Model Quality Assessment and Model Mismatch Detection: A Temporal Smoothness Regularization Approach

Abstract

Most industrial controllers are designed based on process models, and hence the closed-loop performance closely depends on the model quality. Since process dynamics variations are inevitable in practical applications, plant-model quality assessment is necessary so that model mismatch can be detected in time. In this article, a novel method based on temporal smoothness regularization is presented for model quality assessment. The linear time variant (LTV) model structure is applied to approximate the process dynamics. To avoid an overfitted model, temporal smoothness regularization is imposed on the model parameter changes so that model generalization ability is guaranteed. On the basis of the LTV model structure, a data-based model quality assessment approach is proposed, and the applicability is demonstrated through representative case studies.