Dynamic simulator for a grinding circuit

Abstract

The grinding circuit is a primary and indispensable unit of a mineral processing plant. The product from a grinding circuit affects the recovery rate of minerals in subsequent downstream processes and governs the amount of concentrate produced. Because of the huge amount of energy required during the grinding operation, they contribute to a major portion of the concentrator cost. This makes grinding a crucial process to be considered for optimization and control. There are numerous process variables that are monitored and controlled during a grinding operation. The variables in a grinding circuit are highly inter-related and the intricate interaction among them makes the process difficult to understand from an operational viewpoint. Modeling and simulation of grinding circuits have been used by past researchers for circuit design and pre-flowsheet optimization in terms of processing capacity, recovery rate, and product size distribution. However, these models were solved under steady approximation and did not provide any information on the system in real time. Hence, they cannot be used for real time optimization and control purposes. Therefore, this research focuses on developing a dynamic simulator for a grinding circuit. The Matlab/Simulink environment was used to program the models of the process units that were interlinked to produce the flowsheet of a grinding circuit of a local gold mine operating in Alaska. The flowsheet was simulated under different operating conditions to understand the behavior of the circuit. The explanation for such changes has also been discussed. The dynamic simulator was then used in designing a neural network based controller for the semi-autogenous mill (SAG). A two-layer non-linear autoregressive (NARX) neural network with feed to the mill as exogenous input was designed using data generated by the simulator for a range of operating conditions. Levenberg-Marquardt (LM) and Bayesian Regularization (BR) training algorithms were used to train the network. Comparison of both algorithms showed LM performed better provided the number of parameters in the network were chosen in a prudent manner. Finally, the implementation of the controller for maintaining SAG mill power to a reference point is discussed.