Wind Power Output Scenario Generation Based on Long-Short Term Sequence Generative Adversarial Networks
摘要
电力系统需要预警风电出力短期波动性并制定长期调度计划,针对现有风电场景生成模型未能考虑长短期的时间多尺度特征提取问题,提出一种基于长短期序列生成对抗网络(long-short term sequence generation adversarial networks,LST-GAN)的风电出力场景生成混合架构.采用条件深度卷积生成对抗网络作为基础框架,嵌入MambaOUT架构动态调整短期波动特征权重,精细化生成短期出力特性;利用自注意力机制的长期特征提取能力,捕捉场景生成全局变化规律,使得模型能够预测风电出力的短期大幅波动和长期变化趋势,为电力系统制定长短期调度计划提供更准确的参考依据.对比实验表明,在长短期序列风电场景生成任务中,LST-GAN生成场景样本的覆盖率较高且功率区间平均宽度较小,捕捉长短时间序列特征的能力优于其他方法.
Power systems must hedge against short-term wind-power fluctuations while formulating long-term dispatch plans.Existing scenario-generation models,however,rarely extract temporal features on both long and short scales simultaneously.To address this limitation,we propose a hybrid architecture for wind power output scenario generation based on Long-Short Term sequence Generative Adversarial Networks(LST-GAN).Built on a conditional deep convolutional generative adversarial network,the generator embeds an MambaOUT block that dynamically reweights short-term volatility and generates instantaneous output characteristics;a self-attention mechanism,capable of long-term feature extraction,concurrently captures global,long-horizon patterns.The joint architecture allows the model to forecast both short-term,large-amplitude fluctuations and long-term evolution trends in wind power output,thereby providing a more accurate reference basis for formulating long-and short-term dispatch plans in power systems.Experimental results show that,for long-and short-term wind power scenario generation,LST-GAN yields scenario samples with higher coverage and narrower average power-interval widths,and its ability to capture long-and short-term sequence features surpasses that of alternative methods.