Presentatie 4. Jochen Cremer - TU Delft 28 mei 2024

Dutch Power Event
“AI – Navigeren naar de toekomst?”
Op 28 mei 2024 bij Info Support.

1. Thank you
The role of AI to accelerate the
energy transition from
research perspective
Dr. Jochen Cremer
Assistant Professor
[email protected]
Dutch Power Workshop
28-05-2024

2. Topic
Mission & objective
Education
Research
Keyinnovations
Team
Dynamic Sustainable Energy Systems
•combine groundbreaking ML with the reliable theory
of the physical energy system
•make energy systems sustainable, reliable, effective
▪EE4C12 ML for Electrical Engineering
▪SC42150 Statistical Signal Processing
▪SC42110 Dynamic Programming and Stochastic
Control
▪MOOC Digitalization of Intelligent and Integrated
Energy Systems
▪Crash course of “Data-science”
•Supervised learning for real-time grid
assessment
•Distributed learning for power system
congestion management
•Data-driven grid models for electricity load
and weather forecasts
•Characterizing healthy/normal trajectories
of complex dynamical systems using
dictionary learning
•From fast Fourier transform to fast
reinforcement learning
▪AI-based algorithms for grid operation
▪Real-time security assessment and
anomaly detection
▪Real-time learning algorithms for
control and security of complex
dynamical systems
Haiwei Xie
Mert Karaçelebi
Ali Rajaei
Demetris
Chrysostomou
Peyman Mohajerin
Esfahani
Shabnam
Khodakaramzadeh
Mohammad
Boveiri
Viktor
Zobernig
Olayiwola
Arowolo
Basel Morsy
Jochen Stiasny
Team
https://www.tudelft.nl/ai/delft-ai -energy-lab
DAIEnergyLab
[email protected]
Jochen Cremer
Delft AI Energy Lab
Benjamin
Habib
Runyao
Yu
Perine
Cunat

3. Challenge: Highly complex systems
AI: Decentralised multi-agent
algorithms
Challenge: Standard software slow
AI: Real-time-fast algorithms
3
Why AI for Electrical Power Systems (EPS)?

4. Houston, Texas 07 Feb 2021 Houston, Texas 16 Feb 2021
•Damages from the blackouts were estimated at $195 billion
•Seconds away from a total power blackout in Texas
4
Software for control room: Predict Blackouts

5. 5
design an AI-based algorithm that can
predict power blackouts
Define
requirements
Design Test & analyse
Product
Our goal today:

6. 6
•Data
•AI method & model
•Engineers
What do we need for an AI-based algorithm
to predict power blackouts?

7. 7
•Temporal
•Discrete
•Continuous
Different types of energy data

8. 8
•Example: Load profile, renewable generation
Load Profile
Wind generation
https://dataviewer.pjm.com/
Temporal data

9. 10
•Whatpatterndoyouseeintheelectricityconsumptiondata?
https://williamkoehrsen.medium.com/building-energy-data-analysis-part-two-7861b0c6a2d6
Quiz

10. 11
•day of week
•business day vs weekend
•time of day, season
Patterns in energy consumption

11. 12
Sometimes outliers. What happened?
https://blog.scienceandindustrymuseum.org.uk/electricity-half-time-peak/
North WestEngland

12. 13
•Power blackouts happen
very rarely!
Define
requirements
Design Test & analyse Product
Houston, Texas 07 Feb 2021 Houston, Texas 16 Feb 2021
Is this also a problem for our design?

13. •Data
•AI method & model
•Engineers
14
What do we need to design an AI-algorithm to
predict power blackouts?
that includes power blackouts

14. Decision Trees as a model?
????????????
1
insecure
insecuresecure
secure
????????????
2
Two-dimensional example
Decision trees:
•Limited expressive power
•Fantastic interpretability
15

15. Metrics for classification
Predictedand
actual security limits
are different!
16

16. Metrics for classification
TP FN
FP TN
Predicted Class
PositiveNegative
Negative
PositiveTrue
Class
Predictedand
actual security limits
are different!
Confusion Matrix
17

17. Metrics for classification
TP FN
FP TN
Predicted Class
PositiveNegative
Negative
PositiveTrue
Class
Predictedand
actual security limits
are different!
Two types of accurate predictions
:
TN: Is secure and we think it is secure (GOOD)
TP: Is insecure and we think it is insecure (VERY GOOD!)
18

18. Metrics for classification
TP FN
FP TN
Predicted Class
PositiveNegative
Negative
PositiveTrue
Class
This can have a severe effect!
Two types of wrong predictions
:
FP: Is secure but we think it is insecure (BAD)
FN: Is insecure but we think it is secure (VERY BAD!)
Predictedand
actual security limits
are different!
Two types of accurate predictions
:
TN: Is secure and we think it is secure (GOOD)
TP: Is insecure and we think it is insecure (VERY GOOD!)
19

19. Metrics for classification
TP FN
FP TN
TP FN
FP TN
Ratio of correct predictions
Accuracy =
????????????????????????????????????+????????????????????????????????????
????????????????????????????????????+????????????????????????????????????+????????????????????????????????????+????????????????????????????????????
Predicted Class
True
Class
20

20. Metrics for classification
Accuracy =
????????????????????????????????????+????????????????????????????????????
????????????????????????????????????+????????????????????????????????????+????????????????????????????????????+????????????????????????????????????
TP FN
FP TN
TP FN
FP TN
Ratio of correctly found insecure cases to
predictedinsecurepredictions
Precision =
????????????????????????????????????
????????????????????????????????????+????????????????????????????????????
Predicted Class
True
Class
21

21. Metrics for classification
Accuracy =
????????????????????????????????????+????????????????????????????????????
????????????????????????????????????+????????????????????????????????????+????????????????????????????????????+????????????????????????????????????
Precision =
????????????????????????????????????
????????????????????????????????????+????????????????????????????????????
Recall =
????????????????????????????????????
????????????????????????????????????+????????????????????????????????????
TP FN
FP TN
Ratio of correctly found insecure
cases to all insecure cases
Predicted Class
True
Class
22

22. Duality: Precision vs Recall
When do we observe the
highest performance?
Precision
Recall
23

23. Blackout predictions: Precision or Recall?
Houston, Texas 07 Feb 2021 Houston, Texas 16 Feb 2021
24

24. Designed algorithm for predicting blackouts
Data with power
blackouts
25
Recall =
????????????????????????????????????
????????????????????????????????????+????????????????????????????????????
Decision Trees
Ratio of correctly found insecure
cases to all insecure cases

25. How can we move from ideas to implementation?
26
•Interviewed 110 engineers in power systems (with focus on transmission
system operators)
[1] Jochen L. Cremer and Adrian Kelly and Ricardo J. Bessa and Milos Subasic and Panagiotis N. Papadopoulos and Samuel Young andAmar Sagarand
Antoine Marot, “A PioneeringRoadmapforML-DrivenAlgorithmicAdvancementsin Electrical Networks”, https://arxiv.org/pdf/2405.17184, 2024
[2] CIGRE C2.42 Working Group, “The impact of the growing use of ML- AI in the operation and control of power networks from an operational Perspective”

26. Findings
27
•AI-innovations are often rooted in open- source developments
•High testing quality and safety is needed
•Example PowSyBl
N. Omont, Artelys, “Continental-wide coordination of the power grid powered by PowSyBl,” in Linux Foundation Energy Summit,
Paris, 2023.

27. Need to adapt to changing innovation environment
28
Academic and Applied Research Centres
Open-Source Development
AI Testing and Experimentation Facility
Operators
Operators

28. Collaboration between knowledge institutes and private and/or public parties
Shared fundamental research agenda determined together
At least 5 PhD students, a lab manager, and scientific directors
Plus custom made possibilities (postdocs, external PhDs, etc.)
Five year collaboration based on open knowledge development
External focus
PhD students work part of their time at the location of the partner
Knowledge transfer based on the needs (e.g., meetups, internships)
ICAI Labs
29

29. AI for Energy Grids Lab
Flin Verhaasdonk
DecentralAI & Control
Charlotte Cambier van Nooten
GraphNeuralNetworks
Eva De Winkel
ExplainableAI
Shaohong Shi
Risk-based invesmentsand operation
Mission & objective
•enhance the transport capability of distribution grids
•increase the share of renewable energy
•universal access to affordable, reliable and modern
energy services
•combine groundbreaking ML with the reliable theory of
the physical grid
•open- source code and data

30. ICAI is a national network in The Netherlands aimed at technology and talent development between knowledge
institutes, industry, and government in the area of artificial intelligence.
The Netherlands has the talent, the world- class research and the longstanding tradition in AI education to be one
of the world’s top ranked countries in terms of innovation power.
ICAI brings these forces together in a unique national initiative.
ICAI Innovation Centerfor AI
31

31. AI innovation ecosystem
32

32. Thank you for your attention
Jochen Cremer
[email protected]
Reference:
Jochen L. Cremer and Adrian Kelly and Ricardo J.
Bessa and Milos Subasic and Panagiotis N.
Papadopoulos and Samuel Young and Amar Sagar
and Antoine Marot, “A PioneeringRoadmapforML-
DrivenAlgorithmicAdvancementsin Electrical
Networks”, https://arxiv.org/pdf/2405.17184
, 2024