Scholarlink

<p>Short-term photovoltaic (PV) power forecasting is challenged under cloudy conditions. Recent deep learning-based methods utilize time-frequency modeling offer promise, yet they ignore either spectral bias or redundancy, limiting prediction accuracy. Therefore, we propose a complementary time-frequency domain approach for this task, which integrates PV and meteorological data into two learners in both domains. Moreover, both learners are guided by intra- and inter-domain constraints to enhance representation quality, while a specialized attention module fuses features for final prediction. Experiments on two real-world datasets, conducted with GPU-based parallelization, demonstrate that CTFD outperforms fourteen state-of-the-art baselines, reducing MAE by 12.50% and 12.75%, and RMSE by 11.38% and 12.99% on cloudy test sets. These results not only validate the effectiveness of our complementary time-frequency modeling, but also imply that accurate forecasting of PV power fluctuations inherently demands parallel or high-performance computing capabilities, making such computational cost a necessary trade-off in practical deployments.</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

A complementary time-frequency domain approach for short-term photovoltaic power forecasting

  • 期刊论文

  • 发表日期：

    2025年7月17日

查看全文
  • Wentao Wang,
  • Haiyan Lu,
  • Ayesha Ubaid,
  • Jianzhou Wang

摘要

Short-term photovoltaic (PV) power forecasting is challenged under cloudy conditions. Recent deep learning-based methods utilize time-frequency modeling offer promise, yet they ignore either spectral bias or redundancy, limiting prediction accuracy. Therefore, we propose a complementary time-frequency domain approach for this task, which integrates PV and meteorological data into two learners in both domains. Moreover, both learners are guided by intra- and inter-domain constraints to enhance representation quality, while a specialized attention module fuses features for final prediction. Experiments on two real-world datasets, conducted with GPU-based parallelization, demonstrate that CTFD outperforms fourteen state-of-the-art baselines, reducing MAE by 12.50% and 12.75%, and RMSE by 11.38% and 12.99% on cloudy test sets. These results not only validate the effectiveness of our complementary time-frequency modeling, but also imply that accurate forecasting of PV power fluctuations inherently demands parallel or high-performance computing capabilities, making such computational cost a necessary trade-off in practical deployments.