基于新型元启发式反向传播神经网络的500 kV输电线路覆冰厚度预测  

Prediction of icing thickness for 500 kV transmission lines based on back propagation neural network with a new metaheuristic optimization algorithm

作　　者：苏仁斌 熊卫红 刘先珊[2] 李智 邹建明 曾垂辉 SU Ren-bin;XIONG Wei-hong;LIU Xian-shan;LI Zhi;ZOU Jian-ming;ZENG Chui-hui(Central China Branch of State Grid Corporation of China,Wuhan 430072,China;School of Civil Engineering,Chongqing University,Chongqing 400045,China)

机构地区：[1]国家电网有限公司华中分部,武汉430072 [2]重庆大学土木工程学院,重庆400045

出　　处：《兰州大学学报(自然科学版)》2025年第1期17-25,34,共10页Journal of Lanzhou University(Natural Sciences)

基　　金：国家电网有限公司华中分部科技项目(52140023000A)。

摘　　要：实际输电线路的覆冰监测数据为高维非线性时间序列,以某500 kV典型覆冰线路为研究对象,结合监测序列的统计特征,提出基于新型元启发算法瞪羚优化方法(GOA)的反向传播神经网络(BPNN)模型进行覆冰厚度预测.基于10个标准测试函数的最优解求解,验证优化算法对复杂问题全局最优求解的适用性,引入该算法建立GOA-BPNN模型,仿真不同训练样本数的正弦函数,说明该方法能更快地收敛于最优解.根据线路的5 a覆冰监测数据,基于相关系数矩阵及主成分分析法,将6个主控因子降维为4个,作为GOA-BPNN模型的输入层,构建符合线路特征的覆冰厚度GOABPNN预测模型.该模型针对短时间覆冰序列的预测结果比经典BPNN模型的预测值更准确,验证了其对高阶非线性覆冰时间序列的泛化学习能力.以线路的多年覆冰长时间序列为训练集,预测得到5个时刻的覆冰厚度,GOA-BPNN模型相对其他4个模型的预测值最接近实际监测值,模型对“微地形、微气象”环境中的覆冰厚度预测具有较高的可靠性.The real icing monitoring data of transmission lines is a high-dimensional nonlinear time series,taking a typical 500 kV icing line as a case study,a new meta heuristic algorithm based on the gazelle optimization algorithm back propagation neural network(GOA-BPNN) model was proposed based on the statistical characteristics of the monitoring series.10 standard test functions were optimized by GOABPNN model and three other models to demonstrate the former's ability and efficiency and reliability for optimization in complex problems.A GOA-BPNN model was established to simulate the sine function with different training sample numbers,demonstrating its greater convergence and higher efficiency to approach the optimal solution.Using the 5-year ice monitoring data of the proposed line,the correlation coefficient matrix and principal component analysis were applied to reduce six main control factors to four main components as input layers for the GOA-BPNN model,to predict its ice thickness.It could be observed from the prediction of the short-term icing time series that the predicted values with the GOABPNN model were more accurate than with the classic BPNN model,demonstrating its generalization learning ability for high-order nonlinear icing time series.The long-term ice sequence as the training set had been deeply researched into,and it could be seen from the predicted ice thickness at 5 different time that the predicted values were closest to the monitoring values compared to the other four models,indicating the proposed model as being still reliable in predicting ice thickness in "micro topography and micro meteorology" environments.

关 键 词：瞪羚优化算法 GOA-BPNN模型 主成分分析 覆冰厚度 预测模型 

分 类 号：TM721[电气工程—电力系统及自动化]