摘要
电力时间序列数据不足会影响电力系统分析和预测的准确性,而现有数据增强算法常忽视时间序列的时间动力学特性,导致生成数据质量欠佳。为此,提出一种融合自回归模型和生成对抗网络的电力时间序列数据生成方法。采用该方法构建一维时间卷积嵌入层,降维获取潜在特征空间,结合带注意力机制的解码器重建层以实现特征映射;在该框架下建立生成器和鉴别器,设计重建损失、有监督及无监督损失函数。试验表明,该方法在电力数据增强中的稳定性和泛化能力更优,能有效提升生成数据质量。
Insufficient electric power time series data is recognized as a factor that adversely affects the accuracy of power system analysis and forecasting,while existing data augmentation algorithms often overlook the temporal dynamics of time series,resulting in poor quality of generated data.To address this issue,a hybrid method integrating autoregressive models and generative adversarial networks is proposed for electric power time series data generation.The method is employed to construct a one-dimensional temporal convolutional embedding layer,through which dimensionality reduction is performed to obtain a latent feature space,and an attention-equipped decoder reconstruction layer is incorporated to achieve feature mapping.Within this framework,a generator and a discriminator are established,and reconstruction loss,supervised,and unsupervised loss functions are designed.Experiments demonstrate that the stability and generalization capability of the proposed method in power data augmentation are superior,and the quality of the generated data is effectively improved.
作者
张慧虹
陈键
邬凌飞
ZHANG Huihong;CHEN Jian;WU Lingfei(Shanghai Marine Equipment Research Institute,Shanghai 200031,China)
出处
《机电设备》
2025年第5期115-118,共4页
Mechanical and Electrical Equipment
关键词
自回归模型
卷积神经网络
电力数据
数据增强
autoregressive model
convolutional neural network
electric power data
data augmentation
