

CPV Retail
S I L V E R S P R I N G | B R A I N T R E E | S U G A R L A N D
Responsible Energy Starts with Us
Qadir Khan
October 24, 2023
CPV Retail

Responsible Energy Starts With Us
2
www.cpv.com
CPV: Overview
Competitive Power Ventures (CPV) is a leading electric generation project development and asset management company dedicated to
increasing America’s energy sustainability by providing safe, reliable, cost effective and environmentally-responsible electric power.

3
www.cpv.com
CPV Retail Overview
u Value Proposition
ü We are an environmentally focused retailor helping commercial and
industrial customers achieve their sustainability goals.
ü Retail is an additional sales channel for CPV’s generation assets leading to a
Renewable ‘GenTailer’ integrated company
u Vision
ü Our vision is to be a "Greentailer" in the Retail ecosystem by introducing
"E" products to help our customers achieve their environmental goals by
offering renewable energy products and then facilitating our customer’s
reporting for purposes of Scope 2 pursuant to Greenhouse Gas (GHG)
Protocol
u Strategy
ü Focus on large commercial and industrial customers
3

4
www.cpv.com
Customer Sustainability Goals ( Focus on “E” part of ESG)
u 100% Renewable Energy: Purchase enough
energy to match their consumption, it may reduce
some but not all of their emissions
u Carbon Neutral: Purchase carbon offsets to
compensate for the emissions that they produce.
u 24/7 Carbon-free Energy: matches electricity
demand with Carbon-Free Energy generation in
each hour and on the grid where the demand
occurs. This eliminates carbon associated with
an organization’s electricity use.
u Other customer specific initiatives

5
www.cpv.com
Customer approaches to meet sustainability goals
u Onsite self development owned solar and wind
u Grid directed renewables and carbon free electricity
u RECs Renewable Energy Certificates
(environmental attribute of MWh of renewable
energy generated)
u PPAs Power Purchase Agreements 3rd party owned
projects electricity bundled with carbon free
attributes
u vPPAs Virtual Power Purchase Agreements 3rd
party owned off site projects.

6
www.cpv.com
The Power of Ai to Accelerate from REP perspective
• AI starts with DATA, DATA and DATA
• AI improves business outcomes by leveraging data. It automates and personalize at scale.
• AI is helping companies optimize energy consumption, deploy renewable energy sources
• With the proper AI technology and energy expertise, any organization can tap the potential of AI to
reduce operating costs while moving the needle on sustainability
• By analyzing historical data and using predictive modeling, AI can help companies identify trends
and patterns in their carbon emissions and develop strategies to reduce them
• CPV Retail is utilizing tools to predict energy generation ( all sources )
• Using tools to analyze massive volume of data to help optimize the energy fleet from both
grid and trading perspective
• Renewable sources are unpredictable and AI tools can help manage that inefficiency
• CPV Retail uses AI to help with managing load serving obligation
• AI algorithms can analyze historical energy consumption, patterns to predict future energy
demand ( help with their sustainability targets)
• Everything starts from knowing the consumption of customers ( carbon emissions, peak
consumption hours)

7
www.cpv.com
The Power of Ai to Accelerate from REP perspective
• Retail Product
• We are developing a product using AI tool to match customers hourly consumption
with carbon free resources including CPV carbon free assets to help companies
achieve their sustainability goals (24/7 Carbon free product)
• While the 24/7 product is still in works, these intermediate tools can support a path to
other ESG goals
• Use machine learning to analyze customer historical usage, current system load, and
prices to predict future customer usage and determine optimal customer operations
behavior in real-time to reduce carbon intensity
• Calculate locational carbon footprint for customers at hourly granularity through
combination of actual customer volumes, carbon free resource procurement by
customer, and locational marginal emissions attributes from system power
• Create a carbon data and analytics infrastructure, providing high-resolution data on
scope 2 to give customers full visibility into their carbon footprint in order to document
decarbonization with facts
• Help companies manage energy spending and help them achieve their objectives

Tutorial
Decision Points and Practical
Considerations for AI Projects
#VERGE23
CHARLES TRIPP
Senior Scientist:
Artificial Intelligence,
National Renewable
Energy Laboratory

Decision Points and Practical Considerations for AI Projects
VERGE ‘23
Charles Tripp, Ambarish Nag, Sagi Zizman, Jordan Perr-Sauer,
Jamil Garfur, Hilary Egan, Nicholas Wimer

JISEA—Joint Institute for Strategic Energy Analysis 2
Agenda
❖ What is NREL and what AI systems do we work on?
❖ Challenges and Practical Considerations for AI implementations:
Questions, Decisions, Tips and Strategies
❖ Challenges Arising from Input Data
❖Costs, Risks, Biases, Limitations
❖ AI Trust Issues
❖Identifying and Mitigating Risks of AI misbehavior
❖ AI System Costs, Trends and Trade-Offs
❖ Energy, Compute, and Time

Green AI @ NREL
AI Researchers at NREL research and apply AI to address commercial, national, and global
energy efficiency and renewable energy challenges.

Green AI @ NREL

Green AI @ NREL
❖ AI for Energy-Efficient Computing
▪ Grid-Integrated, Carbon-Aware Datacenters
▪ Energy & carbon measurement, estimation,
characterization
▪ Energy-Efficient Algorithms
▪ Deep Learning
❖ AI for Mobility Systems
▪ Connected/autonomous vehicles,
infrastructure
▪ Energy-efficient transit systems
❖ AI for Energy Systems
▪ Grid operations
▪ Renewables
▪ Storage
▪ Cybersecurity
❖ AI for Materials
▪ Materials Discovery
▪ Battery Systems
▪ Semiconductors & Photovoltaics
❖ AI for Building Systems
▪ HVAC operations and coordination
▪ Grid- and Mobility-Integrated Buildings

Green AI @ NREL
Let’s Advance the State-of-the-Art in AI-driven Efficiency and Decarbonization Together!
❖ AI for Energy-Efficient Computing
▪ Grid-Integrated, Carbon-Aware Datacenters
▪ Energy & carbon measurement, estimation,
characterization
▪ Energy-Efficient Algorithms
▪ Deep Learning
❖ AI for Mobility Systems
▪ Connected/autonomous vehicles,
infrastructure
▪ Energy-efficient transit systems
❖ AI for Energy Systems
▪ Grid operations
▪ Renewables
▪ Storage
▪ Cybersecurity
❖ AI for Materials
▪ Materials Discovery
▪ Battery Systems
▪ Semiconductors & Photovoltaics
❖ AI for Building Systems
▪ HVAC operations and coordination
▪ Grid- and Mobility-Integrated Buildings

Data: Collection, Quantity, Quality
– What are the costs of collecting and cleaning the input data?
• Monetary
• Temporal
• Privacy / Data Sharing
• Storage & Retrieval
• Collection Quality
– What limitations does the data impose on the system?
• Performance Domain & Limitations
• Performance Quality: How good of a job can we do with the data we have?
• Are there biases inherent in the training and/or test datasets?
– Are there DEI issues with the datasets? What can we do to mitigate these
issues?
– What are the risks of dirty or malicious training data?
• Could a bad actor inject ‘poisoned’ data to influence system behavior?

• Utilized
• Thermal video cameras (1,304 hours)
• Near-infrared video
• Acoustic detectors
• Radar (3-4 million animals detected)
• Bat behavior
• Many bats passing close to WT
stationary or slow-moving
• Wind speed and blade rotation
influenced behavior
• Approach less frequently with fast
spinning WT
• Bird behavior
• Far out numbered bats (Radar)
• Absence from video observations
• Suggesting no interaction with WT Bats at wind turbines, Paul. M. Cryan et al. Proceedings of the National Academy of Sciences Oct
2014, 111 (42) 15126 15131; DOI:10.1073/pnas.1406672111
Detecting Bat and Bird Activity near Wind Turbines
Work led by John Yarbrough

NREL | 9
Stochastic Soaring Raptor Simulator (SSRS)
Orographic Updraft Field Simulated Eagle Tracks
Relative Presence Density
Turbine-scale Presence
Turbine Control
Avian Detection
Atmosphere/Topography
Plant Siting
Work led by Charles Tripp, Rimple Sandhu, Eliot Quon, Regis Thedin

AI Trustability
Consider the probability and consequences of “bad AI behavior”
– Safety: Could this system waste money, break something, break a contract,
break the law, injure someone?
– Business & Legal Risks
• Data disclosure and security
• Copyright Infringement
• AI-Based Discrimination: Are there DEI challenges facing this system?
– Vulnerability to malicious actors
• Data poisoning
– Out-of-sample behavior
• How likely is the system to encounter untrained scenarios / inputs?
• What might happen if the system does not behave as desired in these scenarios?
• Are there feasible safeguards to mitigate these risks?
• Are there ways of bolstering training data to cover system blind spots?
More Details: Baker et al. Workshop Report on Basic Research Needs for Scientific Machine Learning: Core Technologies for Artificial Intelligence.
United States. https://doi.org/10.2172/1478744

▪ Job Energy Prediction
▪ Energy, Cost, and Carbon-Aware
Scheduling
▪ Anomaly detection
▪ Predictive maintenance
▪ Operational optimization (PUE)
Measured Job Power [W]
Predicted
Job
Power
[W]
System
Power
[kW]
600
500
400
300
200
100
100 200 300 400 500 600
Grid-Integrated, Carbon-Aware Computing
Work led by Caleb Phillips, Hilary Egan

Mitigating AI Trust Issues
– Are there safeguards we can implement to constrain or manage system
outputs?
• Using known-good baseline systems to limit control outputs.
– Can we detect “bad behavior”, or detect “dangerous outputs” before
they cause a problem?
• Anomaly detection systems
– Choosing a Level of Autonomy: What kind of oversight do we need to
mitigate system risks?
– High risk: use AI systems to advise and assist a human practitioner who is
trained to understand and manage the limitations and failure modes of the
system
– Moderate risk: implement anomaly detection and human monitoring
– Low risk: allow day-to-day autonomy but maintain a reasonable level of
oversight, spot-checks, and validation
– Explainability: can we know why it did what it did?

Autonomous Vehicle Fleet Assignment
Work led by Dave Biagioni
• Objective
• Optimize fleet assignment under a
variety of scenarios where all trips in the
city are served by connected
autonomous vehicle (CAVS) fleet
• Impacts
• Reduce empty-passenger miles traveled
• Save energy and operation cost
Reinforcement
Learning
Optimization
Engine
Current trip
demand
Current CAVs
supply
Optimum
solution for fleet
assignment

Materials Discovery
Work led by Peter St. John
• AlphaZero Reinforcement Learning uses self-play to explore large action
spaces and decouples rollouts from policy updates
• Inherently scalable design (demonstrated with thousands of TPUs),
leveraging GPUs in both rollouts (policy evaluations) and policy training
0
100
200
Reward
games r75
0 1 2 3 4
Time (hours)
0.0
0.2
0.4
Policy
Training
value loss prior loss
HO HS
16.5% 14.8%
HO
30.0% 15.3%
CH4
HO
30.4%
HO
15.7%
HO
start
O
SH
final radical
A B C
Molecule
Rollouts
Policy
Model
Data
Buffer
workers
(node 1)
workers
(node n)
in-progress
and final
molecules
and
reward
sample intermediate
states and reward
from most recent
games
predictions
for final
value and
visit
priors

Red AI: Exploding Computational Costs
Historically the
computational cost of AI
grew with our computers.
But, in the last decade AI
growth has far
outstripped the growth in
computing power.
Work led by Charles Tripp

AI Compute Time Doubles Every 4-6 Months
Also growing rapidly:
• AI Compute Costs
• AI Data Requirements

AI Energy Costs Double Every 4-6 Months
Also growing rapidly:
• Inference Energy
• AI Deployment
• Carbon Footprint

The AI Performance – Energy Trade-off
• Larger models can achieve higher performance but are substantially less
efficient.
• Even for achieving lower performance targets.
• We are developing training methods that walk along the optimal frontier
Work led by Charles Tripp, Jordan Perr-Sauer

This work was authored by the National Renewable Energy Laboratory, operated by Alliance for
Sustainable Energy, LLC, for the U.S. Department of Energy (DOE) under Contract No. DE-AC36-
08GO28308. Funding provided by the Joint Institute for Strategic Energy Analysis, and the National
Renewable Energy Laboratory. The views expressed herein do not necessarily represent the views
of the DOE, the U.S. Government, or sponsors.
Thank you!

Making AI More Sustainable:
Innovations for the Enterprise and Data
Center
Jen Huffstetler
Chief Product Sustainability Officer
VP, GM Data Center and AI Group
October 24, 2023

2
ESG Office
Priorities
Carbon Emissions –
Scope 1, 2, and 3
Water Stewardship
Circular Economy
Diversity and Inclusion
Bringing AI Everywhere
Chief Information Office
(CIO)
Priorities
Business Transformation
through AI
Total Cost of Ownership
of AI capabilities
Digital Security
Desired
Outcomes
ü Reduced carbon
emissions through AI-
enabled energy control
ü Lower water
consumption through
reduced energy
ü Increased equipment
recycling
ü Secure deployment of
AI to all departments
ESG and CIO Offices Partnering for More Sustainable AI

3
The Challenge: Energy Consumption for GenAI
Source: Stanford HAI AI Report 2023 page 121
Source: Stanford HAI AI Report 2023 page 120
Parameters
(billion)
GPT-3
Gopher
Training
Power Consumption
(MWh)
GPT-3
Gopher
5.31
Human Life
Avg, 1 year
Car, avg incl fuel
(lifetime)
GPT-3
Gopher
Training
CO2 equivalent emissions
(tonnes)
Data shown for model training
Inference
60%
Training
40%
Google ML Energy
Consumption*
*https://www.nasdaq.com/articles/generative-ais-hidden-cost%3A-its-impact-on-the-environment

4
Making AI More Efficient through
Optimized Models and Software Energy
Consumption
Large Models - Used by Large Cloud Service Providers
answering all the world’s questions
Optimized Models – Used by Enterprise
answering domain specific questions
Right-size the model through optimization
compression, prune, distill
Optimized Software for
Platforms and Frameworks
Intel AI Analytics Toolkit

5
Developing and Deploying AI Hardware More Sustainably
More modular
equipment
=
Less eWaste to
landfills
Processors:
General Purpose
Dedicated
Liquid Cooling:
Cold-plate
Immersion
Modular Design:
Upgradable
Recyclable
Better performance /
watt for AI
=
Lower Scope 2
carbon emissions
More efficient data
centers
=
Lower Scope 2
carbon emissions

6
Six Best Practices for More Sustainable AI
1. Emphasize data quality over data quantity
2. Consider the level of accuracy
3. Leverage domain-specific models
4. Balance your hardware and software from edge to cloud
5. Consider open-source solutions
6. Integrate Carbon Aware Software

8
Customer Success
AI-based auto contouring
for radiation therapy1
1. Source: Intel Case Study
35x
Faster
20%
Less Power
~10%
Reduction in overall power
consumption
AI-based load prediction
Automatic CPU frequency
tuning2
AI-based workload
prediction and scaling of
resources3
28%
Reduction in power
consumption
Results may vary

VERGE 23: A Practical Guide to Harnessing AI for Decarbonization

More Related Content

What's hot

Similar to VERGE 23: A Practical Guide to Harnessing AI for Decarbonization

More from GreenBiz Group

Recently uploaded

VERGE 23: A Practical Guide to Harnessing AI for Decarbonization