基于动态指数时间平滑随机梯度下降的短期风电功率在线概率预测方法  

Online short term wind power probabilistic forecasting approach based on dynamic exponential time smoothed stochastic gradient descent

作　　者：乔如妤 于凯 李泽民 刘雁行 徐恺 李旭东 梁楠 冯军 王耀健 温洪林 顾洁[2] QIAO Ruyu;YU Kai;LI Zemin;LIU Yanxing;XU Kai;LI Xudong;LIANG Nan;FENG Jun;WANG Yaojian;WEN Honglin;GU Jie(Inner Mongolia Power(Group),Power Marketing and Operation Management Division,Hohhot 010090,China;Shanghai Jiao Tong University,Department of Electrical Engineering,Shanghai 200240,China)

机构地区：[1]内蒙古电力(集团)有限责任公司电力营销与运营管理分公司,内蒙古自治区呼和浩特010090 [2]上海交通大学电气工程系,上海200240

出　　处：《供用电》2025年第2期100-107,共8页Distribution & Utilization

基　　金：国家自然科学基金项目(52307119);内蒙古电力(集团)有限责任公司2023年第一批科技项目(2023-5-46)。

摘　　要：为解决风电出力的非平稳性所引起的预测模型参数时变问题,降低短期风电功率预测误差,提出了一种风电功率在线概率预测方法。通过构建分位数神经网络模型及动态更新参数,实现了预测准确性的提升;采用凸神经网络模型,实现了数值天气预报和风电功率间的非线性映射;引入动态局部遗憾度,将凸神经网络模型拓展为在线学习模型,并采用动态指数时间平滑随机梯度下降(dynamic exponentially time-smoothed stochastic gradient descent,DTS-SGD)算法实现参数更新。最后,结合2014年全球能源预测大赛发布的风力发电数据,验证了所提方法的预测准确性以及对超参数选择的鲁棒性,并验证了所提方法可以保证在较高计算效率的同时,能够自适应地应对参数时变问题。To address the issue of time-varying parameters arising from the non-stationarity of wind power,which consequently mitigates short-term wind prediction errors,we present an online probabilistic wind power forecasting approach.The approach employs an input convex neural network and incorporates parameter updates to enhance forecast quality.Specifically,it utilizes an input convex neural network to establish a nonlinear mapping between numerical weather predictions and wind power.The introduction of dynamic local regret transforms the input convex neural network model into an online model.Parameters are updated using dynamic exponential time-smoothed stochastic gradient descent.Finally,the proposed approach is evaluated using wind power generation data from the Global Energy Prediction Competition 2014,assessing prediction accuracy,robustness to hyper parameter selection,and adaptive resolution of the time-varying parameters issue with improved computational efficiency.

关 键 词：短期风电功率预测 概率预测 在线学习 非平稳性 机器学习 

分 类 号：TM614[电气工程—电力系统及自动化]