Model Predictive Control of Fuel Cells Using a Neural Network Based Inverted Model

Fuel cells are multi-physics systems which exhibit nonlinear phenomena that are difficult to capture. Model Predictive Control allows to design an optimal controller based on a prediction of the fuel cell dynamical behavior. However, the nonlinear optimal problem involves a large number of variables and is difficult to solve. A new algorithm is presented which is based on model inversion by deep neural networks (Neural Network Inversion-based Model Predictive Control). The feedforward neural network is able to retrieve a sequence of controls from a string of past and future outputs. Next, a specified trajectory is supplied which can be parametrized by a convergence rate. The algorithm finds the convergence rate that optimizes an optimal criterion such as the efficiency of the fuel cell; the sequence of optimal control is obtained by feeding the neural network with the reference trajectory. Using the neural network based model inversion, the sequence of optimal controls over a receding horizon which optimizes a given criterion can be obtained from the parametrized reference trajectory. As usual in Model Predictive Control, the first control input is applied, and the procedure is repeated at the next time instant. This method allows to work out a predictive controller even for nonlinear systems, with very little optimization variables. The simulation results are displayed in the last part.