Our statistical modelling consulting services can encompass the full spectrum of problem identification, formulation, software development/modelling and appraisal, or a subset of these as appropriate. Software solutions that are developed can either be an application based on existing modelling tools (e.g. R), our in-house statistical modelling package QuinStats, or bespoke custom software such as the statistical modelling package BrickFit developed for EDF Energy. Our preference is to start with a small initial low-risk contract, which builds upon initial discussions and consultations with the client.
Quintessa's general approach to statistical modelling is to work with the client to postulate a number of candidate models for the behaviour of the system in question, the form of which can be motivated by physical understanding. It is important that the models capture the known variability and uncertainty in the system in question. The available historical data are used to determine which of these models are consistent with the observations. It is generally better to work with several models that are consistent with the data, but which may extrapolate differently, rather than to concentrate on a single ‘best’ model. This enables the assessment of model uncertainty as well as parameter uncertainty. The complexity of the models employed needs to be compatible with the information in the data. When the data available are limited it may be possible to justify quite simple models for system evolution, but as more data become available it is often possible to justify more complex models.
An example application of Quintessa's statistical modelling and software development capabilities is CoreStats. This software is used to model the evolution of the graphite core of Advanced Gas-cooled Reactors (AGRs). Blind predictions are made ahead of reactor inspections, where the data are very expensive to obtain but cover only part of the system. The predictions are then compared with the results of the inspection, to enable a judgement to be made on both the performance of the models (and therefore their predictive power) and on the consistency of the new data with the historic data and physical understanding.