基于TCN的双向LSTM光伏功率概率预测  

Probabilistic prediction of photovoltaic power based on bidirectional LSTM with TCN

作　　者：盛万兴[1] 李蕊[1] 赵阳 李鹏丽 张倩[2] SHENG Wanxing;LI Rui;ZHAO Yang;LI Pengli;ZHANG Qian(China Electric Power Research Institute,Beijing 100192,China;School of Electrical Engineering and Automation,Anhui University,Hefei 23060l,China)

机构地区：[1]中国电力科学研究院,北京100192 [2]安徽大学电气工程与自动化学院,安徽合肥230601

出　　处：《安徽大学学报(自然科学版)》2025年第2期39-48,共10页Journal of Anhui University(Natural Science Edition)

基　　金：国家电网有限公司总部科技项目(5400-202355207A-1-1-ZN)。

摘　　要：为更好地描述光伏出力不确定性,该文提出了一种基于时序卷积网络(temporal convolutional network,简称TCN)和双向长短期记忆(bidirectional long short term memory,简称BiLSTM)的光伏功率概率预测模型.首先,基于数值天气预报中的云量和降雨量将历史数据集划分为晴天、多云天和阴雨天3种场景,生成具有相似天气类型的测试集和训练样本集:然后,应用TCN进行集成特征维度提取,利用BiLSTM神经网络建模进行输出功率和天气数据时间序列的双向拟合.针对传统区间预测分位数损失函数不可微的缺陷,引入Huber范数近似替代原损失函数,并应用梯度下降进行优化,构建改进的可微分位数回归(quantile regression,简称QR)模型,生成置信区间.最后,采用核密度估计(kerneldensity estimation,简称KDE)给出概率密度预测结果。以我国华东某地区分布式光伏电站作为研究对象,与现有概率预测方法相比,该文所提出的短期预测算法的功率区间各评价指标都有所改进,验证了所提方法的可靠性。To better describe the uncertainty of photovoltaic output,this paper proposes a photovoltaic power probability prediction model based on temporal convolutional network(TCN)and bidirectional long short term memory(BiLSTM).Firstly,historical data sets are divided into three weather scenarios:sunny,partly cloudy,and cloudy/rainy based on cloud cover and rainfall from numerical weather predictions.This segmentation generates training and testing datasets with similar weather conditions.Then,TCN is applied for integrated feature dimension extraction,and BiLSTM neural network is used for bidirectional fitting of output power and weather data time series.In response to the nondifferentiability issue of traditional interval prediction quantile loss functions,we introduce the Huber norm approximation as a substitute for the original loss function and apply gradient descent for optimization,forming an improved differentiable quantile regression(QR)model to generate confidence intervals.Finally,kernel density estimation is used to give probability density predictions.Taking a distributed photovoltaic power station in East China as the research object,compared with existing probability prediction methods,the proposed ultra-short-term estimation algorithm shows improvements in various evaluation metrics of power intervals,validating the reliability of the proposed approach.

关 键 词：光伏 概率预测 TCN 分位数回归 BiLSTM 

分 类 号：TM73[电气工程—电力系统及自动化]