基于改进深度置信网络的电力系统暂态稳定评估研究

Research on Power System Transient Stability Assessment Based on Improved Deep Belief Network

为了提高深度置信网络的评估性能,提出一种基于稀疏降噪自动编码器和深度置信网络相结合的暂态稳定评估方法。首先,构建一组对系统暂态变化敏感且维数与系统规模无关的原始输入特征;其次,通过稀疏降噪自动编码器的无监督学习过程提取输入特征,用得到的权值和偏置初始化深度置信网络;最后,采用“预训练-微调”2种学习方法训练深度置信网络,获得原始输入特征与系统暂态稳定结果之间的映射关系。与采用随机初始化受限玻尔兹曼机的传统深度置信网络相比,本文提出的改进评估方法在一定程度上克服了由于随机初始化导致评估准确率无法达到最优的弊端。在新英格兰10机39节点系统上的仿真结果表明,该方法比常用的机器学习算法和深度置信网络有更好的评估性能,仿真结果还证明了本文所提方法具有良好的特征提取能力。

In order to improve the evaluation performance of deep belief network,the paper propose a transient stability evaluation method based on sparse denoising auto-encoder and the deep belief network.Firstly,a set of original input features,that are sensitive to system transient changes and whose dimensions are independent of system size,is constructed.Then the input features are extracted by the unsupervised learning process of the sparse denoising auto-encoder,initializing the deep belief network with the obtained weights and offsets.Finally,two learning methods,pretraining and fine tuning,are used to train the deep belief network to obtain the mapping relationship between the original input features and the system transient stability results.Compared with traditional deep belief network using stochastically initialized restricted Boltzmann machine,the improved evaluation method proposed in this paper overcomes the disadvantages that the evaluation accuracy cannot be optimal due to random initialization.Tests on the New England 10-machine 39-bus system show that the proposed method has better evaluation performance than common machine learning algorithms and the deep belief network,and proves that the proposed method has good feature extraction ability.