BP 神经网络在函数估计中的应用.docVIP

摘 要 在系统建模、辨识和预测中，对于线性系统，在频域，传递函数矩阵可以很好地表达系统的黑箱式输入输出模型；在时域，Box-Jenkins方法、回归分析方法、ARMA模型等，通过各种参数估计方法也可以给出描述。对于非线性时间序列预测系统，双线性模型、门限自回归模型、ARCH模型都需要在对数据的内在规律知道不多的情况下对序列间关系进行假定。 传统的非线性系统预测，在理论研究和实际应用方面，都存在极大的困难。相比之下，神经网络可以在不了解输入或输出变量间关系的前提下完成非线性建模。神经元、神经网络都有非线性、非局域性、非定常性、非凸性和混沌等特性，与各种预测方法有机结合具有很好的发展前景，也给预测系统带来了新的方向与突破。建模算法和预测系统的稳定性、动态性等研究成为当今热点问题。目前在系统建模与预测中，应用最多的是静态的多层前向神经网络，这主要是因为这种网络具有通过学习逼近任意非线性映射的能力。利用静态的多层前向神经网络建立系统的输入/输出模型，本质上就是基于网络逼近能力，通过学习获知系统差分方程中的非线性函数。 关键词：人工神经网络，函数估计，非线性预测，MATLAB ABSTRACT In system modeling , identification and prediction , for linear systems , in the frequency domain , the transfer function matrix can be a good expression system black-box input -output model ; in the time domain , Box-Jenkins method of regression analysis , ARMA models through a variety of parameter estimation methods described can also be given . For nonlinear time series forecasting systems, bilinear model , threshold autoregressive model , ARCH models are needed in the internal laws of the data to know the case of small assumption on the relationship between sequences . The traditional nonlinear system predicts the theoretical research and practical applications , there are great difficulties. In contrast , the neural network can not understand the relationship between the input or output variables completed under the premise of nonlinear modeling . Neurons, neural networks are non-linear , non-locality , unsteady , non- convexity and chaos characteristics combine with a variety of forecasting methods have good prospects , but also to bring a new prediction system direction and breakthrough. And predictive modeling algorithm system stability , dynamics and other research become hot issue. Currently in the system modeling and prediction , the most widely used is static multilayer feedforward neural networks, this is mainly because this network has approximate any nonlinear mapping through learning ability . Use static multilayer feedforward neural network sys