Coordinated Ramping Product and Regulation Reserve ... Solar-Forecast… · • P-P Plot score and...
Transcript of Coordinated Ramping Product and Regulation Reserve ... Solar-Forecast… · • P-P Plot score and...
Coordinated Ramping Product and Regulation Reserve Procurements in CAISO and MISO using
Multi-Scale Probabilistic Solar Power Forecasts (Pro2R)
The Johns Hopkins UniversityAward # DE-EE0008215
Solar Forecasting 2 Annual Review & WorkshopOctober 8, 2019
:
Funded by:
Principal Investigator: B.F. Hobbs, The Johns Hopkins University
This presentation may have proprietary information and is protected from public release.
Pro2R Project Team
Johns Hopkins: Benjamin Hobbs (PI), Qingyu Xu, Josephine Wang
NREL: Venkat Krishnan (Co-PI), Elina Spyrou, Paul Edwards, Haiku SkyIBM: Hendrik Hamann (Co-PI), Rui ZhangUniversity of Texas Dallas: Jie Zhang (Co-PI), Binghui Li
Industry Partners: • Amber Motley, Clyde Loutan, Rebecca Webb, Guillermo Bautista
(California ISO)• Blagoy Borissov, Stephen Rose (Midcontinent ISO)
Introduction
2
This presentation may have proprietary information and is protected from public release.
Pro2R Project Summary
Objective: Integrate probabilistic short- (2-3 hr ahead) and mid-term (day-ahead) solar power forecasts into operations of CAISO & MISO
Thrust 2: Integrate probabilistic forecasts in ISO operations for ramp product ®ulation requirements
Thrust 3: Provide situational awareness via visualizations of probabilistic ramp forecasts & alerts
Approach:
Thrust 1: Advanced big data-driven“probabilistic” solar power forecasting technology using IBM Watt-Sun & PAIRS (Big data information processing and machine learning approaches to blend outputs from multiple models).
Introduction
3
This presentation may have proprietary information and is protected from public release.
Thrust 1: SFII Built Upon IBM PAIRS Geoscope Platform
• Distributed computational system• Scales to many hundreds of Petabytes (PB)• Data processing rate at PB/day.
Thrust 1
NAM HRRR-Subhour GOES-R
spatial resolution 12 km 3 km 500 m
temporal resolution 1 hr 15 min 5~10 min
daily data volume 9.4 GB 86 GB 203 GB
4
A: NegativeError
Forecast ErrorGHI
(W/m^2)
C:PositiveError
B:PositiveError
D
16.6
90
Zenith(Deg)
GHI Forecast (W/m^2)0 1.02e3Question: What is the error of the models, when, where, under what weather situation?
NAM GHI Forecast Error (Surfrad BND)– Strongly depends on zenith angle and forecast
irradiance. The two parameters create 4 categories of situations:
Hurricane Ike path forecasts from 8 different weather models*
Situation-Dependent Model Blending for Solar Forecasting
Thrust 1
5
This presentation may have proprietary information and is protected from public release.
Utilization of GOES-R Satellites –In progress– Expected in Watt-Sun 2.0, 2019-20 Q7• Blue band, 0.47 μm: monitor dust, haze, smoke and clouds• Red band, 0.6 μm: detect fog, estimate solar insolation, depict diurnal aspects of clouds• Near-infrared, 0.86 μm: detect daytime clouds, fog, and aerosols; calculate normalized difference vegetation index• Infrared, 10.3 μm: correct atmospheric moisture, estimate cloud particle size, characterize surface properties
Plan:• Translate GOES imagery
to Solar Irradiance raster map
• Use Deep Learning combined with Optic Flow to forecast cloud movement, then translate into forecast solar irradiance map.
Thrust 1
6
7
7Overall Flowchart
Numerical Weather Prediction Models
Model Training
Numerical Weather Prediction Models
Numerical Weather Prediction Models
FANOVA Parameter (feature) selection
Random Forest model to prediction the forecast
error (Train)
Gaussian Mixture Model for clustering into number of weather categories (Train)
Quantile regression
Train
Solar irradiance measurements
Model Forecasting
Numerical Weather Prediction Models
Numerical Weather Prediction Models
Numerical Weather Prediction Models
Use trained Random Forest model to predict
forecast error
Use trained Gaussian Mixture Model & predicted forecast
error to predict weather categories
Use trained Quantile
regression for each category
for probabilistic forecast
Use trained Quantile
regression for each category
for probabilistic forecast
Use trained Quantile
regression for each category
for probabilistic forecast
Quantile regression
Train
Quantile regression
Train
Thrust 1
Used quantile regression to deploy probabilistic forecast models
• Quantiles of solar as function of independent variables• Example results for 2 hr-ahead forecasts
• Distributions are asymmetric need quantile regression techniques• Adjacent days have different distributions; but present CAISO flexiramp requirements are very stable day-to-day
because they don’t reflect weather forecasts need to integrate probabilistic forecasts in requirements
Thrust 1
8
P-P plot-based Error Metric• A P–P plot (probability–probability plot) assesses how closely two data sets agree• By plotting the two cumulative distribution functions against each other
P-P plot (2 hr forecast)Mean absolute difference: 0.0543
Example Empirical CDF Curves (2 hr forecast)
CDF
RankGHI
CDF
Thrust 1
9
P-P plot 2 Hr Ahead Error Metric: Comparison with Bias Corrected HRRR
Empirical CDFMean absolute difference: 0.054
Bias Corrected HRRR (normal distribution)
Mean absolute difference: 0.086
CDF
Rank RankThrust 1
10
• P-P Plot score and MAPE comparison (normalized by max GHI; daylight hours only)• Watt-Sun 1.0 outperforms HRRR Bias Corrected in all sites in terms of MAPE, in most
sites in terms of P-P Plot score
Thrust 1: Forecast Error Report -- CAISO
Thrust 1
11
• P-P Plot score and MAPE comparison (normalized by max GHI, daylight hours only)• Watt-Sun 1.0 outperforms HRRR Bias Corrected in all sites in terms of MAPE, in
most sites in terms of P-P Plot score
Thrust 1: Forecast Error Report -- MISO
Thrust 1
12
Baseline approach CAISO bases FRP requirements on confidence intervals for net load uncertainty by analyzing a rolling window of last 20-40 days for the same hour
0
5,000
10,000
15,000
20,000
25,000
30,000
01.
25 2.5
3.75 5
6.25 7.
58.
75 1011
.25
12.5
13.7
5 1516
.25
17.5
18.7
5 2021
.25
22.5
23.7
5
MW
hour
CAISO RTPD forecast for Net load
5/14/2019 5/15/2019
0
500
1000
1500
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
MW
hours
Uncertainty component (direction down)
5/14/2019 5/15/2019
Thrust 2
Thrust 2: Integration of probabilistic forecasts into ISO operations: Flexible Ramping Product
13
This presentation may have proprietary information and is protected from public release.
Baseline method for FRU uncertainty component potentially leads to under-procurement & price spikes
0
500
1000
1500
2000
2500
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
MW
Max RTD net load – RTPD net load
5/14/2019 5/15/2019 fru
Two days had different net load errors between RTPD & RTD different RTD prices
Source: CAISO OASIS
Estimates of FRP uncertainty component do not necessarily correspond to the target confidence interval
0%
20%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24hours
Percent of 15-min intervals in March-May 2019 that underestimated flexible ramping need (CAISO only)
RTPD FRU < max RTD net load -RTPD net loadRTPD FRD < RTPD net load -min RTD net loadReference: 2.5%
Thrust 2
14
-100
400
900
$/M
Wh
RTD prices 5/15/2019 5/14/2019
This presentation may have proprietary information and is protected from public release.
I. AggregationII. Net load (NL)III. NL forecast errorIV. FRP Requirements
NL forecast error (RTD)
Net load(Binding RTD)
Net load(Advisory RTD)
Convolution
Convolution
Site-level probabilistic solar power
(Advisory RTD)
System-level probabilistic solar power
(Advisory RTD)
Site-level probabilistic solar irradiance
(Advisory RTD)PVlib Convolution
System-level probabilistic load, wind power (Advisory
RTD)
Site-level probabilistic solar power
(Binding RTD)
Site-level probabilistic solar irradiance
(Binding RTD)PVlib Convolution
System-level probabilistic load, wind power (Binding
RTD)
System-level probabilistic load, wind power (Binding
RTD)
Convolution
FRP in RTD
Thrust 2
Probabilistic net load forecast-based ramping product 15
This presentation may have proprietary information and is protected from public release.
Value of probabilistic forecast-based FRP procurement
Compare baseline to probabilistic requirements
Probabilistic 𝐹𝐹𝐹𝐹𝐹𝐹≈ 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐹𝐹𝐹𝐹𝐹𝐹
Probabilistic 𝐹𝐹𝐹𝐹𝐹𝐹> 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐹𝐹𝐹𝐹𝐹𝐹
Actual RTD flex need≤ 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐹𝐹𝐹𝐹𝐹𝐹
Actual RTD flexibility need> 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐹𝐹𝐹𝐹𝐹𝐹
Probabilistic 𝐹𝐹𝐹𝐹𝐹𝐹< 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵 𝐹𝐹𝐹𝐹𝐹𝐹
Actual RTD flex need> 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐵𝐵𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝑃𝑃𝐵𝐵𝑃𝑃 𝐹𝐹𝐹𝐹𝐹𝐹
Actual RTD flex need≤ 𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝑃𝐵𝐵𝑃𝑃𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝑃𝑃𝐵𝐵𝑃𝑃 𝐹𝐹𝐹𝐹𝐹𝐹
Production cost ≈ Increase DecreasePower balance violations (scarcity events)
≈ Decrease Increase
a b c d e
F: Frequency
Fa
Fb Fc Fd Fe
Compare requirements
Actual need (realization)
Performance of prob. forecasts
Thrust 2
Under procurement (Risky)
Over procurement(Conservative)
16
This presentation may have proprietary information and is protected from public release.
Daily profile BaselineExtreme ramps
realized?(Prob Req -Baseline
Req) / (Baseline Req)19th May: (b) intervals Risky Occurred (11%) +40%
6th May: (c) intervals Risky Did not occur
(42%) +79%
12th May: (e) intervals Conservative Did not occur
(44%) -13%
Market simulations: Probabilistic vs. Baseline
Test system: Modified 118 bus IEEE Reliability Test System, mimicking CAISO gen mix (1/10th CAISO system)
Thrust 2.2
Reliability (# Gen. scarcity)
Production cost
Uncertainty-induced costs
Decrease(2713)
Greater(+$2,601) Lower (-60%)
Same(5 5)
Greater(+$920) Higher (+5%)
Same(0 0)
Lower(-$497) Lower (-19%)
• When baseline is riskier we expect reliability benefits when high ramps realized• When baseline is conservative we expect production cost reductions when ramps not realized
0
2
4
6
0
2000
4000
6000
8000
6.50 6.58 6.67 6.75 6.83 6.92 7.00 7.08 7.17
MW
h
$/M
Wh
RTD market results (Baseline)
RTD price RTD unserved energy
Power balance violations during 7 RTD market intervals:Baseline Probabilistic
00.20.40.60.81
-
20
40
60
6.50 6.58 6.67 6.75 6.83 6.92 7.00 7.08 7.17
MW
h
$/M
Wh
RTD market results (Probabilistic)
RTD price RTD unserved energy
No power balance violations in RTD market:
NREL | 18
The Ramp Visualization for Situational Awareness (RaViS) tool provides situational awareness and visualization capabilities using probabilistic solar and net-load time series.
Features:• Integrates IBM forecast data• Refresh rate of 60 seconds• User interface: Single page web application and
open source• Shows site specific metadata via hover• Highly flexible, easily configurable
Future work:• Net-load ramps• Adaptable to other kinds of events: outage/trip,
cyber threatsEvent Outlook
Geolocation All-regions
Geolocation Region Detail
Forecast Detail
Thrust 3: RaViS - Visualization tool
Thrust 3
18
This presentation may have proprietary information and is protected from public release.
Pro2R: What’s Ahead
• Thrust 1: Probabilistic Watt-Sun Versions 2.0 & 3.0: integrates GEOS-R, multi-expert machine learning, & blended models
• Thrust 2:• Statistical/ML modeling of relationship of probabilistic solar forecasts
to net load uncertainty on regulation & FRP timescales• Simulation-based testing on ISO-scale systems of improved
requirements: cost & reliability• Interaction with CAISO & MISO on data, method value, simulations, &
implementation pathways• Thrust 3: RaVIS development, including integrating latest Watt-Sun
methods, net load data, market information, and ISO feedbackConclusion
19
This presentation may have proprietary information and is protected from public release.
Wrap-Up: Pro2R Impact
• Identified the major methodological issues in integrating probabilistic forecasts into system products• BP2 will address those issues and possible improvements in forecasting and
industry practice (e.g., product definition, timeline) • Expected outcome: a blueprint for research in forecasting and industry practice
• Project results expected to highly influence industry practice • ISO staff confirmed that they are evaluating potential improvements to their
existing requirements approach• Integrating probabilistic forecasts is the most promising way to address needs
for requirements to reflect up-to-date weather forecasts
20
Conclusion
This presentation may have proprietary information and is protected from public release.
Questions? 21
This presentation may have proprietary information and is protected from public release.
EXTRA SLIDES
This presentation may have proprietary information and is protected from public release.
Percentage of time when reduction is over 10% FRU FRD
May 27, 2019 38% 25%May 28, 2019 42% 50%May 29, 2019 46% 50%May 30, 2019 58% 38%May 31, 2019 42% 46%
• RTPD• Requirements are increased occasionally due to greater
uncertainties of NL forecast errors• 25%-58% of hours see >10% reductions in FRU and FRD
requirements
Thrust 2.1: Requirements for ramping product
00:00 06:00 12:00 18:00 00:00
May 27, 2019
-2000
-1000
0
1000
2000
MW
Prob
Baseline
00:00 06:00 12:00 18:00 00:00
May 27, 2019
10-2
10-1
100
101
102
Rat
io o
f Pro
b/Ba
selin
e
FRD
FRU
Thrust 2.1