基于粒子群优化深度置信网络的气体绝缘金属封闭开关设备局部放电模式识别

Partial Discharge Pattern Recognition in Gas Insulated Metal-enclosed Switchgear Based on Particle Swarm Optimization Deep Belief Network

气体绝缘金属封闭开关设备(gas insulated metal-enclosed switchgear,GIS)局部放电模式识别是其绝缘缺陷诊断和状态评估的重要部分,为实现放电类型的准确识别,提出了一种基于粒子群优化(particle swarm optimization,PSO)深度置信网络(deep belief network,DBN)的局部放电模式识别方法。该方法通过PSO算法对DBN网络的权值参数进行优化,提高网络对局部放电特征的学习能力。首先,选取现场多平台的4种GIS局部放电类型监测数据组成样本集,用于对所提方法进行分析;其次,用改进的PSO算法结合样本数据确定DBN网络的初始最优权值参数,建立初始DBN网络;然后,利用训练样本对初始DBN网络进行训练,得到局部放电识别模型。最后,基于渤海油田岸电海上动力平台GIS的局部放电数据,采用多种不同局部放电识别模型对数据样本进行算例分析,结果表明:所提的PSO-DBN模型可有效识别GIS设备局部放电类型,相较于传统的DBN网络、多层前馈神经网络(back propagation,BP)、支持向量机(support vector machine,SVM)和卷积神经网络(convolutional neural networks,CNN)具有更高的准确识别率。

Gas insulated metal-enclosed switchgear(GIS) partial discharge pattern recognition is an important part of the insulation fault diagnosis and state evaluation.To achieve accurate identification of discharge types,a method based on particle swarm optimization deep belief network(DBN) was proposed.The weight parameters of DBN network were optimized by particle swarm optimization(PSO) algorithm to improve the learning ability of the network for partial discharge characteristics.Firstly,a sample set of GIS monitoring data of four types of partial discharge was selected to analyze the proposed method.Secondly,the improved PSO algorithm combined with the sample data was used to determine the initial optimal weight parameters of the DBN network and establish the initial DBN network.Then,the partial discharge recognition model was obtained by training the initial DBN network with training samples.Finally,based on the partial discharge data of GIS equipment of offshore power platform in Bohai oilfield,a variety of different partial discharge identification models were used to analyze the data samples.The results show that the proposed PSO-DBN model can effectively identify the type of partial discharge of GIS equipment,and has a higher accurate recognition rate than the traditional DBN network,back propagation(BP),support vector machine(SVM) and convolutional neural network(CNN).