Machine Learning for a dynamic sizing of Elia balancing reserves!
1. Dynamic Dimensioning of Balancing
Reserves with Machine Learning Algorithms
Jose Gonzalez Pastor – Adequacy Analyst - Elia System Operator
Guillaume Leclercq – Senior Consultant Grid & Markets - N-SIDE
Grid Analytics Europe 2018
London, September 26th
1
2. Agenda
How to balance power systems with increasing
uncertainty driven by renewable generation?
How can machine learning provide solutions to
predict and cope with this uncertainty ?
Implementation of a dynamic dimensioning method
for reserve capacity to balance the power system
3. • Imbalances result in frequency deviations, and large deviations result in
protective disconnection of generation units and loads, and eventually system
black-out.
➢ Interconnected system: each TSO is responsible for the balance in its control zone
➢ Unbundled system: each market party is responsible for the balance in its portfolio
Balancing increasing uncertainty in power systems
∑Pload = ∑Pgeneration
• Power plant outages
• Unforeseen variations
of renewable and
sustainable generation
CHALLENGE: Increasing
shares of (offshore) wind
and photovoltaic power
• Unforeseen demand
variations
CHALLENGE:
Electrification of heating
and mobility
Uncertainty in Generation Uncertainty in Demand
• Transmission line
outages
CHALLENGE:
integration of HVDC-
interconnectors
Uncertainty in the Network
4. Elia incentivizes market players to balance their portfolio (imbalance tariffs)
- Facilitated by using day-ahead and intra-day markets
- Facilitated by using balancing markets (allowing the participation of new technologies for flexibility)
Elia contracts reserve capacity (to ensure availability), which is activated in real-time to cover residual imbalances
• aFRR (R2): fast-response reserves to react on sudden imbalances and variations.
• mFRR (R3): slower-response reserve to cover larger imbalances, and for longer periods.
Elia calculates each year the required reserve capacity needs to cover the residual imbalances :
• The current ‘static’ reserve capacity dimensioning fixes the reserve capacity only once a year.
• Does not recognize that risk for imbalances in the system may depend on system conditions.
Reserve capacity for balancing the power system
Elia, as transmission system operator, is responsible for covering the residual system
imbalances between injection and off-take in its control zone by using reserve capacity
5. • Cover risk of imbalance due to power plant outage
• Cover volatility of system imbalance
‘Static’ Reserve Dimensioning
5
Loss of power pdf due to outages in generation
Historical imbalance
Convolution
Loss of power (MW)
Probability
Imbalance – no Forced Outage (MW)
Expected Imbalance (MW)
99%
Imbalance pdf
6. • Resulting dimensioning is, currently, mainly driven by the power plant outage
• Loss of one nuclear reactor (1GW)
• Increasing renewable sources penetration
• 2018: 1 GW of off-shore wind
• 2020: 2,3 GW of off-shore wind
This will shift the ‘static’ dimensioning result as it will be driven by the increasing volatility of system
imbalance
‘Static’ Reserve Dimensioning
6
7. • Quick win, assess the risk of power plant outage depending on their availability
Dynamic Dimensioning – Dynamic outage risk assessment
7
Historical
Outages
Analysis
Simulation
Historical
Outages
Analysis
Simulation
Loss of power (MW)
Loss of power (MW)
(unit not scheduled)
Probability
Probability
Yearly Dimensioning
of outage risk
Dynamic Dimensioning
of outage risk
8. Dynamic Dimensioning – Dynamic imbalance volatility
assessment
8
System Conditions
High Wind – Medium Solar
Is the system imbalance correlated to the system conditions?
Wind forecast
Solar (PV)
forecast
9. • System imbalances are mainly driven by :
• Prediction risks: wind, photovoltaics, load forecast errors
• Outage risks: power plant and transmission line outages
• Market risks: ability of market players to balance their portfolio within 15’
• System imbalances are correlated with predicted system conditions (e.g. power
plant and renewable generation schedules) :
• E.g. higher risk for shortages when predicting higher wind conditions
• E.g. reduced outage risks when power plants are scheduled as unavailable
• Predicting the system imbalance risk allows Elia to dimension its reserve
capacity needs to the risks of the system. This requires :
• Mapping correlations between imbalances and predicted system conditions
• Leveraging these correlations to day-ahead predicted system conditions
Predicting the System Imbalance Risk
9
10. Agenda
How to balance power systems with increasing
uncertainty driven by renewable generation?
How can machine learning provide solutions to
predict and cope with this uncertainty ?
Implementation of a dynamic dimensioning method
for reserve capacity to balance the power system
11. 11
How to size the reserve dynamically ?1
RISK linked to FORCED OUTAGES RISK linked to Forecast Errors
UP
FRR
DOWN
FRR
• Power plant outage have
effect on SI
• Considered separately as
rare event for BE system
• Modelled via a “Monte
Carlo” simulator
HOW TO MAKE IT DYNAMIC?
Considering a variable input
in the Monte Carlo, e.g.:
• DA Power units schedule
• HVDC interconnector
schedule
• …
HOW TO MAKE IT DYNAMIC?
Need to map system
conditions (wind forecast, PV
forecast…) to the imbalance
risk…
→ How to map the system
conditions to system
imbalance?
Risk is represented
through past system
imbalance data
1 Approach proposed in the DRS study and proof of concept performed by ELIA with N-SIDE support
12. 12
How to map the system conditions to system imbalance?
Feature 1 e.g. WIND FORECAST
Step 1: sort the imbalance
measures depending on
features (e.g. PV & WIND)
Historical measures of imbalance
e.g. SI = -25MW
e.g. SI = 122MW
e.g. SI = 172MW
Feature2e.g.PVFORECAST
13. 13
How to map the system conditions to system imbalance?
Feature 1 e.g. WIND FORECAST
Feature2e.g.PVFORECAST
STATIC sizing would just
use all the data without
considering the features
14. 14
How to map the system conditions to system imbalance?
Feature 1 e.g. WIND FORECAST
Feature2e.g.PVFORECAST
SCENARIO 2
high wind – high PV
SCENARIO 1
low wind – high PV
SCENARIO 3
low wind – low PV
SCENARIO 4
high wind – low PV
To make it DYNAMIC, you need to
create “several scenarios”
• They represent several
situations that could occur
• To each scenario are
associated past imbalance
data and sizing corresponding
to predefined reliability
15. 15
How to map the system conditions to system imbalance?
Feature 1 e.g. WIND FORECAST
Feature2e.g.PVFORECAST
This approach seems to work qualitatively… The problem is how to
quantitatively compute these scenarios?
• Each features is split in 2? 3? 4?...
• How to compute each interval?
How to design a methodology to compute these parameters smartly?
Even if an agreement is found on these parameters, how to
automatically update them with new data (lot of data)?
More features are needed to properly make the mapping → with this
basic method, the number of scenarios will grow exponentially, e.g. if
each features is cut in 2:
• 2 features → 4 scenarios
• 10 features (which is reasonable) → 1024 scenarios
These are key concerns that can be
addressed with MACHINE LEARNING
What are the issues & open
questions with such approach?
17. 17
99.9%
reliability
COMPLEXITY
“DISCRETE” MAPPING
“Qualitativ
e
clustering”
“KMEANS”
“KNN”
K-NN/deep learning
(function estimation)
Feature 1
Feature2
Feature 1
Feature2
Feature 1
Feature2
Automatic
and smart
“clustering
” (i.e.
scenarios)
“CONTINUOUS” MAPPING
Local
grouping (no
predefined
scenarios)
Machine learning offers powerful tools to smartly map the
system conditions to imbalance
• Neural network and deep learning are
powerful machine learning approaches, also
considered for dynamic dimensioning of
reserve
• However, they are not the most suited for this
application for technical reasons linked to the
methodology (i.e. the need to convolve a
distribution with FO distribution…)
• Though they could be applicable considering
some adaptations
• More advanced technics relying on
hybridation of several ML algorithms are also
considered in the industrialization phase
DEEP LEARNING & MORE
ADVANCED APPROACHES?
18. 18
Machine learning Methodology: from Model Design to Prediction
- 1 -
MODEL
DESIGN
OUTPUT
theoretical ML
model:
• # clusters?
• Features?
• Parameters of
the algo…
MODEL 1
MODEL 2
MODEL 3
MODEL N
…
KPI 1 KPI k…❖ Define a list of possible models
❖ TRAIN & VALIDATE on historical data
(get the most out of the data)
19. 19
Feature 1
Feature2
- 1 -
MODEL
DESIGN
OUTPUT
theoretical ML
model:
• # clusters?
• Features?
• Parameters of
the algo…
- 2 -
TRAINING
OUTPUT
The practical
trained ML
model
DATA INPUTS
Historical data (>1 year) of SI and
system features (DA forecast of
wind, PV…)
…
…
…
Cluster 2
Cluster 1
Cluster 3
Cluster 4
Cluster 5
…
Up reserve = xxx
Down reserve = xxx
Machine learning Methodology: from Model Design to Prediction
20. 20
- 1 -
MODEL
DESIGN
OUTPUT
theoretical ML
model:
• # clusters?
• Features?
• Parameters of
the algo…
- 2 -
TRAINING
OUTPUT
practical trained
ML model
- 3 -
PREDICTION
Imbalance
risk for the
next day
DATA INPUTS
Historical data (>1 year) of SI and
system features (DA forecast of
wind, PV…)
DATA INPUTS
Day-ahead system conditions for
the next day
Feature 1
Feature2
For each hour of tomorrow,
check in which cluster we are.
E.g. Tomorrow at 10am is cluster 3!
Cluster 1
Cluster 2
Cluster 4
Cluster 5
Cluster 3 Up reserve = xxx
Down reserve = xxxOrange point = features
prediction for next day
Machine learning Methodology: from Model Design to Prediction
21. Agenda
How to balance power systems with increasing
uncertainty driven by renewable generation?
How can machine learning provide solutions to
predict and cope with this uncertainty ?
Implementation of a dynamic dimensioning method
for reserve capacity to balance the power system
22. • A better reliability management with higher FRR during higher risk period
• Financial gains following average FRR reductions (outweighing the implementation costs)
• A robust methodology towards the middle and long term system with more uncertainty driven by renewables
➢ With increasing advantages with higher renewable generation shares in 2027
Proof of Concept on 2020 and 2027 shows
22
23. Proof of Concept confirms the potential of
dynamic sizing to save volumes while ensuring better the reliability
VOLUME
SAVINGS
Dynamic sizing can save
up to 200 MW
and is 85%/time < static sizing
Savings on both upward and
downward side
DYNAMIC
BEHAVIOUR
Dynamic sizing
potential
confirmed
1100
1200
1300
1400
1500
1600
1700
CONTINUOUS NEIGHBORS
static
1320
1340
1360
1380
1400
1420
1440
UPWARD
1180
1190
1200
1210
1220
1230
1240
1250
1260
DOWNWARD
60MW
savings
45MW
savings
KNN
MORE
CONSTANT
RELIABILITY
INSTANTANEOUS RELIABILITY UPWARD
1304
1457
1417
1417
99.91%
99.96%
99.89%
99.86%
LOWRISKHIGHRISK
STATIC SIZE DYNAMIC SIZE
More constant reliability ensured with
dynamic sizing, regardless
whether high or low risk period
24. • Difference between laboratory and operational, real-time, environment
• How can the operator be confident on the result of the machine learning algorithm?
• Monthly training of the algorithm, with validation of performance on past dates
• Ex-post assessment of algorithm performance
• What has to be done in case of failure, or strange results, technical problem?
• Foresee fallback scenarios to be applied on real-time
Implementation – Business Impact
24
Requires review of business processes, both operational, ex-ante and ex-post
25. 1250
1350
1450
1550
1650
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Upward Reserve
7000
12000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Forecasted Load
0
5000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Forecasted Wind
-2000
0
2000
4000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Forecasted PV
0
2000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Nemo Import Volume to be covered
Week 24 September – 30 September
• Night Time
• Low Wind
• No PV
Saturday 24/09 2am to 6am
static
Max load
Max Wind
Max PV
Lower upward FRR need
Time series analysis brings intuition on the sizing
generated with the Machine Learning methods …
26. 1250
1350
1450
1550
1650
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Upward Reserve
static
7000
12000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Forecasted Load
0
5000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Forecasted Wind
-2000
0
2000
4000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Forecasted PV
0
2000
24.09.16 25.09.16 26.09.16 27.09.16 28.09.16 29.09.16 30.09.16 01.10.16
Nemo Import Volume to be covered
• Significant Load
• Nemo import covered
• Very High Wind !!!
• Significant PV
Thur. 29/09 12pm to 4pm
Max load
Max Wind
Max PV
Week 24 September – 30 September
… which is an major requirement for adoption by the users
Displayed in the GUI of the industrial tool such that the
user can get confidence into the results !!
Very High Upward FRR
Need
27. Typical steps to develop an
industrial « machine learning » tool
27
POTENTIAL ASSESSMENT &
PROOF OF CONCEPT
INDUSTRIALIZATION
CONTINUOUS
IMPROVEMENT
• Identify and quantify the
potential
• Implement & test several
algorithms with high level
differences
• Assess feasibility of the
approaches
• Ensure no black-box model and
intuition is ensured for potential
users/stakeholders
• Industrialization is NOT just
“translating” the PoC !
• Implies the use of mature and stable
advanced analytics modules
Effort TO BUILD A PLATFORM
• Automatize data ecosystem
management
• Robustness (tool needs to run every
day!): fall-back solution.
• Build an intuitive interface
• Adapt and improve the tool as the
market changes continuously
• Improve the advanced analytics
methods according to the
technological improvements
• Improve the process to use the
tool and the tool based on the
feedback of the key users
1 2 3
28. Architecture of the dynamic reserve dimensioning
industrial tool
• Tool built on a platform managing the data connections internally and externally (other
tools, DB)
• User Interface for the users (training user, prediction user)
• Specific analytic module where the advanced analytics algorithms used in the POC are
developed
29. Challenges and lessons learnt
29
INDUSTRIALIZATION
• Better to gradually increase the the complexity of tool to allow gradual acceptation of the tool and an efficient usage in
operation/production
• In our case, move from static sizing every year > sizing based on planned outages every day > full dynamic sizing based
on machine learning
• Data quality issues are very important
• Quick fixes can be easily developed in a PoC study while it is not the case in an industrial tool
• Integration challenges with existing tools
• Non-existing in PoC
• Very important and time consuming in industrial tool
• Robustness of the tool
• Need for many tests
• Programming code quality, review, , robustness and documentation
• Not so much time in applying the advanced analytics methods, but rather everything around
2
30. Conclusions
30
• Power system operation faces increasing uncertainty with large-scale integration of new technologies.
• Machine learning can help capturing this system behavior, and allow to better cope with these evolutions
• Efficiency : reducing average reserve capacity
• Reliability : ensuring adequate reserve capacity during high needs
• Sustainability : facilitating the integration of renewable energy
• Elia is pursuing implementation of this method with 2020 as target for operational
• Transition from Proof Of Concept towards operational solution is not straightforward and requires
• Review of business processes
• Build confidence of users in new method
• Ensure reliability of data and infrastructure