A novel mechanism-guided residual network for accurate modelling of scroll expander under noisy and sparse data conditions
Abstract Accurate and practical modelling of the scroll expander is essential to improve energy conversion efficiency and reduce energy losses. To improve the generalizability and physical consistency of the data-driven models for the scroll expander under noisy and data scarcity conditions, a novel...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-08-01
|
| Series: | Complex & Intelligent Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s40747-025-02059-5 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Accurate and practical modelling of the scroll expander is essential to improve energy conversion efficiency and reduce energy losses. To improve the generalizability and physical consistency of the data-driven models for the scroll expander under noisy and data scarcity conditions, a novel mechanism-guided residual network (MGResNet) model is proposed in this study. Firstly, the overall framework of MGResNet is presented. This framework is based on the architecture of residual network, where the mechanistic laws are embedded as constraints in the training of the network through an improved loss function. Then, a hybrid optimization algorithm is detailed, which can achieve efficient and accurate updating of the parameters of the network and mechanistic equations. Finally, comparative prediction experiments are carried out to validate the proposed MGResNet. It has the ability to incorporate mechanistic constraints within the data-driven approach, setting it apart from conventional machine learning and deep learning methods that often disregard underlying physical laws. Experimental results demonstrate that MGResNet significantly outperforms traditional models, achieving over 14.324% improvement in volume flow rate prediction and 3.937% in torque prediction under noisy conditions. Even with a 90% reduction in training data, MGResNet maintains superior accuracy, showing up to 45.983% better performance than other models. This proves that the proposed MGResNet exhibits better forecasting accuracy and stronger robustness in the noisy environments and data sparse conditions due to embedded mechanistic constraints, while generating outputs consistently with physical laws. |
|---|---|
| ISSN: | 2199-4536 2198-6053 |