Machine learning for supporting automated performance ... › Websections... · Smart Buildings...
Transcript of Machine learning for supporting automated performance ... › Websections... · Smart Buildings...
![Page 1: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/1.jpg)
dr.ir. Roel Loonen [email protected]
Smart Buildings Symposium – TU Delft – 07/02/2020
Machine learning for supporting automated performance analysis of rooftop PV
![Page 2: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/2.jpg)
Install and forget
► PV systems can fail in many ways:
• Delamination, hotspots, connectors, ageing, hailstorms, etc.
► Sub-optimal PV performance can be detrimental for return-on-investment and sustainability targets
► O&M is essential, but often, there are no clear signs if a system is malfunctioning or not
► Huge opportunity for data analytics and large-scale system monitoring
2
![Page 3: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/3.jpg)
How to quantify performance?
3
• Efficiency
• Yield per installed capacity (kWh/Wp)?
![Page 4: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/4.jpg)
How to quantify performance?
4
![Page 5: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/5.jpg)
How to quantify performance?
5
Typical building applications vs. STC
• Non-optimal tilt and orientation
• Low light conditions
• Temperature effects
• Partly shaded sites
• Year-to-year variability
![Page 6: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/6.jpg)
How to quantify performance?
6
• Efficiency
• Yield per installed capacity (kWh/Wp)
• Performance ratio (PR)
![Page 7: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/7.jpg)
Performance ratio
7
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
![Page 8: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/8.jpg)
Performance ratio
8
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
![Page 9: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/9.jpg)
Performance ratio
9
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
![Page 10: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/10.jpg)
Performance ratio
10
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of multiple PV systems
![Page 11: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/11.jpg)
11
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
Measured power
Performance ratio
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of multiple PV systems
![Page 12: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/12.jpg)
12
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
Measured power
STC power corrected with actual irradiance
Performance ratio
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of multiple PV systems
![Page 13: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/13.jpg)
13
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑟𝑎𝑡𝑖𝑛𝑔
Performance ratio
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of multiple PV systems
![Page 14: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/14.jpg)
14
Available data:
On site:- Psite measured [W]- Site details (Tilt, orientation,
etc)
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑆𝑇𝐶
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of multiple PV systems
Performance ratio
![Page 15: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/15.jpg)
15
Available data:
On site:- Psite measured [W]- Site details (Tilt, orientation,
etc)
At TU/e Solar Measurement Station:- Irradiation data (GHI,DNI, DHI)- Solar position (Azimuth, Zenith)
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑆𝑇𝐶
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of multiple PV systems
Performance ratio
![Page 16: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/16.jpg)
16
Available data:
On site:- Psite measured [W]- Site details (Tilt, orientation, etc)
At TU/e Solar Measurement Station:- Irradiation data (GHI,DNI, DHI)- Solar position (Azimuth, Zenith)
G site calculated
Tool for performance assessment:
PR =𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝑃𝑆𝑇𝐶 ∗𝐺𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝐺𝑆𝑇𝐶
=
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑐𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑆𝑇𝐶
PR is originally invented to be able to assess the quality of themodules. Typical values: 0.7 - 0.9.
It compensates for the effect of different irradiance conditions,but still depends on other variables, such as:Temperature, Degradation, Incident angle, Spectral properties etc.
We intend to use it to compare real life on-site performance of CIGS and C-Si.
Performance ratio
![Page 17: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/17.jpg)
17
Analemma Daily sun path
Sun position at noon in the summer
Sun position at noon in the winter
Visualization
![Page 18: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/18.jpg)
18
Analemma Daily sun path
Sun position at 3 PM in the
winter
Sun position at
3 PM in the summer
Visualization
![Page 19: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/19.jpg)
19
Analemma Analemma - PR
Visualization
![Page 20: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/20.jpg)
20
Analemma Analemma - PR on a shaded site
Visualization
![Page 21: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/21.jpg)
21
Analemma Analemma - PR on a shaded site
OBSTRUCTION
Visualization
![Page 22: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/22.jpg)
22
Analemma Analemma - PR on a shaded site
OBSTRUCTION
OBSTRUCTION
Visualization
![Page 23: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/23.jpg)
Creating analemma from measurement data
23
TU/e SMS
Eliminating datapoints with cloud-shadingBinarizing training dataTrain SVM – find unshaded solar positionsBring back all datapoints and classify them
Performance Ratio
PR =
𝑃𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑𝐺𝑠𝑖𝑡𝑒 𝑐𝑎𝑙𝑐𝑢𝑙𝑎𝑡𝑒𝑑
𝑃𝑆𝑇𝐶𝐺𝑆𝑇𝐶
=𝜂𝑠𝑖𝑡𝑒 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑
𝜂𝑆𝑇𝐶
Evaluate datapoints without local shading
![Page 24: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/24.jpg)
Training Prediction Assessment
24
Shad
e-fr
ee
at
no
on
Shad
e-fr
ee in
th
e m
orn
ing
Shad
e-f
ree
in t
he
even
ing
![Page 25: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/25.jpg)
Conclusions
25
► The proposed method is scalable and is ready to be integrated in automated workflows
• Works with only data that is commonly available
• No human intervention needed
► Support vector machines (SVM) are powerful for pattern detection in problems with geometrical features
![Page 27: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/27.jpg)
27
![Page 28: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/28.jpg)
28
c-Si sites
![Page 29: Machine learning for supporting automated performance ... › Websections... · Smart Buildings Symposium –TU Delft –07/02/2020 Machine learning for supporting automated performance](https://reader030.fdocuments.us/reader030/viewer/2022011902/5f0be8347e708231d432ce28/html5/thumbnails/29.jpg)
29
CIGS sites