摘要
针对光伏发电可预测性低的问题,提出了一种综合使用通径分析(path analysis,PA)、k近邻算法(k-Nearest Neighbor,KNN)、神经网络分位数回归(quantile regression neural network,QRNN)和核密度估计(kernel density estimator,KDE)的光伏出力概率分布估计方法,构造出未来1 d任意时刻的光伏出力概率密度函数,可以得到比点预测和区间预测更多的有用信息。首先由通径分析对气象因素进行约减,在降低模型输入维数的基础上减小变量间的耦合作用。然后通过K-means算法按天气类型对历史样本进行聚类,进一步提高相似样本的筛选效果。最后利用神经网络分位数回归和核密度估计对光伏出力的概率分布进行估计。实验结果表明,相比于核密度估计和传统的正态分布估计方法,采用所提方法估计出的概率分布的可靠性和锐度更高。
Considering poor predictability of PV power generation, this paper proposes a method based on path analysis （PA）, k-Nearest Neighbor （KNN）, quantile regression neural network （QRNN） and kernel density estimator （KDE） to estimate probabilistic distribution of PV generation at any moment in a day. Probability density function of PV generation provides more useful information than point and interval predictions. Firstly, path analysis is employed to reduce dimensions of meteorological factors and coupling influence between variables. Then K-means method is used to divide historical data according to weather types and improve effect of similar sample selection. Finally, quantile regression neural network and kernel density estimator are used to estimate probabilistic distribution of PV generation. Results show that the proposed method improves both reliability and sharpness of PV generation probabilistic forecast compared with probabilistic forecast methods based on normal distribution and kernel density estimator.
出处
《电网技术》
EI
CSCD
北大核心
2017年第2期448-454,共7页
Power System Technology
基金
国家自然科学基金项目(61374122)~~
关键词
神经网络分位数回归
通径分析
核密度估计
光伏发电
概率分布
quantile regression neural network
path analysis
kemel density estimator
solar power generation
probabilistic distribution
