Data-driven plant-model mismatch quantification for MIMO MPC systems with feedforward control path

In this paper, an autocovariance-based technique is proposed to estimate plant-model mismatch in unconstrained model predictive control (MPC) systems with feedforward control path and time-varying set-points. Compared with previous works, the variance of output noise (assumed to be white noise) is considered unknown. Moreover, we assume there exist mismatches in both feedback and feedforward control paths. Only input, output data and the MPC tuning parameters are assumed to be known. Based on the unconstrained MPC formulation, the predicted autocovariance matrices of output signals are expressed in terms of the plant-model mismatches and unknown output noise variance. On the other hand, the sampled autocovariance matrices of measured output signals can be calculated using routine operating sampled input data. Estimates of model mismatch and output noise variance are obtained by minimizing the discrepancy between the predicted and sampled autocovariance matrices. The performance of the proposed strategy is demonstrated using a numerical simulation.