A Novel Data-Driven Algorithm for Prediction Horizon Estimation in Model Predictive Control.

Model predictive control (MPC) is a widely used advanced control strategy in industrial applications. The prediction horizon is one of its most influential tuning parameters, as it directly affects both control performance and computational demand. Despite its importance, systematic methods for its configuration remain scarce in the literature. This paper proposes a novel algorithm for prediction horizon estimation based on cross-correlation analysis of input and output data simulated by a trained long short-term memory (LSTM) network model of the controlled process. The use of LSTM networks allows the method to simulate process behaviour directly, eliminating the need for experiments on the physical system. Furthermore, this enables the method to work entirely offline, allowing the prediction horizon to be determined prior to deployment. The proposed algorithm is evaluated on two representative benchmark systems: a continuous stirred-tank reactor and a single tank system. LSTM models are trained for both benchmark systems and are subsequently integrated into an MPC framework. Closed-loop simulations demonstrate that MPC controllers designed with the estimated prediction horizons achieve strong control performance across both benchmark systems. The results suggest that cross-correlation analysis of LSTM-simulated data provides a reliable and systematic basis for prediction horizon estimation, contributing a practical tool for MPC tuning in industrial process control.