摘要
神经网络对于非线性模型的辨识和非平稳信号的预测,与传统预测模型相比具有较明显的优势,但是神经网络的结构对于信号预测或模型辨识的精度具有较大影响,本文针对广泛使用的BP神经网络预测模型,以太阳黑子数据为例,分析了网络预测的拓扑结构(输入节点数、隐层节点数)及网络允许的训练误差MSE(Mean ofSquared Error)对其预测能力的影响。发现最优网络模型对应于一定的拓扑结构,收敛于某个由MSE目标值决定的最优位置,该收敛位置并不是网络的全局最优点。在此基础上,利用遗传算法,对输入节点数、隐层节点数和MSE目标值进行了优化,得到了最优的网络预测模型。最后,用算例验证了本文对BP网络模型预测精度影响因素分析的正确性。
In identifying of non-linear model and forecasting non-even signal, Artificial Neural Network （ANN） has obvious advantages over traditional forecasting models, but ANN structure has great influence on forecasting and identifying precision. In this paper BP neural network model which is used most widely is aimed at, the sunspot data is utilized, and the structure of ANN （including input layer node number and hidden layer node number） and the admitted training error MSE （Mean of Squared Error） are analyzed in order to make out how they affect the forecasting precision of ANN. It is found that the optimum network possesses a specific structure （including input layer node number and hidden layer node number） and converges on an optimum position which is direct related to MSE target value and is not the whole optimum position. On this basis, the Genetic Algorithm （GA） is utilized to optimize ANN model. The node number of input layer, the node number of hidden layer for 3-layer BP network, and MSE target value are optimized, then the optimum forecasting model of BP network is obtained. At last, an example shows that the analysis of influence factors for forecasting precision of ANN model is correct.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2005年第5期528-534,共7页
Pattern Recognition and Artificial Intelligence
关键词
预测
神经网络
模型
影响因素
遗传算法
优化
Forecasting, Neural Network, Model, Influence Factors, Genetic Algorithm, Optimizing
