國立虎尾科技大學 |

Short-Term Solar Forecasting from All-Sky Images Using Deep Learning.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Short-Term Solar Forecasting from All-Sky Images Using Deep Learning./
Author:	Nie, Yuhao.
Description:	1 online resource (308 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-06, Section: B.
Contained By:	Dissertations Abstracts International85-06B.
Subject:	Data processing. -
Online resource:	click for full text (PQDT)
ISBN:	9798381019902

Short-Term Solar Forecasting from All-Sky Images Using Deep Learning.
Nie, Yuhao.

Short-Term Solar Forecasting from All-Sky Images Using Deep Learning. - 1 online resource (308 pages)

Source: Dissertations Abstracts International, Volume: 85-06, Section: B.

Thesis (Ph.D.)--Stanford University, 2023.

Includes bibliographical references

Integration of renewable resources, such as solar photovoltaics (PV), has been recognized as a crucial component in transition to a decarbonized energy system. However, the intermittent nature of solar power has challenged the large-scale deployment of PV. This variability is partly caused by short-term and local cloud events. Ground-based all-sky images captured with high temporal and spatial resolution have shown great promise as a source of input to forecast such fluctuations. In recent years, the development of deep learning has provided powerful tools to extract information from sky images and enhanced short-term solar forecasting capabilities. Despite these advancements, several major challenges have been identified: (1) sky image data are often imbalanced due to the tendency of installing PV systems in sunny locations; (2) cloud dynamics are not well captured by end-to-end deep solar forecasting models and uncertainty of predictions are rarely quantified; and (3) high-quality standardized sky image datasets for solar forecasting method development and benchmark are limited. In this dissertation, we explore ways to address these challenges by focusing on predicting the power output of a roof-top 30 kW PV system up to 15-minute ahead using a locally collected sky image dataset.The first part of the dissertation focuses on how to make effective use of imbalanced sky image data for PV output prediction. We explore methods from two different perspectives, namely, modeling framework improvement and data re-balancing. In Chapter 3, we develops a classification-prediction framework to ameliorate PV output prediction under cloudy conditions. The proposed framework first classifies input images into different sky conditions and then the classified images are sent to tailored sky-condition-specific sub-models for PV output prediction. Under the best design, the proposed framework improves cloudy prediction performance by 6% while requires 6% fewer trainable parameters compared to the end-to-end convolutional neural network (CNN) baseline. Chapter 4 examines the efficacy of using resampling and data augmentation methods to enrich the minor cloudy data for model development. We employ three-stage greedy search to determine the optimal resampling approach, data augmentation techniques and over-sampling rate. We find that for the nowcast problem, resampling and data augmentation can effectively enhance the model performance, while for the forecast problem it nearly overlaps the baseline performance. The optimal resampling approach expands on the original dataset by over-sampling the minor cloudy data, with the best results from 4x ∼ 6x over-sampling rate.The second part of the dissertation explores how to enhance cloud dynamics modeling and quantification of prediction uncertainty. In Chapter 5, we leverage the recent advances in deep generative models to synthesize visually plausible yet diversified sky videos for probabilistic solar forecasting. We introduce a physics-constrained stochastic video prediction model named SkyGPT. By using past sky image sequences as input, SkyGPT is able to generate multiple possible future sky images with diverse cloud motion patterns. Extensive experiments demonstrate its effectiveness in capturing cloud dynamics and generating future sky images of high realism and diversity. We feed the collection of generated future sky images for 15-minute-ahead probabilistic solar forecasting and observe better PV output prediction reliability and sharpness by using the predicted sky images from SkyGPTcompared with other benchmark models.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798381019902Subjects--Topical Terms:

840972
Data processing.
Index Terms--Genre/Form:

554714
Electronic books.

Short-Term Solar Forecasting from All-Sky Images Using Deep Learning.
LDR:04819ntm a22003257 4500 001 1148748
005 20240930100135.5
006 m o d
007 cr bn ---uuuuu
008 250605s2023 xx obm 000 0 eng d
020 $a 9798381019902
035 $a (MiAaPQ)AAI30726863
035 $a (MiAaPQ)STANFORDbm790hj4850
035 $a AAI30726863
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Nie, Yuhao. $3 1474792
245 1 0 $a Short-Term Solar Forecasting from All-Sky Images Using Deep Learning.
264 0 $c 2023
300 $a 1 online resource (308 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-06, Section: B.
500 $a Advisor: Brandt, Adam.
502 $a Thesis (Ph.D.)--Stanford University, 2023.
504 $a Includes bibliographical references
520 $a Integration of renewable resources, such as solar photovoltaics (PV), has been recognized as a crucial component in transition to a decarbonized energy system. However, the intermittent nature of solar power has challenged the large-scale deployment of PV. This variability is partly caused by short-term and local cloud events. Ground-based all-sky images captured with high temporal and spatial resolution have shown great promise as a source of input to forecast such fluctuations. In recent years, the development of deep learning has provided powerful tools to extract information from sky images and enhanced short-term solar forecasting capabilities. Despite these advancements, several major challenges have been identified: (1) sky image data are often imbalanced due to the tendency of installing PV systems in sunny locations; (2) cloud dynamics are not well captured by end-to-end deep solar forecasting models and uncertainty of predictions are rarely quantified; and (3) high-quality standardized sky image datasets for solar forecasting method development and benchmark are limited. In this dissertation, we explore ways to address these challenges by focusing on predicting the power output of a roof-top 30 kW PV system up to 15-minute ahead using a locally collected sky image dataset.The first part of the dissertation focuses on how to make effective use of imbalanced sky image data for PV output prediction. We explore methods from two different perspectives, namely, modeling framework improvement and data re-balancing. In Chapter 3, we develops a classification-prediction framework to ameliorate PV output prediction under cloudy conditions. The proposed framework first classifies input images into different sky conditions and then the classified images are sent to tailored sky-condition-specific sub-models for PV output prediction. Under the best design, the proposed framework improves cloudy prediction performance by 6% while requires 6% fewer trainable parameters compared to the end-to-end convolutional neural network (CNN) baseline. Chapter 4 examines the efficacy of using resampling and data augmentation methods to enrich the minor cloudy data for model development. We employ three-stage greedy search to determine the optimal resampling approach, data augmentation techniques and over-sampling rate. We find that for the nowcast problem, resampling and data augmentation can effectively enhance the model performance, while for the forecast problem it nearly overlaps the baseline performance. The optimal resampling approach expands on the original dataset by over-sampling the minor cloudy data, with the best results from 4x ∼ 6x over-sampling rate.The second part of the dissertation explores how to enhance cloud dynamics modeling and quantification of prediction uncertainty. In Chapter 5, we leverage the recent advances in deep generative models to synthesize visually plausible yet diversified sky videos for probabilistic solar forecasting. We introduce a physics-constrained stochastic video prediction model named SkyGPT. By using past sky image sequences as input, SkyGPT is able to generate multiple possible future sky images with diverse cloud motion patterns. Extensive experiments demonstrate its effectiveness in capturing cloud dynamics and generating future sky images of high realism and diversity. We feed the collection of generated future sky images for 15-minute-ahead probabilistic solar forecasting and observe better PV output prediction reliability and sharpness by using the predicted sky images from SkyGPTcompared with other benchmark models.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Data processing. $2 sears $3 840972
650 4 $a Deep learning. $3 1127428
650 4 $a Forecasting. $3 632960
655 7 $a Electronic books. $2 local $3 554714
690 $a 0800
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a Stanford University. $3 1184533
773 0 $t Dissertations Abstracts International $g 85-06B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30726863 $z click for full text (PQDT)

	Simple Search
	Advanced Search
	Expert Search
	Public Library Lists
	Public Reader Lists

	AcademicReservedBook [CH]
	BookLoanBillboard [CH]
	BookReservedBillboard [CH]
	SubmaintainWeb [CH]
	Classification Browse [CH]
	NewBook [CH]

	Personal Details
	Bookmarks
	Recommendations
	Penalties
	Messages