Scholarlink

<p>To address the nonlinearity and non-stationarity inherent in power load series, this paper proposes an Adaptive Distribution-Shift- and Frequency-Aware Generative Adversarial Network (DFGAN) for power load forecasting. These characteristics often manifest as temporal statistical distribution shifts and complex frequency variations, which lead to reduced forecasting accuracy. To mitigate these issues, an adaptive distribution shift module based on Reversible Instance Normalization (RevIN) is integrated into the DFGAN generator, enabling learnable normalization and denormalization that mitigate statistical distribution inconsistencies. In addition, a frequency-aware mechanism implemented via Frequency-domain MLPs for Time Series (FreTS) is incorporated into the DFGAN generator to explicitly capture informative frequency components of power load series. Extensive experiments conducted on a real-world Suzhou power load dataset demonstrate the effectiveness of the proposed DFGAN. The results show that DFGAN achieves a mean absolute error (MAE) of 14.8332, a root mean squared error (RMSE) of 20.6534, and an R² of 92.25%, outperforming state-of-the-art benchmark models.</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 GAN-based power load forecasting method with adaptive distribution shift and frequency-aware mechanisms

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月5日

查看全文
  • Jie Sun,
  • Yong Zhang,
  • Haolei Hu,
  • Feng Xu

摘要

To address the nonlinearity and non-stationarity inherent in power load series, this paper proposes an Adaptive Distribution-Shift- and Frequency-Aware Generative Adversarial Network (DFGAN) for power load forecasting. These characteristics often manifest as temporal statistical distribution shifts and complex frequency variations, which lead to reduced forecasting accuracy. To mitigate these issues, an adaptive distribution shift module based on Reversible Instance Normalization (RevIN) is integrated into the DFGAN generator, enabling learnable normalization and denormalization that mitigate statistical distribution inconsistencies. In addition, a frequency-aware mechanism implemented via Frequency-domain MLPs for Time Series (FreTS) is incorporated into the DFGAN generator to explicitly capture informative frequency components of power load series. Extensive experiments conducted on a real-world Suzhou power load dataset demonstrate the effectiveness of the proposed DFGAN. The results show that DFGAN achieves a mean absolute error (MAE) of 14.8332, a root mean squared error (RMSE) of 20.6534, and an R² of 92.25%, outperforming state-of-the-art benchmark models.