Energy-based Neural Network Controllers for DC-DC Converters

In this paper, our objective is to use notions of system energy to formulate closed-loop dc-dc converter controls that are nonlinear and satisfy passivity properties that guarantee stability. Port-Hamiltonian models are a particular form of models which can be used to describe the total energy in a converter, much like a Lyapunov function. In our approach, we first formulate a port-Hamiltonian model that represents the desired closed-loop dynamics we seek. However, the solution to this model is generally quite difficult for even the simplest of converters. To bypass this challenge, we offer a framework where a neural-network-based controller is trained to estimate the solution to this design problem. Essentially, our objective is to ensure that the energy dynamics of the dc-dc converter with a neural network as a controller closely match that of the target port-Hamiltonian model. This method circumvents the mathematical difficulties encountered when attempting to solve the closed-loop port-Hamiltonian model directly and gives a generalized framework. Our paper illustrates this approach and its versatile application towards boost, buck, buck-boost, and Ćuk converters.