基于CNN-QRLightGBM的短期负荷概率密度预测  

Short-term Load Density Prediction Based on CNN-QRLightGBM

作　　者：许佳辉 王向文 杨俊杰 XU Jiahui;WANG Xiangwen;YANG Junjie(College of Electrical Engineering,Shanghai University of Electric Power,Pudong New District,Shanghai 201306,China;Shanghai Dianji University,Pudong New District,Shanghai 201306,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海市浦东新区201306 [2]上海电机学院,上海市浦东新区201306

出　　处：《电网技术》2020年第9期3409-3415,共7页Power System Technology

基　　金：国家自然科学基金项目(61401269,61572311)。

摘　　要：为了应对智能电网时代用电需求的活跃性、不稳定性,提出了一种基于卷积神经网络(convolutional neural network,CNN)结合分位数回归轻量梯度提升机的短期负荷概率密度预测方法。首先通过建立Copula模型分析变量之间的相关性、采用CNN进行特征提取;其次通过树状结构Parzen估计方法对回归预测模型进行超参数调优;然后用调优后的模型在不同分位点上进行预测,将预测结果进行核密度估计得到概率密度函数。最后在美国新英格兰地区缅因州数据集上进行仿真验证,结果表明所提方法可更好地量化短期负荷概率密度预测。In order to cope with the active and unstable demand of electric power in the era of smart grid,a new method of short-term public building load probability density prediction is proposed based on convolutional neural network and combined with quantile regression Light Gradient Boosting Machine(QRLightGBM).Firstly,the Copula model is used to analyze the correlation between variables and the feature extraction is performed by the convolutional neural network.Then the parameters of the regression prediction model are tuned by the Tree-structured Parzen Estimator(TPE).Finally,the tuned model is used to predict the different quantile points,and the prediction results are estimated by kernel density to obtain the probability density function.The simulation results on the Maine dataset in New England,USA show that the proposed method can better quantify the load probability density prediction.

关 键 词：COPULA函数 卷积神经网络 分位数回归 LightGBM 核密度估计 TPE 

分 类 号：TM715[电气工程—电力系统及自动化]