Short Term Solar Forecasting with Skycams
INTERNATIONAL CONFERENCE OF ENERGY & METEOROLOGY - JUNE 2015, BOULDER, COLORADO
Sam West
Senior Experimental Scientist
CSIRO Energy, Newcastle, Australia
Contents
1.
2.
3.
4.
5.
6.
What & Why
Cloud Classification Model
Cloud Presence Detection Demo
Cloud Motion Vector Forecasting Demo
Shade Event Forecast Performance
Skycam Network Rollout
What?
• (Very) short-term solar forecasting
• 10-30 min forecast horizon, 10s update
• Inexpensive whole-sky cameras ($800)
• Low cost hardware means:
• Widespread deployment more practical, but
• Image processing more challenging
Why?
•
•
•
•
Remote Area Power Systems (RAPS): save fuel, but avoid blackouts
Solar Farms: avoid ramp effects, improve market operation
Concentrating Solar Thermal: avoid thermal stress and component damage
Distributed PV: improve grid power/load forecasts, encourage solar penetration
Cloud Classification – Unprocessed Image
Red-Blue Ratio (RBR) Classifier
+ Good performance for overcast & dark clouds
- Misclassifies near sun
- Misses thin and near-horizon clouds
Random Forest Classifier
+ Sensitive to cloud edges, thin & distant clouds
+ Good near-sun performance
Misses dark & non-textured cloud areas
Combined RBR + Random Forest Classifiers
• Bright red = models agree
• Darker red = only one model detects cloud
+ The two models are complementary
New Model: Final Combined Result
Random Forest Supervised Classifier Training
• Human manually identifies cloud and sky pixels
• Train classifier using inputs for each pixel:
–
–
–
–
–
–
Pixel Red-Green-Blue
Hue-Saturation-Brightness
Red Blue Ratio & Difference
Sun Position
Distance to Sun
Pixel movement
• 97% classification accuracy on 1,000,000 pixel test-set
Cloud Presence Detection Demo
•
•
•
•
Sudden cloud formation event
Left chart: measured DNI, GHI, PV Power
Red histogram bars show cloud % in concentric rings around sun
Sharp increase, ~7 min before gives ample warning of shading event
Cloud Motion Vector Forecasting Demo
• Clear morning, intermittent afternoon, approaching cloud front
• Detected 25 minutes in advance of shade event
• Left chart: measured DNI, GHI, Diffuse, PV Power, and forecast Cloud Pixel Fraction
Cloud Front Timing Forecasts
Ramp Timing Prediction – Binary Shading Model
• Evaluation of ramp-event timing forecast on 2 month dataset
Misprediction Rate (%)
Model
Aggressive detection
Conservative detection
Persistence Model
Detection Sky State
Threshold Correct (%)
4600
96
200
61
N/A
88
False Positive
(predicted sky
when cloud)
2.3
0.04
5.9
False Negative
(predicted cloud
when sky)
2.2
39
5.6
On average, 0.04% false positive = all shade events predicted
for 42 days straight, before a single ramp event is missed.
More Details: S. R. West, D. Rowe, S. Sayeef, and A. Berry, “Short-term irradiance forecasting using
skycams: Motivation and development,” Sol. Energy, vol. 110, pp. 188–207, Dec. 2014.
Camera Network
• Cover whole city with handful of cameras
• 15 sites around Canberra & Newcastle for wide-area ramp and
irradiance forecasting
Canberra
Newcastle
Camera Spatial Range Verification
• To plan locations for network of cameras, need camera’s spatial range
• Methodology:
•
•
•
•
•
Two cameras with known distance
Find clouds moving between cameras
Calculate ground velocity from time between cameras
Calculate cloud base height = f(image angular velocity, ground velocity)
Verified height with ceilometer
Camera Spatial Range Verification
Cloud Angle vs Horizontal Distance
55
50
HORIZONTAL DISTANCE (KM)
45
40
Low cloud (cumulus): sky is visible for
10km radius around camera
35
30
25
20
15
10
5
0
-5
0
5
10
15
Cirrostratus1 10-12km
20
25
30
35
40
45
ZENITH ANGLE
Cirrostratus2 10-11.5km
Cirrus 7-7.5km
50
55
60
Altocumulus 5-6km
65
70
75
Cumulus 2.5-3km
80
Practical Challenges: Vandals!
Thanks!
Sam West
[email protected]