Abstract:
A new space-time neural network is proposed using the function expansion technique and the linear impulse response filtering theory in this paper. It consists of function expansion and linear delay pulse. Net input space is mapped into a high dimensional space by function expansion. Therefore, nonlinear mode in low dimensional space can be converted to linear mode in high dimensional space. Linear delay pulse is equivalent to the temporal linear impulse response filter, which is responsible for fitting linear model in space-time series. Space-time neural network fast learning algorithm is proposed by using Levenberg-Marquardt optimization method. Simulation results show that space-time neural network has the characteristics of fast convergence and high precision. Compared with Space-time Autoregressive Moving Average(STARMA) and multilayer perceptron neural network, the prediction accuracy of the space-time neural network is significantly improved.
Key words:
space-time neural network,
Functional Link Artificial Neural Network(FLANN),
linear impulse response filtering,
Space- time Autoregressive Moving Average(STARMA) model,
space-time series,
airport noise
摘要： 针对时空序列建模分析问题，利用函数扩展技术，结合线性脉冲响应滤波原理，提出一种新型时空神经网络。该网络由函数扩展功能模块和线性延时脉冲模块组成。函数扩展功能模块将网络输入空间映射到高维空间，实现时空序列非线性模式到高维映射空间线性模式转换；线性脉冲延时模块等效于时空线性脉冲响应滤波器，用于拟合时空序列中的线性模式。采用Levenberg- Marquardt最优方法对网络进行训练，设计时空神经网络的快速学习算法。机场噪声仿真结果表明，该网络具有快速收敛和高精度的特点，预测精度高于时空自相关移动平均模型和多层感知器神经网络。
关键词:
时空神经网络,
函数链接人工神经网络,
线性脉冲响应滤波,
时空自相关移动平均模型,
时空序列,
机场噪声
CLC Number: