Estimating the Prediction Uncertainty of Dynamic Neural Network Process Models

Assessing the Predictions of Dynamic Neural Networks

“…Further research is necessary to evaluate the proposed methods when (defined) plantmodel mismatches are present. The results shown here and in (Dadhe et al, 2004) promise that the bootstrap methods perform properly even with highly nonlinear systems and that the application within neural network based NMPC schemes can improve the controller performance.…”

“…In (Dadhe et al, 2004) it was stressed that the variance of the error distribution varies with the input of the neural network.…”

“…The data generating process is a nonlinear AR(2) model with exogenous input u originally trained with data obtained from simulation studies of a highly nonlinear biochemical reactor (Dadhe et al, 2004…”

“…This paper introduces the application of the GARCH error model and performs a more extensive simulation study compared to the preceding paper (Dadhe et al, 2004) considering different non-normal error processes, different methods to estimate the resulting prediction intervals and a further application of the bootstrap method to dynamic neural networks. The paper is structured as follows.…”