摘要
目前,准确的风速预测是解决大规模风电安全高效并网的关键基础问题之一。本文将深度学习理论引入风速预报的回归问题研究,基于堆栈降噪自动编码器建立了多个具有不同隐含层数的混合型深信度网络回归模型,并利用风电场的实测风速数据进行四组不同季节风速日前预测实验,包括相同隐含层模型在不同训练测试上的对比实验和不同隐含层模型在同一测试集上的对比实验。实验结果表明:相同隐含层模型在4个测试集上的回归误差(MSE和MAE)随着预报步长的变化都相对比较平稳、波动性不大,即深信度网络模型的鲁棒性较好对数据集的敏感性不强;预报误差随着隐含层数变化具有一定规律,即深信度网络模型对于同一测试集上存在一个最优的隐含层数能够使预报误差最小。最后,针对相同的训练测试集,采用典型常用的支持向量回归机进行风速日前日前对比实验,预报误差统计效果均比混合型深信度网络的差。因此,深信度网络可以通过其强大的非线性映射能力自动提取风速的复杂变化模式,从而有效提高回归模型的预测精度和鲁棒性。这为深度特征学习方法在风速预测中的实际应用奠定了一定的基础。
Currently, the accurate wind speed prediction is one of the fundamental problems to meet the safe and efficient interconnection of large scale wind power. This paper introduces the deep learning theo- ry to wind speed forecasting regression problems, and build the mixed convinced regression model with different hidden layers based on stack noise reduction automatic encoder. In addition, we have done four wind speed prediction experiments in four different seasons with the actual wind speed data, including the contrast experiment that the same underlying layer model on different training tests and different hiddenlayers model on the same test set. Experimental results show that the regression error （ MAE and MSE） of the model with the same underlying layer regression error is relatively stable and in low volatility as the prediction step change. This indicates that convinced network model has good robustness and the sensitiv- ity of the data set is not strong. Meantime, the prediction error has a certain regularity of the change of hidden layer, so there is a best number to make the prediction error least. Finally, a typical common support vector regression machine model is also used for wind speed prediction as a group of comparative experiments with same training and test sets, whose statistical error is poorer than that of hybrid type deep belief network. Therefore, the convinced network model can automatically extract the complex changes in wind speed relying on its strong nonlinear mapping ability to improve the prediction accuracy and robust- ness of the regression model effectively. This study in deep learning will provide the basis of practical ap- plication in wind speed prediction.
出处
《节能技术》
CAS
2016年第1期81-86,共6页
Energy Conservation Technology
基金
国家电网公司科技项目(NY71-13-043)
国家重点基础研究发展计划973资助项目(2012CB215201)
关键词
风速
日前预测
深度学习
自动编码器
混合型
深信度网络
wind speed
day - ahead prediction
deep learning
autoencoder
hybrid
deep belief network
