引用本文: | 李彬,彭曙蓉,彭君哲,黄士峻,郑国栋.基于深度学习分位数回归模型的风电功率概率密度预测[J].电力自动化设备,2018,(9): |
| LI Bin,PENG Shurong,PENG Junzhe,HUANG Shijun,ZHENG Guodong.Wind power probability density forecasting based on deep learning quantile regression model[J].Electric Power Automation Equipment,2018,(9): |
|
摘要: |
针对风电功率预测问题,在现有预测方法和概率性区间预测的基础上,提出基于深度学习分位数回归的风电功率概率预测方法。该方法采用Adam随机梯度下降法在不同分位数条件下对长短期记忆神经网络(LSTM)的输入、遗忘、记忆、输出参数进行估计,得出未来200 h内各个时刻风电功率的概率密度函数。根据美国PJM网上的风电功率实际数据的仿真结果表明,所提方法不仅能得出较为精确的点预测结果,而且能够获得风电功率完整的概率密度函数预测结果。与神经网络分位数回归相比,其精度更高,且在同等置信度下的预测区间范围更小。 |
关键词: LSTM 核密度估计 风电功率概率预测 LSTM分位数回归 概率密度分布 |
DOI:10.16081/j.issn.1006-6047.2018.09.003 |
分类号:TM614 |
基金项目:湖南省教育厅创新平台开放基金资助项目(17K001) |
|
Wind power probability density forecasting based on deep learning quantile regression model |
LI Bin, PENG Shurong, PENG Junzhe, HUANG Shijun, ZHENG Guodong
|
School of Electrical & Information Engineering, Changsha University of Science & Technology, Changsha 410114, China
|
Abstract: |
Aiming at the problem of wind power forecasting, a deep learning quantile regression based method is proposed based on the existing forecasting methods and probabilistic interval forecasting. The method adopts Adam random gradient descent method to estimate the input, forgetting, memory and output parameters of LSTM(Long Short Term Memory neural networks) under different quantiles. The probability density function of wind power at each moment within 200 h in the future is obtained. According to the actual wind power data from PJM network in the United States, the simulative results show that the proposed method can obtain accurate point forecasting results and complete probability density function forecasting results of wind power, compared with the quantile regression of neural network, it has higher accuracy and smaller range of forecasting interval under the same confidence. |
Key words: LSTM nuclear density estimation wind power probability prediction LSTM quantile regression probability density distribution |