Presentations Advanced Modeling and Building Simulations

Duct network with multiple branch ducts is widely used to distribute conditioned air to some indoor spaces. Reducing flow resistance while assuring the identical flow rate to each branch duct is needed. However, these dual objectives may conflict between each other. This investigation proposed to adopt a two-stage shape optimization to realize the both objectives: in the first stage the flow resistance is minimized by optimizing the geometric shapes without addressing the requirement of identical airflow rate in each duct; in the second stage, the airflow rate in each duct is adjusted. A duct network containing one main duct and four arc-shaped branch ducts are optimally designed. The designed duct network was manufactured by the 3D printing for measurement. The obtained results reveal that the flow resistance was reduced by at least 30% after the design while achieving the nearly identical airflow rate in each branch duct.

Buildings are responsible for one-third of global energy consumption and greenhouse gas emissions. A major part of building energy consumption is related to HVAC systems which aim to provide thermal comfort. Hence, improving the efficiency of HVAC systems using optimization and fault detection techniques is crucial not only to mitigate climate change but also to satisfy occupant needs. To achieve the best performance, an accurate model is often required. Previous studies have highlighted potential limitations focusing on error minimization for model evaluation. This study (i) introduces the concept of reliability as a complementary criterion and (ii) analyzes its importance in building energy modeling using the concept of k-fold cross validation. To this end, models of a water-cooled chiller are developed and compared considering reliability. The results suggest that in addition to training accuracy, reliability is another important factor to consider in model selection and evaluation processes.

Building energy consumption accounts for over 40% of total primary energy use, making accurate thermal models essential for energy efficiency and carbon neutrality in buildings. Traditional physics-based models are complex, while purely data-driven models lack interpretability. To address this, we propose a Physics-informed Neural Network (PINNs) combining a simplified Resistance-Capacitance (RC) grey-box thermal model with a fully connected neural network. The model integrates physical knowledge to enhance the prediction performance. Results show the PINNs model outperforms both pure neural networks and the 2R2C model in three days prediction task, achieving a temperature CVRMSE of 4.42% compared to 9.82% (NNs) and 8.17% (2R2C). This highlights the value of incorporating physical knowledge into data-driven models, offering a generalizable solution for building energy management systems.

A fast building performance simulation method is proposed for evaluating renovation strategies of Dutch terraced houses. It utilizes limited EnergyPlus simulation data generated by data sampling to train surrogate models. Two data sampling approaches (simple random sampling (SRS) and Latin hypercube sampling (LHS)) and three data-driven models (multiple linear regression, extreme gradient boosting, and artificial neural network (ANN)) are investigated. A Dutch terraced house is applied for performance verification. The results show that surrogate models trained with simulation samples from LHS achieve accuracy similar to those trained with samples from SRS. Furthermore, it is found that limited simulation data (400 samples) are sufficient to train an accurate surrogate model. ANN is the most accurate model in simulating indoor air temperatures (R2 = 0.99, MAE = 0.12 ºC) and thermal loads (R2 = 0.95, MAE = 700 W). The computational time of ANN accounts for only 0.03% of that of EnergyPlus.

In the application of computational fluid dynamics (CFD), a mesh is used to describe boundaries of the computational domain and discretize the domain where the governing equations are solved. In built environments, typical flow features such as wall-bounded flows and flows over large flat surfaces can be well modelled with a hexahedral mesh. Besides, under the same grid size, using hexahedral mesh better controls the overall number of grid cells. Nevertheless, in the generation of a hexahedral mesh, connectivity modifications propagate through the mesh, which results in complex spatial division of blocks. Whether using Ansys ICEM CFD or blockMesh, the workload caused by block division is nearly unavoidable. Therefore, this study developed a scripting tool for the fast and batch generation of orthogonal hexahedral grids. The tool can automatically generate blocks for uniform grid by inputting characteristic geometric coordinates, solid coordinates, and mesh scales. It also supports the input of the first layer mesh scale, mesh growth rate, and maximum mesh scale to automatically generate sub-blocks for non-uniform meshes. The definition of a boundary only requires inputting the diagonal coordinates of the surface. Additionally, we innovatively use a staggered addition method that can quickly identify whether the definition of boundary condition is correct and output the reasons and locations of errors. The tool is fully coded and does not require pre-judgment of the relative position of geometric bodies. It is very suitable for cases with many hexahedral bodies in fluid domain such as the office in Figure 1. Under the batch model, a series of geometries and hexahedral meshes could be generated automatically as Figure 2 shows. The exported meshes were tested with commercial software Ansys Fluent and open-source platform OpenFOAM. The tool will be made open source on GitHub.