From multi-physics models to neural network for predictive control synthesis - IMT Atlantique Accéder directement au contenu
Article Dans Une Revue Optimal Control Applications and Methods Année : 2021

From multi-physics models to neural network for predictive control synthesis

Résumé

The aim of this document is to present an efficient and systematic method of model-based predictive control synthesis. Model predictive control requires using a model of a dynamical system, that can be linear, time-varying, non-linear or identified from data. Finding a model that is both precise and simulatable at low computational cost can be challenging and time consuming due to requiring extensive knowledge of the system and physics as well as a large volume of data with relevant scenarios and sometimes a complicated identification work (filtering noises and bias, data formatting, etc.). The proposed methodology begins with fine-scale multi-physics modelling, which is possible thanks to open model libraries (see Modelica). The obtained model is then simulated by considering ad hoc scenarios to generate data, which are then used to identify a neural network, that will support the predictive control syntheses. The systematic methodology is detailed and applied to the widely used control benchmark known as the quadruple tanks process. Results show that the methodology is accurately applied to optimize hyperparameters in finding a neural network model and to control the quadruple tanks process with the predictive controller.
Fichier principal
Vignette du fichier
From_multi_physics_models_to_neural_network_for_predictive_control_synthesis.pdf (903.16 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03510668 , version 1 (04-01-2022)

Identifiants

Citer

Pierre Clément Blaud, Philippe Chevrel, Fabien Claveau, Pierrick Haurant, Anthony Mouraud. From multi-physics models to neural network for predictive control synthesis. Optimal Control Applications and Methods, 2021, ⟨10.1002/oca.2845⟩. ⟨hal-03510668⟩
90 Consultations
489 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More