Approach to validation of stochastic dynamic models with initial state uncertainty

It has been well realized that model validation plays an important role in a modelling and simulation development process. One important validation approach is to directly compare the experimental data with the data produced by the simulation. A problem with this approach is that the residual used in the analysis is usually nonstationary and contaminated by unknown initial conditions. A new approach is proposed in this paper, in which a Luenberger observer or a Kalman filter is used to generate the residual. In this way, the generated residual conveys the information of modelling errors, and meanwhile the effect of unknown initial conditions upon it is minimized. Furthermore, the residual is Gaussian and white when no modelling errors exist, a property which can be easily tested. The approach is illustrated with the validation of a biological model, which is typically hard to validate.