基于生成型深信度网络的回归算法鲁棒性分析

摘要

Currently, good result is gained in classification problems by making use of deep learning method that also began to be applied in parts of regression tasks.However, more research have focused on forecasting accu-racy compared with other regression algorithm and the robust problem of forecasting regression algorithm with practi-cal application demand is ignored.A Deep Belief Networks ( DBN) model with three hidden layers is constructed by the Restricted Boltzmann Machine ( RBM) for multistep regression task.Then three typical multistep regression models are built including support vector regression ( SVM) , single hidden layers neural networks ( SHL-NN) and neural networks with three hidden layers ( THL-NN).And day-ahead prediction experiments were carried out with actual wind speed data of wind farm in Ningxia of China.By comparing the experimental results, it’ s found that the prediction error, consisting of both mean value and variance, of the DBN model with only three hidden layers is less than those of the other three typical approaches.Consequently, the generation type DBN is not only of high forecast accuracy, but its algorithm robustness is better than the other three algorithms.And the generation type DBN is generally better able to meet the needs of the actual engineering application of wind speed forecasting.%目前，深度学习在分类问题中取得了很多很好的效果，并开始在部分回归任务得到应用。然而，绝大部分研究重点都集中在相对其他回归算法的预报精度上，而忽视了有实际应用需求的回归算法预报鲁棒性问题。首先基于受限的玻尔兹曼机建立了一个具有3个隐含层的生成型深信度网络多步预测模型；然后，建立了基于单隐含层神经网络、三个隐含层的神经网络以及单核支持向量的典型多步预测模型，并利用4组宁夏地区不同季节的风速数据进行回归算法的稳定性对比实验。实验结果显示，基于受限玻尔兹曼机建立的具有三个隐含层的深信度网络模型的多步预报误差的均值和方差都是最小的。因此，基于生成型深信度网络的回归模型不仅预报精度高，而且此预报算法的鲁棒性也比较好；相对其他三种典型回归算法来说，可以更好地满足风电场风速预报问题的实际工程应用需求。