Working with uncertainty in adaptive process optimisation

This contribution is concerned with adaptive processing decisions, where the process parameters are optimally adapted to the varying properties of the material input stream. Our starting point is the geometallurgical paradigm that the varying properties of the input stream are considered known e.g. from a geometallurgical model of the mined ore body, and optimal processing parameters are computed from them, by finding the parameter seaming optimal in a computer simulation.
This approach however has to work with a lot of uncertainties: The prediction of the geometallurgical ore parameters can only be done with some geostatistical uncertainty. The parameters themselves are only proxies for true ore properties. Model prediction can differ relevantly from actual process results, due to model simplifications. Due to these uncertainties the computed processing choices can turn out to be inferior to simple non-adaptive processing.
We systematically analysed this effect, by modelling this uncertainty effect mathematically and in computer simulations.
The most important findings are:
(a) Processing choices not taking into account the uncertainty sometimes even perform worse than simple non-adaptive processing, for the sole reason of ignoring the uncertainty effect.
(b) Ore properties, not adequately reflected in the ore description, might require different approaches, in which the observed processing behaviour feeds back into process control.
(c) Bayesian decision theory allows computing optimal processing choices combining the information from the mine (geostatistical predictions) and from the process feedback. These choices give much more robust choices and do not suffer from the drawbacks described for the simple approach we started from.
This new approaches can substantially improve the performance of adaptive processing in existing plants.