Transfer-Learning-Enabled LSTM-Attention for Thyristor Temperature Prediction in UHVDC Converter Valves
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Accurate temperature prediction of converter-valve thyristors is critical for ensuring the thermal stability and safe operation of ultra-high-voltage direct-current (UHVDC) transmission systems. Direct temperature measurement is often constrained by high installation cost, limited sensor deployment, and difficulties in continuous online monitoring. To address these issues, this paper proposes a data-driven thyristor temperature prediction method based on a long short-term memory network integrated with a soft attention mechanism (LSTM-Attention). Valve current, coolant flow rate, and valve-hall ambient temperature are selected as key input features to characterize the dominant thermal dynamics of liquid-cooled converter valves. A semi-physical thermal simulation model is developed to generate long-term multivariate time-series data under varying load and cooling conditions. The proposed model is evaluated against a standard LSTM, and a transfer learning strategy is further introduced to enable rapid adaptation to valve-specific conditions using limited samples. Experimental results show that the LSTM-Attention model achieves lower prediction errors and reduced lag during transient heating and cooling processes. The fine-tuned model also exhibits improved generalization performance under distribution shifts, demonstrating its effectiveness and practicality for online thermal monitoring in UHVDC systems.
Fields of Study & Subjects
HVDC Systems and Fault Protection
Thermal Analysis in Power Transmission
Silicon Carbide Semiconductor Technologies
Publication details
- Journal / source
- International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
- Pages
- 141–147
- Publication date
- 16 January 2026
- Language
- EN