Influence of experimental observations on n-propylbenzene kinetic parameter estimates
- Kinetic parameters of an n-propylbenzene shock tube oxidation model are determined by weighted least-squares optimisation.
- The influence of individual experimental measurements on the parameter estimates is quantified.
- These regression influence diagnostics are based on omission of data points.
- The methods studied are computationally affordable and do not require the use of surrogate models.
- The diagnostics offer insights into, for example, which observations determine which parameters and to what extent.
We calculate the derivatives of best estimates of kinetic parameters of an n-propylbenzene shock tube oxidation model, as determined by a weighted least-squares optimisation, with respect to experimental observations and compare these derivatives to some influence diagnostics based on omission of data points which are widely used in linear regression analysis. The considered data set comprises of 2378 measured concentrations of 37 stable species at various temperatures, pressures, and equivalence ratios. The methods studied are computationally affordable, as they require only a single optimisation and do not require the use of surrogates. We find that the diagnostics offer many insights into how individual observations influence parameter estimates, such as which observations determine which parameters to what extent. Additionally, the significance of non-linearities is investigated. While we observe that they can be of substantial importance to the derivatives, and improve the numerical conditioning of the involved matrix inversion, we find that results obtained from the linear omission-based diagnostics frequently agree, at least qualitatively, with those obtained from the derivatives.
- This paper draws from the preprint: Influence of experimental observations on n-propylbenzene kinetic parameter estimates.
- Access the article at the publisher: http://dx.doi.org/10.1016/j.proci.2014.05.061
Keywords: n-propylbenzene, regression influence diagnostics,
Associated Project: Numerics