• Continuous Conditional GAN
• Deep Learning
• Digital Twin
• Physics Informed Neural Networks

The dynamics of complex systems are often difficult to predict, and the optimization of their behavior is a challenging task mainly due to the high number of variables and their mutual interaction. Modeling such systems frequently requires advanced and dynamic models like Digital Twins. Digital Twins are virtual representations of physical systems that mirror their real-world counterpart in real time. When incorporated within a feedback loop, these predictions can be used to optimize the performance of the system. In this thesis, we aim at developing a Digital Twin of a warehouse store, with the ultimate goal of optimizing energy consumption, reducing environmental impact, and enhancing the overall well-being of workers and clients. Accurate predictions from a digital twin require numerous temperature measurements throughout the store, but placing sufficient sensors is infeasible due to industrial constraints and security concerns. We propose to use a Deep Learning model capable of instantaneously predicting temperature fields based on one or a few input values. As a feasibility study, we apply this model to a simplified room setup, representing the warehouse store. We propose to use a Computational Fluid Dynamics (CFD) simulation to generate a dataset of temperature images, utilizing the Ansys Fluent software. These simulations are considered as the ground truth, and are used to train the model. We employ a Continuous Conditional Generative Adversarial Net (CcGAN) to generate the temperature images. Furthermore, we enhance this model by enforcing the physics underlying the generated temperature fields, introducing the Physics Informed Neural Network (PINN). Here, the governing equation is the energy equation, which describes heat transfer processes. Experimental results demonstrate the feasibility of the proposed approach, and the benefits of properly implementing the physical equation.