摘要
由于变电站现有的状态监测方法分类情况较差、故障误警率高,为此研究基于机器学习的变电站运维状态实时监测方法。首先,运用三比值法对变电站设备中的气体浓度进行分析,判断变电站设备内部的运行情况。其次,运用小波分解算法对变压器变化特征曲线进行分解并重构,计算不同频段时域所获得能量,相加后得到其特征量。再次,在决策树中获得最大的信息增益,计算不同数据信息的梯度值并将其进行关联。从次,运用遗传算法对模型进行优化,对粒子的位置寻优。将最优粒子结果带入模型中完成学习,输出状态值。最后,计算状态向量的误差值,设置状态告警阈值。如果误差值超过阈值则进行告警,从而完成监测。实验结果表明,故障监测阈值为-0.025时,实验组的故障误警率值为0%,识别分类准确,具有较好的监测性能。
Due to the poor classification and high false alarm rate of existing condition monitoring methods for substations,a real-time monitoring method for substation operation and maintenance status based on machine learning is studied.Firstly,the three-ratio method is used to analyze the gas concentration in substation equipment,so as to judge the internal operation of substation equipment.Secondly,wavelet decomposition algorithm is used to decompose and reconstruct the characteristic curve of transformer,and the energy obtained in different frequency bands in time domain is calculated,and its characteristic quantity is obtained after adding.Once again,the maximum information gain is obtained in the decision tree,and the gradient values of different data information are calculated and correlated.Secondly,the genetic algorithm is used to optimize the model and optimize the position of particles.The optimal particle results are brought into the model to complete the learning and output the state values.Finally,the error value of the state vector is calculated and the state alarm threshold is set.If the error value exceeds the threshold,it will give an alarm,thus completing the monitoring.The experimental results show that when the fault monitoring threshold is-0.025,the fault false alarm rate of the experimental group is 0%,which is accurate in identification and classification and has good monitoring performance.
作者
张萌
ZHANG Meng(Ultra-high Voltage Branch of State Grid Chongqing Electric Power Company,Chongqing 400000,China)
出处
《通信电源技术》
2023年第14期220-222,共3页
Telecom Power Technology
关键词
机器学习
变电站
运维状态
实时监测
machine learning
transformer substation
operation and maintenance status
real time monitor
