A Comparison of Two Artificial Neural Networks for Modelling and Predictive Control of a Cascaded Three-Tank System

Process industries are confronted with multifaceted problems, including a high degree of nonlinearity and integrated processes, high energy costs, and stringent environmental regulations. The traditional methods for solving these problems are suboptimal. The quest for an optimal solution for industrial processes with reduced product variability and increased profit margin has since birthed the need to develop efficient design methods. Hence, this study investigated two artificial neural networks (ANN) applications for modelling and predictive control of an experimental cascaded three-tank system - a 3-by-3 multivariable and nonlinear process. To achieve this, the tank process was excited by well-designed input signals to obtain input-output data at a sampling time of 5s. The datasets obtained were used to fit recurrent neural network (RNN) and feedforward neural network (FFNN) models for the system. Thereafter, the identified models were used in the design of predictive controllers. Validation results showed that FFNN gave a better fit than RNN. The closed-loop experimental results also showed the FFNN-based predictive controller displaying an overall superior performance for both servo and regulatory control problem.