Contribution to the analysis of the performance of dynamic neural networks in the modeling of nonlinear systems and classification problems

Abstract

This doctoral work explores the advantages of "black-box" modeling for nonlinear dynamic systems by using neural networks within a state-space framework, an approach that stands apart from traditional input-output models. The thesis demonstrates the power and versatility of this approach through two main contributions, both applied to the complex case study of photovoltaic (PV) systems. The first contribution focuses on the task of modeling. We developed a complete methodology to create a high-fidelity "digital twin" of a PV system using a BiLSTM network. The results show exceptional predictive accuracy (with an R² score over 0.997) and, more importantly, physical coherence, which we validated by analyzing the model's hidden states. This approach turns the "black box" into an interpretable "grey box." The second contribution tackles the task of classification for fault diagnosis. We proposed a robust hybrid framework that combines DWT for feature extraction, NCA for feature selection, and an optimized BiLSTM classifier. This model achieves a classification accuracy of over 96%, showing its superiority in distinguishing between faults with very similar electrical signatures. In summary, this thesis validates the hypothesis that the state-space approach is superior for "black-box" modeling, as it facilitates the development of models that combine precision with a clear physical interpretation. It also confirms the remarkable versatility of dynamic neural networks for both modeling and classification tasks, opening up essential avenues for predictive maintenance and the intelligent control of energy systems.