基于集成深度神经网络的配电网联络关系辨识技术

Distribution Network Connectivity Recognition Based on Ensemble Deep Neural Network

随着城市配电网络规模不断扩大,配电网实时拓扑难以获取已成为观测其运行状态的主要瓶颈。为了解决传统拓扑辨识方法噪声敏感性高、在线运行难等问题,提出了一套基于集成深度神经网络的配电变压器(简称配变)联络关系辨识方案。首先,依据配电网测量的横纵连续性,对历史数据进行二维小波阈值去噪,降低噪声对辨识结果的影响。为提高深度学习算法的精度上限,采用搜索、生成与评价的策略对数据进行特征提取与选择。然后,以选择的特征为输入,构造交叉熵深度神经网络,通过网格搜索优化深度神经网络的超参数。采用集成学习的策略训练同质深度神经网络,保证模型的在线拓扑辨识能力。最后,通过在TensorFlow上进行的实验验证了提出的集成深度神经网络模型在配变联络关系辨识中的精确度与鲁棒性。

With the development of the urban distribution network, the difficulty in obtaining the real-time distribution network topology has become an important bottleneck for observing its operational status. To tackle high noise sensitivity and difficult online operation in traditional topology recognition methods, this paper proposes a distribution transformers(DTs) connectivity recognition method based on ensemble deep neural network. Firstly, based on the horizontal and vertical continuity of distribution network measurements, the two-dimension wavelet threshold de-noising method is implemented on historical measurements to reduce the impact of noise on the recognition results. To improve the upper limit of deep learning algorithm, this work proposes the search, generation and evaluation strategy for feature extraction and feature selection. Secondly, we use the extracted features as inputs for the deep neural network with cross-entropy function. The grid search algorithm is used to optimize the hyperparameters of deep neural network. An ensemble learning strategy is utilized to train homogeneous deep neural networks to ensure the ability of online topology recognition. Finally, the experiments implemented on TensorFlow certificate the accuracy and robustness of proposed ensemble deep neural network models in distribution transformers connectivity recognition.