用神经网络估计模型误差的预测滤波算法

Predictive filters with model error estimations using neural (networks)

针对时不变非线性系统,提出一种用神经网络进行模型误差估计的预测滤波算法.该算法用寻优的方法离线获得与当前状态和下一步输出测量相对应的模型误差估值,并作为样本训练神经网络;实际滤波中,用训练好的神经网络进行模型误差估计.该方法与原预测滤波算法相比没有动态过程,不会因为滤波器初始误差太大而振荡或发散,且稳态精度与计算步长无关.通过对一个二阶非线性系统的仿真验证了神经-预测滤波器的优越性.

A new algorithm of predictive filters for time-invariable nonlinear systems is proposed. A neural network is used to estimate the model error, which is trained off-line by a given data set composed of state estimations, outputs and model error estimations. Unlike the original predictive filter, this algorithm does not have dynamic processes and is not vibrating or divergent while initial estimation errors are big. The precision is high and does not decrease with the increase of calculation steps. A simulation on a two-dimension system shows the advantage of the neuro-predictive filter.