TY - JOUR

T1 - Feasibility Analysis of POD and Deep-Autoencoder for Reduced Order Modelling

T2 - 18th IBPSA Conference on Building Simulation, BS 2023

AU - Wang, Shaofan

AU - Chen, Xia

AU - Geyer, Philipp

N1 - Funding Information: The author greatly acknowledges the EU project TheGreeFa, the DFG researcher unit FOR2363 EarlyBIM and the DFG Heisenberg grant with grant on (for TheGreeFa online, EarlyBIM: GE 1652/3-2, Heisenberg: GE 1652/4-1

PY - 2023

Y1 - 2023

N2 - The computational effort of CFD simulation remains a prohibitive barrier to real-time prediction for design assistance and fast simulation. This study examines and compares the feasibility of two dimensionality reduction algorithms, proper orthogonal decomposition (POD) and deep-autoencoder (AE), for efficient feature extraction from indoor airflow simulation, and investigates how they contribute to fast and accurate prediction. The performance of these two methods is evaluated in various typical scenarios conducted in a 2-dimensional (2D) rectangular chamber, with air velocity distribution as the dependent variable, and tested using a set of different inlet velocities. The research pipeline involves using computational fluid dynamic (CFD) generated data as input for POD and AE to obtain compressed representation, then fed into a neural network (NN) method to train surrogate models for prediction and reconstruction. The results indicate that the combination of AE and neural network method (AE-NN) has excelled in the robust prediction of flow pattern transition than a combination of POD and neural network method (POD-NN). AE-NN accurately predicted that the leading edge of the airflow gradually moved forward as the inlet velocity increased, and the move displacement distances agreed with CFD results.

AB - The computational effort of CFD simulation remains a prohibitive barrier to real-time prediction for design assistance and fast simulation. This study examines and compares the feasibility of two dimensionality reduction algorithms, proper orthogonal decomposition (POD) and deep-autoencoder (AE), for efficient feature extraction from indoor airflow simulation, and investigates how they contribute to fast and accurate prediction. The performance of these two methods is evaluated in various typical scenarios conducted in a 2-dimensional (2D) rectangular chamber, with air velocity distribution as the dependent variable, and tested using a set of different inlet velocities. The research pipeline involves using computational fluid dynamic (CFD) generated data as input for POD and AE to obtain compressed representation, then fed into a neural network (NN) method to train surrogate models for prediction and reconstruction. The results indicate that the combination of AE and neural network method (AE-NN) has excelled in the robust prediction of flow pattern transition than a combination of POD and neural network method (POD-NN). AE-NN accurately predicted that the leading edge of the airflow gradually moved forward as the inlet velocity increased, and the move displacement distances agreed with CFD results.

UR - http://www.scopus.com/inward/record.url?scp=85179514937&partnerID=8YFLogxK

U2 - 10.26868/25222708.2023.1227

DO - 10.26868/25222708.2023.1227

M3 - Conference article

AN - SCOPUS:85179514937

VL - 18

SP - 647

EP - 654

JO - Building Simulation Conference Proceedings

JF - Building Simulation Conference Proceedings

SN - 2522-2708

Y2 - 4 September 2023 through 6 September 2023

ER -