基于广义回归神经网络的特高频局部放电定位法

局部放电的检测和定位是变电站电力设备状态监测和诊断的重要手段。现有的基于时差法的特高频局部放电定位技术，由于高昂的设备成本限制了其应用范围。提出的基于广义回归神经网络和接收信号幅值强度（RSSI）指纹图的局部放电定位法，分为2个阶段。在算法的离线阶段，建立被测区域的RSSI指纹图；在线阶段，利用广义回归神经网络（GRNN）实现对局部放电源的定位。现场测试表明：提出的方法平均定位误差为0.51 m，定位误差小于1 m的累积概率为81.6%。和基于RSSI信号衰减模型定位法的克拉美罗下界（CRLB）最小均方误差相比，均方误差小于0.6 m²的GRNN定位误差累积概率为66.7%，要优于基于信号衰减模型定位方法的CRLB。该方法解决了传统方法定位精度低、成本高的缺点，具有较低的硬件成本和良好的环境适应性。

UHF Partial Discharge Localization Methodology Based on Generalized Regression Neural Network

Partial discharge(PD) detection and localization is an important means for condition monitoring and diagnosis of power equipment. The existing time-difference based ultra-high frequency(UHF) PD localization techniques are limited in application due to their high costs. A novel PD localization method is proposed based on generalized regression neural network(GRNN) and received signal strength indicator(RSSI) fingerprint, which consists of two stages. In the off-line stage of algorithm, a RSSI fingerprint map is built. In the on-line stage, the GRNN is used to calculate the position of the PD source. The field testing shows that the proposed UHF PD localization method has an average localization error of 0.51 m, and a cumulative probability of 81.6% for the localization error of less than 1 m. Compared to the minimum mean square error(MSE) of the Cramér-Rao lower bound(CRLB),which is based on RSSI log normal shadowing model positioning method, the cumulative probability of the GRNN localization error with the mean square error less than 0.6 m2 is 66.7%, which is better than CRLB. The proposed method overcomes the shortcomings of low positioning accuracy and high costs of the traditional methods, and has the characteristics of low hardware cost and good environmental adaptability.