考虑训练样本分布不均衡的超短期风电功率概率预测  

Probability Prediction of Ultra-short-term Wind Power Considering Unbalanced Distribution of Training Samples

作　　者：李丹[1] 方泽仁 缪书唯 胡越 梁云嫣 贺帅 LI Dan;FANG Zeren;MIAO Shuwei;HU Yue;LIANG Yunyan;HE Shuai(Electric and New Energy Faculty,China Three Gorges University,Yichang 443002,Hubei Province,China;Hubei Provincial Collaborative Innovation Center for New Energy Microgrid,Yichang 443002,Hubei Province,China;Hubei Provincial Key Laboratory for Operation and Control of Cascaded Hydropower Station,Yichang 443002,Hubei Province,China)

机构地区：[1]三峡大学电气与新能源学院,湖北省宜昌市443002 [2]梯级水电站运行与控制湖北省重点实验室,湖北省宜昌市443002 [3]新能源微电网湖北省协同创新中心,湖北省宜昌市443002

出　　处：《电网技术》2024年第3期1133-1145,共13页Power System Technology

基　　金：国家自然科学基金项目(51807109)。

摘　　要：提出一种考虑训练样本分布不均衡的超短期风电概率预测方法。首先构建深度信念混合密度网络,通过深度信念网络独特的预训练和微调机制提取输入变量的隐特征,利用Beta混合概率分布的有界性准确表征风电预测功率的概率分布,实现隐特征与预测功率概率分布参数之间的非线性映射;然后引入训练样本分布平滑策略,其中特征分布平滑技术用于校准输入特征,标签分布平滑技术用于对各样本误差赋予差异化权重,从输入和输出两方面改善训练样本分布不均衡现象对预测结果的不利影响。实际算例结果表明,与常见风电功率概率预测模型相比,所提模型在点预测和概率预测方面均能获得较高的预测精度,尤其能有效提高低密度样本区域的预测精度。A probability prediction method of ultra-shortterm wind power is proposed considering the unbalanced distribution of training samples.Firstly,a deep belief mixture density network is built.The hidden features of input variables are extracted through the deep belief network’s unique pre-training and fine-tuning mechanism.A bounded Betamixed PDF is used to accurately characterize the probability distribution of forecasting wind power and realize the nonlinear mapping from hidden features to PDF parameters.Then,a distribution smoothing strategy of training samples is introduced to mitigate the negative consequences of an unbalanced distribution of training samples from both input and output distribution aspects.The feature distribution smoothing technology is used to correct the input features,and the label distribution smoothing technology is employed to assign different weights to sample errors.The actual case results show that compared with the popular probability prediction models,the proposed model can achieve higher accuracy in both point and probability prediction,especially in low-density sample areas.

关 键 词：风电功率概率预测 深度信念网络 混合密度网络 训练样本分布不均衡 特征分布平滑 标签分布平滑 

分 类 号：TM914[电气工程—电力电子与电力传动]