Machine learning-based reduced-order reconstruction method for flow fields
•Design of the ROR model framework based on partial differential operators.•Extraction of low-dimensional flow field features using an autoencoder.•Combination of flow field and spatial features in low-dimensional space using a cross-fitting algorithm.•Elimination of high-dimensional redundancies an...
Gespeichert in:
| Veröffentlicht in: | Energy and buildings Jg. 320; S. 114575 |
|---|---|
| Hauptverfasser: | , , , |
| Format: | Journal Article |
| Sprache: | Englisch |
| Veröffentlicht: |
Elsevier B.V
01.10.2024
|
| Schlagworte: | |
| ISSN: | 0378-7788 |
| Online-Zugang: | Volltext |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Zusammenfassung: | •Design of the ROR model framework based on partial differential operators.•Extraction of low-dimensional flow field features using an autoencoder.•Combination of flow field and spatial features in low-dimensional space using a cross-fitting algorithm.•Elimination of high-dimensional redundancies and noise, reducing dataset quality requirements.•Fast high-fidelity flow field data acquisition without prior physical knowledge.
The real-time prediction of flow fields has scientific and engineering significance, although it is currently challenging. To address this issue, we propose a nonintrusive supervised reduced-order machine learning framework for flow-field reconstruction, referred to as ROR, to achieve real-time flow-field prediction. The model predicts a signed distance function of the domain and uses a typical flow field as feature extraction objects. Utilizing a cross-fit method, it efficiently combines these features, enabling rapid prediction of the full-order flow field. During the model validation phase, we assess the performance of our model by reconstructing steady-state two-dimensional indoor flows in different room layouts. The results indicate that our model accurately predicts the flow field in the target indoor layout within a short timeframe (approximately 5 s) and demonstrates robustness. To delve deeper into the model performance, we discuss the specific parameters of the model framework and test the effectiveness of the flow-field reconstruction under different air supply modes, with the results showing a mean squared error (MSE) of less than 1.5 %. Additionally, we compare our model with the fourier neural operator (FNO) model and find that it exhibited superior performance with the same number of training steps. The outcomes of this study not only bear significant theoretical implications for the field of flow-field prediction but also provide robust support for practical engineering applications. |
|---|---|
| ISSN: | 0378-7788 |
| DOI: | 10.1016/j.enbuild.2024.114575 |