Sviluppo di Tecniche di Intelligenza Artificiale nei Convertitori di Potenza e negli Elementi di Accumulo

In recent years, artificial intelligence (AI) has experienced extraordinary growth, profoundly transforming numerous industrial and scientific sectors. Among the most significant innovations are generative neural networks, which have shown revolutionary potential in data analysis, language processing, and image recognition. These advanced tools have been successfully adopted in fields such as medicine, finance, and automated design, leading to new frontiers of efficiency and predictive capabilities. However, the application of AI in power electronics remains relatively recent, with significant room for further development and refinement. In this context, the present thesis aims to explore the use of innovative AI techniques to address challenges related to the dynamic black-box modeling of power converters and, applying the knowledge gained in terms of design and optimization of neural networks, to extend the research to the estimation of the state of health (SoH) of lead-acid batteries. The first objective of this research is the development of dynamic black-box models for power converters using artificial neural networks (ANN)s. These models are initially applied to a boost converter and later extended to other dc-dc converter topologies. The focus is on nonlinear autoregressive with exogenous inputs (NARX) ANNs, which have shown the best trade-off between accuracy and computational complexity. The analysis starts with a comparison between different datasets since the quality and quantity of data are crucial for the performance of the models. Afterwards, the hyperparameters of the NARX-ANNs such as the number of neurons, activation functions, number of delays, and sampling frequency are evaluated to improve the accuracy of the models while keeping the computational complexity low. The outcomes demonstrated the high effectiveness of NARX-ANNs in capturing the complex dynamic behaviors of converters even with strong nonlinearities in the system and different operating conditions. This represents a significant advancement in the modeling of complex systems that have traditionally relied on physics-based approaches and extends the literature about the use of AI for the modeling of power converters. The second objective focused on the analysis of transfer learning (TL) capabilities for estimating the output impedance of power converters, with the aim of improving accuracy and reducing data requirements. This technique, still relatively unexplored in the context of power electronics, showed a remarkable generalization ability, allowing models to adapt to different converter topologies with minimal retraining. The results indicate that TL can be a powerful tool for optimization and reducing development time in real-world applications. The third objective concentrated on the SoH of lead-acid batteries using AI techniques applied to measurements obtained from power converters. Although this field has been extensively studied for lithium-ion batteries, its application to lead-acid batteries is less developed. In this scenario, the aim is to apply the knowledge gained from the previous objectives to develop accurate SoH estimation models. In particular, starting from the analysis of the battery performance obtained from the converter measurements, the research focuses in the analysis of the capabilities of different AI algorithms to estimate the SoH of these batteries.