Indoor Temperature Field Modeling Based on Neural Network without Real Data

Accurately capturing and reconstructing the indoor temperature field is the key to meet the requirements of indoor environmental comfort and energy saving. Traditional computational fluid dynamics (CFD) methods and surrogate models are time-consuming and have challenges in simulating inlet and outlet boundary conditions and initial conditions, which may lead to inaccurate results. Artificial neural networks (ANNs) can reconstruct airflow fields directly from measurements, eliminating the need for large and potentially inaccurate CFD databases. However, because ANNs are purely data-driven, they can sometimes produce unphysical results. Incorporating physical laws into neural networks can address this issue. In this study, physical laws are integrated into ANN s. To improve the accuracy of ANN predictions, boundary conditions are strictly enforced using a hard constraint method, ensuring that the edge points comply with physical laws. This paper sets up a virtual indoor environment to determine the inlet and outlet positions as well as the fluid flow rate. The loss function is divided into the mean square error (MSE) of data and the losses of physical information calculated by different components. The experimental results indicate that, compared to the traditional CFD method, the neural network has higher prediction accuracy and faster calculation speed under hard constraint.