Data-Driven Prediction of Unknown Power Electronic Systems Using a Self-Organizing Adaptive Deep Belief Network

The increasing integration of power electronic converters into the power grid, driven by the need to meet diverse power requirements of end-users, has led to the emergence of complex nonlinear systems. Developing accurate mathematical models and predicting the future states of these Unknown Nonlinear Power Electronic Systems (UNPESs) is highly challenging and requires deep knowledge of various aspects, including converter topologies, control strategies, modulation techniques, filter modeling, and the characteristics of grids or electrical machines. To address this issue, a data-driven prediction method is proposed. The approach is based on a Deep Belief Network (DBN) with adaptive learning features that reduce the extended training time often required for Restricted Boltzmann Machines (RBMs). The method also features a self-organizing structure to improve the accuracy of next-state prediction for dynamically changing UNPESs. Furthermore, Partial Least Squares (PLS) regression is applied as a supervised learning technique in place of gradient-based backpropagation, which helps to shorten training time and avoid convergence to suboptimal solution for weight parameters. To validate the effectiveness of the proposed method, two case studies are conducted: one involving the general nonlinear Mackey-Glass system, and the other focusing on next-state prediction of a grid-forming converter.