Model Uncertainty in Fire Safety Engineering

Abstract: Summary: Traditionally, fire safety has often been addressed with methods based on prescriptive recommendations. The opportunity to use an alternative analytical approach has led to the development of fire safety engineering, beginning with structural fire safety design in the 1960's and today including fire safety design and fire risk analysis in general. The prediction of reality using model calculations and dealing with the errors and uncertainties associated with the calculations are two key tasks for a professional practitioner using an analytical approach. In fire safety engineering, smoke transport models are commonly used to predict the conditions caused by a fire. This is done despite the fact that knowledge of the errors and uncertainties associated with the models is lacking and there are insufficient means available to take them into account. The licentiate dissertation "Model Uncertainty in Fire Safety Engineering" is part of the project "Design Based on Calculated Risk", which is financed by The Swedish Fire Research Board (BRANDFORSK) and The Development Fund of the Swedish Construction Industry (SBUF). The objective of this part of the project was to evaluate the predictive capability of smoke transport models quantitatively, in terms of model error and uncertainty in the model error. The result is an adjustment model that can be used to take model error into account in future model predictions and thereby increase the predictive capability of the model. To exemplify the results of this study, model predictions of the smoke temperature and smoke layer height by the computer model CFAST 2.0 are evaluated by means of multi-scenario analysis. A single-scenario analysis is also carried out on smoke temperature predictions by the models FAST 3.1, FASTLite 1.0 and FPETool 3.2. The analysis shows that the predictive capability of the two-zone models can be questioned and that the model results should not be used uncritically, without consideration of the model error. In the analysis of the scenarios it is concluded that the smoke transport model CFAST 2.0 overpredicts the temperature and underpredicts the smoke layer height. Whether or not this can be considered as a conservative prediction in a specific application depends on how hazardous conditions are defined in that situation. The analysis also shows that the model error can be quantified and taken into account, thus increasing the accuracy of the model predictions. A general all-scenario adjustment factor can not be derived, due to the variation in the predictive capability with the type of scenario. For a prediction in a specific scenario the adjustment model can be used to derive a conservative estimate of the model output to be used in a deterministic analysis. The adjusted temperature can also be expressed as a distribution, if the prediction is to be used in a probabilistic uncertainty analysis. Even if the model error is taken into account, there will still be some bias and uncertainty in the adjusted predictions, but substantially less then before. If uncertainty is not taken into account, it is impossible to ensure that quantitative design criteria are fulfilled, which can lead to unsafe designs or unnecessarily expensive buildings. The choice of model can have severe effects on the result. To be able to evaluate the severity of the model uncertainty in relation to the total uncertainty in the assessment, a comparison with the other types of uncertainties included in the calculations, e.g. the uncertainty in input data, uncertainty in other predicted variables, etc., must be performed. This is possible quantitatively if the statistical method presented in this dissertation is used.