The main goal of estimating models for industrial applications is to guarantee the cheapest system identification. The requirements for the identification experiment should not be allowed to affect product quality under normal operating conditions. This paper deals with ensuring the required liquid levels of the cascade system tanks using the model predictive control (MPC) method. The MPC strategy was extended with the Kalman filter (KF) to predict the system’s succeeding states subject to a reference trajectory in the presence of both process and measurement noise covariances. The main contribution is to use the application-oriented input design to update the parameters of the model during system degradation. This framework delivers the least-costly identification experiment and guarantees high performance of the system with the updated model. The methods presented are evaluated both in the experiments on a real process and in the computer simulations. The results of the robust MPC application for cascade system water levels control are discussed.