摘要
为减少光伏发电的不确定性对电力系统造成的影响,描绘更为准确、清晰的光伏出力区间,提出一种基于天气相似聚类与分位数回归神经网络(QRNN)单调模型的短期光伏功率区间概率预测模型。首先,QRNN单调模型在预测过程中引入单调性,保证单调的分位数结果,并采用Huber范数近似替代原损失函数,弥补了传统区间预测分位数交叉及损失函数不可微的缺陷。其次,基于气象信息的数据特征,采用动态自组织映射聚类算法,通过神经元竞争与更新确定神经元邻域权重结构,并根据其邻域权重大小将天气聚类为晴天、多云天与阴天,得到相似天气下的数据集。然后,在不同天气条件下对影响光伏出力的因素进行相关性分析,得到不同天气下的天气影响特征,并作为输入因子输入神经网络中。最后,以澳大利亚沙漠知识太阳能中心实际数据集为例进行区间预测,并采用核密度估计给出概率预测结果,验证了所提方法的可靠性。
In order to reduce the influence of the uncertainty of photovoltaic power generation on the power system and describe the photovoltaic power output interval more accurately and clearly,a short-term interval probability prediction model of photovoltaic power based on the weather similarity clustering and quantile regression neural network(QRNN)monotone model is proposed.Firstly,the monotonicity of QRNN monotone model is introduced in the prediction process to ensure the monotonicity of quantile results,and the Huber norm is used to replace the original loss function,which makes up the defects of the traditional interval prediction of quantile crossing and the non-differentiable loss function.Secondly,based on the data characteristics of meteorological information,the dynamic self-organizing mapping clustering algorithm is used to determine the neuronal neighborhood weight structure through neuronal competition and updating.According to the neighborhood weight,the weather is classified into sunny,cloudy and overcast days,and the data sets under similar weather are obtained.Then,the correlation analysis of the factors affecting the photovoltaic output under different weather conditions is carried out,and the weather influence characteristics under different weather are obtained and input into the neural network as input factors.Finally,the actual data set of Australian Desert Knowledge Solar Energy Center is taken as a case to make interval prediction,and the probability prediction results are given by kernel density estimation,which verifies the reliability of the proposed method.
作者
赵耀
高少炜
李东东
林顺富
杨帆
黄学勤
ZHAO Yao;GAO Shaowei;LI Dongdong;LIN Shunfu;YANG Fan;HUANG Xueqin(School of Electrical Engineering,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《电力系统自动化》
EI
CSCD
北大核心
2023年第23期152-161,共10页
Automation of Electric Power Systems
基金
上海市启明星计划项目(21QC1400200)
上海市自然科学基金资助项目(21ZR1425400)
国家自然科学基金资助项目(52377111)。
关键词
区间预测
分位数回归
神经网络
核密度估计
光伏功率预测
interval prediction
quantile regression
neural network
kernel density estimation
photovoltaic power prediction
