École Nationale Supérieure de
Techniques Avancées

Résumé

Uncertainty quantification (UQ) is a critical process for improving the accuracy and reliability of computational models by identifying and reducing uncertainties. Traditionally, UQ has relied on physics-based models, but recent advancements in machine learning, particularly deep learning, have introduced a data-driven perspective. Deep learning, a subset of machine learning, leverages neural networks with multiple layers to model complex patterns within data. In this study, we present a comprehensive approach to UQ for the Thermal Energy Delivery System (TEDS), a complex thermal system at Idaho National Laboratory (INL). The data will be generated using a digital twin made with Modelica/Dymola. Our methodology involves perturbing three parameters—”shape factor”, porosity, and inlet mass flow rate—within specified boundary conditions, with the tank temperature as the output of interest. The UQ method that will be used is called variance-based sensitivity analysis. The sensitivity analysis is focused on the thermal storage tank, with other components of TEDS, such as the heat exchanger, pipes, and heaters, excluded from consideration. A dataset comprising over 1,000 samples was extracted from the digital twin, and machine learning models were developed and evaluated using Python. Through the pyMAISE library, we identified optimal models with minimal prediction error, ultimately selecting a feed-forward neural network with over 1,300 neurons. This model achieved an R² score exceeding 0.99 and a mean absolute error (MAE) below 1 Kelvin. Subsequent sensitivity analysis revealed that inlet mass flow rate at initial timestamps and porosity exert the most significant influence on the predicted temperatures.