Scholarlink

<p>The lack of historical data makes it difficult to accurately forecast photovoltaic (PV) generation for newly established power stations. In response to this issue, the current literature extensively examines transfer learning (TL) and data augmentation strategies to train prediction models. However, two challenges remain: 1) While TL aims to enrich forecasting in a target domain (i.e., a newly established power station) by training models with knowledge learned from a similar source domain (previously established power station), it can overlook the rich data available from other relevant sources; 2) despite achieving efficient learning from the source domain, the domain adaptation of TL may still encounter challenges owing to the persistent issue of data scarcity in the target domain, which necessitates the development of an effective data augmentation strategy. To overcome these challenges, this study introduces a novel methodology that integrates learned knowledge from multiple source domains by leveraging two fusion methods: average weights and an advanced evolutionary optimization-based approach. It also adapts a generative adversarial network (TimeGAN) to enhance the robustness of the proposed TL approach by supplementing it with sufficient synthetic time-series data for domain adaptation. In particular, the study evaluates different data augmentation strategies, single-source-based and multi-source-based augmentation, to enhance forecasting accuracy. The proposed methodology improves the PV generation forecasting performance by 8%–51% (in terms of MAE), compared to the baseline model trained solely on limited target domain data. These improvements were observed across multiple real-world datasets under varying data scarcity conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generative adversarial networks and transfer learning for renewable energy forecasting: a multi-source knowledge fusion approach

  • 期刊论文

  • 发表日期：

    2025年9月21日

查看全文
  • Sarah Almaghrabi,
  • Mohammad Saiedur Rahaman,
  • Mashud Rana,
  • Margaret Hamilton

摘要

The lack of historical data makes it difficult to accurately forecast photovoltaic (PV) generation for newly established power stations. In response to this issue, the current literature extensively examines transfer learning (TL) and data augmentation strategies to train prediction models. However, two challenges remain: 1) While TL aims to enrich forecasting in a target domain (i.e., a newly established power station) by training models with knowledge learned from a similar source domain (previously established power station), it can overlook the rich data available from other relevant sources; 2) despite achieving efficient learning from the source domain, the domain adaptation of TL may still encounter challenges owing to the persistent issue of data scarcity in the target domain, which necessitates the development of an effective data augmentation strategy. To overcome these challenges, this study introduces a novel methodology that integrates learned knowledge from multiple source domains by leveraging two fusion methods: average weights and an advanced evolutionary optimization-based approach. It also adapts a generative adversarial network (TimeGAN) to enhance the robustness of the proposed TL approach by supplementing it with sufficient synthetic time-series data for domain adaptation. In particular, the study evaluates different data augmentation strategies, single-source-based and multi-source-based augmentation, to enhance forecasting accuracy. The proposed methodology improves the PV generation forecasting performance by 8%–51% (in terms of MAE), compared to the baseline model trained solely on limited target domain data. These improvements were observed across multiple real-world datasets under varying data scarcity conditions.