Self-tuning moving horizon estimation of nonlinear systems via physics-informed machine learning Koopman modeling

Abstract

In this paper, we propose a physics-informed learning-based Koopman modeling approach and present a Koopman-based self-tuning moving horizon estimation design for a class of nonlinear systems. Specifically, we train Koopman operators and two neural networks - the state lifting network and the noise characterization network - using both data and available physical information. The two neural networks account for the nonlinear lifting functions for Koopman modeling and describing system noise distributions, respectively. Accordingly, a stochastic linear Koopman model is established in the lifted space to forecast the dynamic behavior of the nonlinear system. Based on the Koopman model, a self-tuning linear moving horizon estimation (MHE) scheme is developed. The weighting matrices of the MHE design are updated using the pre-trained noise characterization network at each sampling instant. The proposed estimation scheme is computationally efficient because only convex optimization is involved during online implementation, and updating the weighting matrices of the MHE scheme does not require re-training the neural networks. We verify the effectiveness and evaluate the performance of the proposed method via the application to a simulated chemical process.

31 pages, 7 figures

References in corpus (9)

• What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?
• A Data-Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition
• Deep learning for universal linear embeddings of nonlinear dynamics
• Linear predictors for nonlinear dynamical systems: Koopman operator meets model predictive control
• Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics
• Deep Neural Networks with Koopman Operators for Modeling and Control of Autonomous Vehicles
• Robust Tube-based Model Predictive Control with Koopman Operators--Extended Version
• Physics-Informed Neural Nets for Control of Dynamical Systems
• Reduced-order Koopman modeling and predictive control of nonlinear processes