基于卷积神经网络和状态时间序列的参数辨识

Parameter estimation based on convolutional neural network and state sequence

针对参数辨识过程运算时间长的问题,本文提出一种基于卷积神经网络的参数辨识方案。该方案避免了对数值模型的大量迭代,能够根据多个连续时间步的实测系统状态对多参数进行快速估计,实现参数辨识;同时,为了帮助神经网络更好地提取特征,还引入一种双向标准化的方法对数据进行处理。以Lorenz63为实例,对其参数进行了分析、实验。实验结果表明:该方案能够有效地估计当前物理场状态对应的模型参数,并且计算时间仅为传统方法的4%,大大提升了计算效率。

To reduce the time cost in parameter estimation processes,a new parameter estimation scheme based on a Convolutional Neural Network is proposed.This scheme can quickly estimate multiple parameters based on a temporal sequence of system states by avoiding time-consuming iterations of numerical models.At the same time,a two-way standardization method is used to help the neural network extract features better.The scheme has been tested in a Lorenz63 nonlinear system.Results show that it can effectively estimate model parameters corresponding to the current physical state with a calculation time only 4%of the particle swarm optimization method.