On the Development of Predictive Models with Applications to a Metallurgical Process

A methodology for constructing useful predictive models, primarily intended for use on line by the process supervisors as a basis for decision making, is presented. The methodology mainly relies on a sound choice of model structure for black-box modelsa choice based on a clear definition of the modeling goal. Some simple approaches to the treatment of a combination of event-driven and continuous information in predictive models are also proposed. Furthermore, directional considerations in the tracking of time-variant parameters are illustrated to be of considerable practical importance for obtaining accurate models, and such considerations are possible to implement without prior knowledge of the underlying dynamics. The topics treated are finally applied to a prediction problem of industrial importance:  The success of the case study clearly motivates the careful consideration of different modeling options.