摘要
随着光伏发电在电力系统中的日益普及,表征其固有的不确定性变得越来越重要。针对现有光伏场景生成方法过分依赖统计假设和模型训练不稳定等问题,提出了一种含有梯度惩罚的改进生成对抗网络光伏功率场景生成方法。该方法首先以Wasserstein距离作为损失函数设计生成器和判别器两个深度神经网络进行对抗训练,其次在损失函数中引入一种梯度惩罚策略增强模型的Lipschitz连续性约束并创新性地应用于光伏功率的场景生成,提高了场景生成模型的收敛速度和生成场景的质量。所提方法能够捕获光伏出力的非线性,且无需建模假设和复杂的采样技术。算例分析表明,该方法能够精准捕捉光伏功率的分布特性,具有很强的泛化能力,并且优于其他先进的场景生成方法。
With the increasing popularity of PV power generation in power systems,the characterization of its inherent uncertainties becomes increasingly important.Aimed at the problems that the existing PV scenario generation methods rely too much on statistical assumptions and the model training is unstable,a PV power scenario generation method based on an improved generative adversarial network with gradient penalty is proposed.First,two deep neural networks(i.e.,a generator and a discriminator)are designed with the Wasserstein distance as a loss function for adversarial training.Second,a gradient penalty strategy is introduced into the loss function to enhance the Lipschitz continuity constraint of the model,which is innovatively applied to PV power scenario generation,thus improving the convergence speed of the scenario generation model and the quality of generated scenarios.The proposed method can capture the nonlinearity of PV output without modeling assumptions or complex sampling techniques.The results of an example demonstrate that the proposed method can accurately capture the distribution characteristics of PV power while maintaining a strong generalization capability.In addition,it is superior to other state-of-the-art scenario generation methods.
作者
胡石峰
朱瑞金
唐波
HU Shifeng;ZHU Ruijin;TANG Bo(School of Electrical Engineering,Tibet Agricultural and Animal Husbandry University,Linzhi 860000,China)
出处
《电力系统及其自动化学报》
CSCD
北大核心
2022年第11期109-115,共7页
Proceedings of the CSU-EPSA
基金
西藏自治区自然科学基金资助项目(XZ202001ZR0093G)。
关键词
生成对抗网络
场景生成
深度学习
光伏功率
梯度惩罚
generative adversarial networks(GANs)
scenario generation
deep learning
PV power
gradient penalty
