Modelado de una celda de combustible de carbonato fundido mediante el uso de aprendizaje automático.
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Dentro de este estudio se aborda el modelado de una celda de combustible de carbonato fundido (MCFC) mediante la implementación de técnicas de aprendizaje automático, con énfasis en modelos híbridos que incorporan conocimiento físico en el proceso de entrenamiento. Las MCFC representan una tecnología de generación de energía con alta eficiencia y bajo impacto ambiental, gracias a su capacidad para operar con diferentes tipos de combustibles ricos en hidrógeno y su integración en esquemas de cogeneración o captura de CO₂. Sin embargo, su modelado tradicional mediante modelos físicos suele requerir una caracterización detallada del sistema, elevados tiempos de desarrollo y una alta especialización técnica.
Frente a este desafío, se propone el uso de redes neuronales artificiales (ANN) como alternativa para modelar el comportamiento de una MCFC a partir de datos experimentales. Además, se implementó un modelo de redes neuronales guiadas por la física (PGNN), el cual incorpora parte de la física del sistema mediante una función de pérdida modificada en la ANN. Este enfoque busca mejorar la generalización fuera del dominio de entrenamiento, superando las limitaciones de modelos basados solo en datos.
Durante el desarrollo del trabajo se exploraron diferentes configuraciones de red, y luego se comparó el desempeño de una ANN tradicional, con regularización y la PGNN, mediante métricas como MSE, MAE y R², evaluando tanto el desempeño dentro del dominio de entrenamiento como su capacidad predictiva en condiciones no vistas.
Los resultados exponen que el modelo PGNN mejora la capacidad de extrapolación fuera del dominio (R2 = 0.7881), reduciendo el sobreajuste y logrando un mejor balance entre precisión y generalización (R2 = 0.9591 en entrenamiento) en comparación con los modelos basados solo en datos (R2 = 0.9854 en entrenamiento y R2 = -3.5138 fuera del dominio) o con regularización de los pesos de la red neuronal (R2 = 0.9504 en entrenamiento y R2 = 0.1291 fuera del dominio), destacando la utilidad de integrar conocimiento físico en la función de pérdida para optimizar el desempeño del algoritmo. También, se comenta su utilidad frente a otros tipos de modelos, que no son de aprendizaje automático, para disminuir el costo computacional asociado a hacer el modelo y obtener las predicciones.
Finalmente, se discute la proyección de este tipo de esquemas en contextos industriales reales, como en plantas piloto, sistemas de cogeneración y sistemas de captura de carbono acoplados a fuentes de emisión. Como trabajo futuro se plantea explorar modelos PINN más sofisticados, los cuales integran dinámicas temporales del sistema mediante el uso de datos secuenciales.
Este estudio busca así contribuir al avance del modelado de celdas de combustible mediante herramientas basadas en datos y conocimiento físico, sentando las bases para soluciones más eficientes, escalables y sostenibles en la transición hacia tecnologías energéticas limpias.
This study addresses the modeling of a molten carbonate fuel cell (MCFC) through the implementation of machine learning techniques, with an emphasis on hybrid models that incorporate physical knowledge into the training process. MCFCs represent a high-efficiency, low-emission power generation technology due to their ability to operate with various hydrogen-rich fuels and their integration into cogeneration or CO₂ capture systems. However, traditional modeling using physical models often requires detailed system characterization, lengthy development times, and high technical expertise. To address this challenge, the use of artificial neural networks (ANN) is proposed as an alternative for modeling the behavior of an MCFC based on experimental data. In addition, a physics-guided neural network (PGNN) model was implemented, which incorporates part of the system's physics through a modified loss function in the ANN. This approach aims to improve generalization beyond the training domain, overcoming the limitations of purely data-driven models. During the study, different network configurations were explored, and the performance of a traditional ANN, an ANN with regularization, and the PGNN was compared using metrics such as MSE, MAE, and R², evaluating both in-domain performance and predictive capacity under unseen conditions. The results show that the PGNN model improves extrapolation capability outside the training domain (R² = 0.7881), reduces overfitting, and achieves a better balance between accuracy and generalization (R² = 0.9591 in training) compared to purely data-driven models (R² = 0.9854 in training and R² = -3.5138 out of domain) or models with weight regularization (R² = 0.9504 in training and R² = 0.1291 out of domain). These results highlight the value of integrating physical knowledge into the loss function to optimize algorithm performance. The study also discusses its usefulness compared to other non-machine learning models in reducing the computational cost associated with modeling and prediction. Finally, the potential application of such frameworks in real industrial contexts is discussed, such as in pilot plants, cogeneration systems, and carbon capture systems coupled to emission sources. As future work, the study proposes exploring more sophisticated PINN models, which incorporate the system’s temporal dynamics using sequential data. This study thus seeks to contribute to the advancement of fuel cell modeling through data- and physics-based tools, laying the groundwork for more efficient, scalable, and sustainable solutions in the transition toward clean energy technologies.
This study addresses the modeling of a molten carbonate fuel cell (MCFC) through the implementation of machine learning techniques, with an emphasis on hybrid models that incorporate physical knowledge into the training process. MCFCs represent a high-efficiency, low-emission power generation technology due to their ability to operate with various hydrogen-rich fuels and their integration into cogeneration or CO₂ capture systems. However, traditional modeling using physical models often requires detailed system characterization, lengthy development times, and high technical expertise. To address this challenge, the use of artificial neural networks (ANN) is proposed as an alternative for modeling the behavior of an MCFC based on experimental data. In addition, a physics-guided neural network (PGNN) model was implemented, which incorporates part of the system's physics through a modified loss function in the ANN. This approach aims to improve generalization beyond the training domain, overcoming the limitations of purely data-driven models. During the study, different network configurations were explored, and the performance of a traditional ANN, an ANN with regularization, and the PGNN was compared using metrics such as MSE, MAE, and R², evaluating both in-domain performance and predictive capacity under unseen conditions. The results show that the PGNN model improves extrapolation capability outside the training domain (R² = 0.7881), reduces overfitting, and achieves a better balance between accuracy and generalization (R² = 0.9591 in training) compared to purely data-driven models (R² = 0.9854 in training and R² = -3.5138 out of domain) or models with weight regularization (R² = 0.9504 in training and R² = 0.1291 out of domain). These results highlight the value of integrating physical knowledge into the loss function to optimize algorithm performance. The study also discusses its usefulness compared to other non-machine learning models in reducing the computational cost associated with modeling and prediction. Finally, the potential application of such frameworks in real industrial contexts is discussed, such as in pilot plants, cogeneration systems, and carbon capture systems coupled to emission sources. As future work, the study proposes exploring more sophisticated PINN models, which incorporate the system’s temporal dynamics using sequential data. This study thus seeks to contribute to the advancement of fuel cell modeling through data- and physics-based tools, laying the groundwork for more efficient, scalable, and sustainable solutions in the transition toward clean energy technologies.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Mecánico/a.
Keywords
Recursos energéticos, Automatización