Abstract:
This work builds on existing observable dynamic models of froth flotation circuits, aimed at on-line
parameter estimation and model-based control. The models are analysed and two main limitations are
identified and addressed: the lack of explicit modelling of reagent effects and the need for dynamic
validation on large-scale industrial plant data.
The feasibility of expanding a froth flotation model to include reagent effects is investigated. A Sobol
sensitivity analysis is used to identify the crucial parameters. The model is expanded with two different
reagent effect models. Both expansions include mass balance models of the frother concentration in
each cell. The first model expands an empirical parameter in the air recovery model, related to the froth
height at which peak air recovery (PAR) is achieved, as a linear function of frother concentration. The
second model adds a linear frother concentration term to the existing air recovery model to modify the
steady-state air recovery directly. Observability analyses of the expanded models show that all states
and the important time-varying model parameters are observable (and identifiable) from the available
on-line measurements. Most importantly, the frother concentrations are shown to be observable without
concentration measurements. Simulations of the model expansions show that the second model can
qualitatively predict the impact of increased frother dosage on air recovery, grade and recovery, while
the first model can only predict the correct effect under certain conditions.
The implementation of a Moving Horizon Estimator (MHE) based on the model (excluding reagent
effects) on data from an industrial rougher bank is investigated with the aim of validating the model
and parameter estimation approach. The available plant data and its limitations are discussed and
additional model analysis is conducted. An expanded observability analysis of the model identifies
groups of parameters for which identifiability is linked. It is shown that without on-line compositional
measurements only a reduced model that lumps all recovery mechanisms into a single empirical
equation is observable.
The reduced model is used to develop the MHE which is implemented on data from the Mogalakwena
North Concentrator (MNC) historian. The state and parameter estimates are then used to evaluate
the model prediction accuracy over a shifting control horizon, as would be done in model predictive
control (MPC). Estimation results show that there are substantial amounts of unmodelled dynamics
and/or disturbances. Parameter estimates compensate somewhat, but the model predictions are only
accurate over some sections of the data.
The lack of on-line compositional measurements as well as uncertainty regarding the validity of
calculated measurements and assumptions prevented a fair evaluation of the full potential of the model,
but served to highlight drawbacks and challenges that will need to be addressed in future work.
