Nonlinear Model Predictive Control for Concentrating Solar Power Receiver based on a Transformer Neural Network

Kurzfassung

The purpose of this thesis is to explore the use of Transformer Neural Network (TNN) to model complex system dynamics, to evaluate their potential for use in modern control techniques, and to investigate their applicability in a Model Predictive Control (MPC) controller for the Solar Tower Power Plant Jülich (STJ) at the Deutsches Zentrum für Luft- und Raumfahrt e. V. (German Aerospace Center, DLR) in Jülich. The multi-step ahead prediction capability of a Neural Network (NN) with transformer architecture is used to model the dynamics of the nonlinear multivariable system of the receiver based on real data. By incorporating information about future disturbances or setpoint changes, the predictive behavior of the model facilitates counteracting expected external influences. This property is then applied in a MPC through an Optimal Control Problem (OCP) formulation. To solve the OCP, the PyTorch and SciPy libraries are tested with different optimizers. The first, while being an unconstrained optimizer, is used to solve the constrained problem by means of proposed barrier functions to explore its potential. It is also observed that despite the state of the art NN, its accuracy depends on the training data and that interpolation outside the training region leads to inaccurate or unexpected predicted dynamics. In this work, the PyTorch approach provides similar solutions to constrained optimizers while having faster computational times and better performance based on simulation results. Within the distribution of data on which the NN was trained, it is shown that the proposed Transformer Neural Network (T-NN) enabled MPC controller is capable of tracking the reference while satisfying the constraints and rejecting disturbances or setpoint changes in the proposed test scenarios and in the presence of measurement noise. Moreover, prediction errors are fitted in a Gaussian Process Regressor (GPR) to obtain Uncertainty Quantification (UQ) information insights to be displayed to the operator of the system. For our application, the results show the feasibility of this type of data-based controller and its potential to increase the efficiency and resilience of the system. In-situ test campaigns are needed to confirm these results.