Uncertainty-Guided Reduced-Order Models for Real-Time System Prediction

Abstract

This study addresses the high computational cost associated with full-order models (FOMs) in simulating complex dynamical systems, which limits their applicability in real-time prediction and control. To address this issue, we present uncertainty-controlled reduced-order modeling (ROM) framework, which combines model reduction and probabilistic uncertainty estimation. The suggested methodology creates a low-dimensional system representation and incorporates the uncertainty-based mechanisms to improve the predictability with respect to different operating conditions. Threefold are the main contributions of this work. To begin with, the framework is greatly superior in accuracy in prediction over the traditional methods of ROM because it uses information of uncertainty in the training and inference of the model. Second, it offers consistent measure of uncertainty, making it possible to be confident in decision-making using well-calibrated prediction intervals. Third, the model has significant computation efficiency, thus can be applied to real time applications. The standard accuracy and uncertainty measures are used to assess the performance of the proposed method. Relative L2 Error and Root Mean Square Error (RMSE) are used to measure prediction accuracy and Prediction Interval Coverage Probability (PICP) and Continuous Ranked Probability Score (CRPS) are used to gauge the quality of uncertainty. Also, computational efficiency is measured in terms of speed-up ratios over full-order simulations. Findings indicate that the developed uncertainty-driven ROM has an acceptable trade-off between the accuracy, reliability, and calculation time, which supports the possibility of real-time system prediction using the new uncertainty-guided ROM in engineering.