Model Estimation Ensuring Passivity by Using Port-Hamiltonian Model and Deep Learning

An accurate model is necessary for highly accurate control, but it is not always easy to obtain the model via first principles. One of the methods for creating models is to represent the model by neural networks and train them in accordance with the data. However, the model created by machine learning cannot always satisfy the physical properties of the system. If some prior knowledge can be imposed on the estimation, it can be beneficial in the application of the obtained model and the reduction of the burden needed for the training. In this paper, we propose the new method to reflect the passivity of the system by using a port-Hamiltonian form. The effectiveness of the proposed method is shown via numerical examples.