A neural internal model control scheme for an industrial rotary calciner

Rotary kilns are used for a wide variety of processes involving gas-solid or solid-solid reactions where extensive mass and heat transfer take place. Sodium carbonate, also known as soda ash, is produced by calcining the natural trona ore in direct fired rotary kilns. The large scale of the calcination process and high cost of operation do not leave much room for experimentation. To evaluate such a process and design appropriate control systems, mathematical modeling and computer simulation are needed.; Chemical processes that exhibit nonlinearities and time delays are demanding in control requirements. The soda ash kiln has a nonlinear nature which possesses a long dead time, and is often under the influence of large, long-term process disturbances. For economical and product quality reasons, the kiln operating conditions must be controlled in order to obtain a stabilized kiln temperature profile which in turn leads to a steady product quality.; The ability of multilayer feedforward neural networks (NN) to model several nonlinear function suggests that they may provide a promising approach for modeling nonlinear processes. A neural network is trained to learn the inverse dynamics of the process. This neural net is then used as the nonlinear controller within an internal model control loop.; The primary target of this control scheme is to maintain product quality, but the response in calciner shell and heat shield temperatures is also important. Because the kiln is direct fired, the shell temperature might go beyond the critical temperature of mild steel causing deformation in the kiln shell. The heat shield lowers the shell temperature, and thus avoids shell overheating and deformation while reducing the heat lost to the surroundings. However, even in the presence of the heat shield, calciner shell temperature needs to be monitored. The desired control system should maintain product quality without consuming an amount of natural gas that would lead to shell overheating.; The performance of the neuro-controller is tested with respect to various load and set point disturbances, and compared with the performance of three feedback control schemes developed previously. Looking at the resulting energy consumption profiles and Integral of Time Absolute Error (ITAE) values for the controlled variable, it can be concluded that the designed neural control scheme can provide good control for the system.