【摘要】 接地网是保障电网完全运行的重要部件,但接地网材料易被腐蚀,甚至发生断裂。鉴于接地网腐蚀存在样本数目少、非线性强的特点,在引入对比学习(contrastive learning,CL)和度量学习(metric learning,ML)的基础上,将对比学习和度量学习进行了组合优化,使输出结果变为参与拟合锚点的腐蚀速率系数,在此基础上提出了基于改进CL-ML方法的接地网不开挖腐蚀速率预测模型。该方法把已经采样的珍贵样本作为锚点,显著减小了拟合函数的压力与复杂性,并充分挖掘了稀有样本的内在相关性。试验结果表明,采用该模型的预测结果比采用广义回归神经网络和BP神经网络具有更高的精度。更多还原

【Abstract】 Grounding grid is an important component to ensure the safety operation of power grid. However, the grounding grid material is easy to be corroded or even broken. In view of the small number of samples and strong nonlinear characteristics of grounding grid corrosion, the idea of contrastive learning(CL) and metric learning(ML) was introduced in this paper. The combined optimization of the CL and ML method was carried out to change the output result into the corrosion rate coefficient involved in fitting the anchor point. Thus, a new non-excavation corrosion rate prediction model of grounding grid based on improved CL-ML method was proposed. This method takes the precious samples as anchor points, which not only significantly reduces the pressure and complexity of fitting function but also fully excavates the intrinsic correlation of rare samples. The experimental results show that the prediction results of this model are more accurate than those of generalized regression neural network and BP neural network.更多还原