Scholarlink

<p>Current efficiency (CE) is a key performance metric in the aluminum electrolysis process, directly reflecting both energy utilization and the operational condition of electrolytic cells. However, accurately predicting CE remains a significant challenge due to the inherently multi-scale, non-stationary nature of auxiliary process variables. To overcome these limitations, we propose a novel Multi-Scale Time-Frequency Domain Network (MSTFNet) specifically designed for robust CE prediction under real-world industrial conditions. It features a dual-branch architecture. The first branch leverages discrete wavelet transform to perform multi-resolution decomposition of input signals into trend and detail components. It introduces a tailored Time-Feat Gated Mixing (T-FGM) mechanism with a self-gated attention unit to mine complex cross-temporal and cross-variable relationships, while the cross-scale convolutional interaction (CSCI) block promotes interaction between the two types of subsequences. The second branch employs a Global Channel Additional Attention (GCAA) module to extract high-level global temporal patterns and refine cross-variable interactions. The key innovations of MSTFNet lie in wavelet-guided multi-scale feature extraction and attention-enhanced temporal fusion, which enable the model to capture intricate dependencies across time and variables. Extensive experiments on real industrial datasets demonstrate that MSTFNet achieves high performance in CE prediction.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MSTFNet: A Multi-scale Time-Frequency Domain Network for Current Efficiency Prediction in Aluminum Electrolysis Process

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Lihui Cen,
  • Jieli Guo,
  • Yuming Wu,
  • Xiaofang Chen

摘要

Current efficiency (CE) is a key performance metric in the aluminum electrolysis process, directly reflecting both energy utilization and the operational condition of electrolytic cells. However, accurately predicting CE remains a significant challenge due to the inherently multi-scale, non-stationary nature of auxiliary process variables. To overcome these limitations, we propose a novel Multi-Scale Time-Frequency Domain Network (MSTFNet) specifically designed for robust CE prediction under real-world industrial conditions. It features a dual-branch architecture. The first branch leverages discrete wavelet transform to perform multi-resolution decomposition of input signals into trend and detail components. It introduces a tailored Time-Feat Gated Mixing (T-FGM) mechanism with a self-gated attention unit to mine complex cross-temporal and cross-variable relationships, while the cross-scale convolutional interaction (CSCI) block promotes interaction between the two types of subsequences. The second branch employs a Global Channel Additional Attention (GCAA) module to extract high-level global temporal patterns and refine cross-variable interactions. The key innovations of MSTFNet lie in wavelet-guided multi-scale feature extraction and attention-enhanced temporal fusion, which enable the model to capture intricate dependencies across time and variables. Extensive experiments on real industrial datasets demonstrate that MSTFNet achieves high performance in CE prediction.