摘要
针对普通神经网络的梯度消失和易陷入局部极值的问题,提出一种基于多元宇宙优化算法(multi-verse optimizer,MVO)的BP神经网络优化方法(MVO-BP),利用MVO全局寻优的特性求取BP神经网络各层之间可靠的神经元阈值与连接权,从而使神经网络预测模型具备更高的预测精度。建立基于MVO-BP算法的GNSS高程异常拟合预测模型,并采用实际工程中少量高程异常数据进行算法可行性检验。结果表明,相较于常规的BP神经网络法及多面函数法,MVO-BP法精度更高、适用性更强,可为实际工程测量中正常高的求取提供参考。
We are concerned with the problem of gradient vanishing and ease of falling into local extremum of ordinary neural network.To enable the neural network prediction model to be more accurate in prediction,we apply the global optimizing feature of multi-verse optimizer(MVO)to retrieve the reliable neuron threshold and connection weight between each layer of BP neural network.We build the prediction model of GNSS height anomaly fitting based on the MVO-BP method,then we carry out the feasibility test of the algorithm by adopting a limited amount of height anomaly data in practical engineering.The results show that MVO-BP method is more accurate and versatile than the conventional BP neural network method and the multifaceted function method,and it has a certain reference value for the acquisition of normal height in practical engineering measurements.
作者
蒙金龙
唐诗华
张炎
何广焕
刘银涛
MENG Jinlong;TANG Shihua;ZHANG Yan;HE Guanghuan;LIU Yintao(College of Geomatics and Geoinformation,Guilin University of Technology,319 Yanshan Street,Guilin 541006,China;Guangxi Key Laboratory of Spatial Information and Geomatics,319 Yanshan Street,Guilin 541006,China)
出处
《大地测量与地球动力学》
CSCD
北大核心
2022年第12期1233-1238,共6页
Journal of Geodesy and Geodynamics
基金
国家自然科学基金(41864002)
广西空间信息与测绘重点实验室基金(2018GXNSFAA281279)
广西中青年教师基础能力提升项目(2021KY0268)。
