基于改进WGAN考虑特征分布相似性的小样本负荷预测方法

Small sample load forecasting method considering characteristic distribution similarity based on improved WGAN

中文关键词  负荷预测  迁移学习  小样本  改进wasserstein生成对抗网络  特征分布  最优传输

英文关键词  load forecast  transfer learning  small sample  improved wasserstein generative adversarial nets  characteristic distribution  optimal transport

中文摘要

      对于综合能源系统中新接入用户, 其往往由于历史数据匮乏而难以建立精准的短期负荷预测模型. 本文基于迁移学习理论, 提出了一种基于改进wasserstein生成对抗网络(WGAN)的小样本负荷预测方法. 首先, 本文采用最大信息系数法量化各影响特征与负荷的相关性强弱. 接着, 将源域特征序列进行分割, 计算各分割子序列与目标域小样本的实序列编辑距离确定初始源域. 然后, 引入卷积神经网络和长短期记忆模型建立源域预测网络. 通过WGAN对齐目标域和源域负荷特征的空间分布, 并在最优传输代价函数中加入局部特征损失以提高训练的稳定性和快速性. 最后, 将对抗训练后网络用于目标域负荷预测. 采用该方法对某地区小样本负荷进行实验, 结果表明, 本文所提算法与其他预测模型相比能达到更高精度

英文摘要

      For a new user of an integrated energy sysytem, it is much more difficult to develop an accurate load forecasting model due to the lack of historical data. A small sample load forecasting method based on the improved wasserstein generative adversarial nets(WANG) is proposed based on the transfer learning theory. First, the maximal information coefficient method is used to quantify the correlation among the impact characteristics and the load. Next, the source domain characteristic sequence is segmented and the edit distance on real sequence between each segmented sequence and the small sample in the target domain is calculated to determine the initial source domain. Then, the convolution neural network and long short-term memory model are introduced to establish the source domain prediction network. The spatial distribution of load characteristics both in target domain and source domain is aligned by WANG, and the local characteristic loss is added to the optimal transport cost function to improve the stability and rapidity of training process. Finally, the network after adversarial training is used for the target domain load forecasting. The method proposed is used to test a small sample in a certain area and the result shows that the algorithm proposed in this paper turns out to be more accurate compared with other prediction models.