Scholarlink

<p>In the industrial fuel ethanol fermentation process, accurate prediction of out-of-tank ethanol concentration, i.e., its main performance, is essential for its control and optimization. However, ethanol fermentation is a long-period batch process, whose performance is affected by multiple static variables, i.e., initial tank entry information, and dynamic variables, including time-series operational variables and compositional variables sampled at different time points. Existing prediction models are usually based on static variables and dynamic variables. However, dynamic variables generally require sampling over a period of 0–40&#xa0;h, resulting in delays in the timely prediction of fermentation performance. To address this challenge, this paper proposes a novel model called the Temporal Fusion Enhanced Network (TFEN). First, the enhancement module extracts temporally fused features from historical data, including static variables at the beginning of production, low-frequency dynamic variables sampled at different hours, and high-frequency dynamic variables. Second, the reconstruction module learns the hidden relationship between temporally fused features and static variables. Third, static variables from the dataset are passed through the reconstruction module to generate new features, which are used to train the model. Finally, TFEN can accurately predict the ethanol fermentation performance only based on static variables at the beginning of production. Comparative experiments indicate that the predictive performance of TFEN not only surpasses that of other algorithms, but also exceeds the results of the model employing static variables and dynamic variables at 8&#xa0;h and 24&#xa0;h as the input variables. Moreover, its accuracy is close to the model with other dynamic variables at 40&#xa0;h. Also, the more comprehensive the historical data is, the more accurate TFEN is. The proposed TFEN achieves the best prediction accuracy with an RMSE of 0.0526 (normalized), corresponding to 0.339&#xa0;g/100&#xa0;mL in physical units.</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

Temporal Fusion Enhanced Network for Early Prediction of Ethanol Fermentation Performance Using Initial Tank Information

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Yichao Xia,
  • Yifei Sun,
  • Yufeng Dong,
  • Xuefeng Yan

摘要

In the industrial fuel ethanol fermentation process, accurate prediction of out-of-tank ethanol concentration, i.e., its main performance, is essential for its control and optimization. However, ethanol fermentation is a long-period batch process, whose performance is affected by multiple static variables, i.e., initial tank entry information, and dynamic variables, including time-series operational variables and compositional variables sampled at different time points. Existing prediction models are usually based on static variables and dynamic variables. However, dynamic variables generally require sampling over a period of 0–40 h, resulting in delays in the timely prediction of fermentation performance. To address this challenge, this paper proposes a novel model called the Temporal Fusion Enhanced Network (TFEN). First, the enhancement module extracts temporally fused features from historical data, including static variables at the beginning of production, low-frequency dynamic variables sampled at different hours, and high-frequency dynamic variables. Second, the reconstruction module learns the hidden relationship between temporally fused features and static variables. Third, static variables from the dataset are passed through the reconstruction module to generate new features, which are used to train the model. Finally, TFEN can accurately predict the ethanol fermentation performance only based on static variables at the beginning of production. Comparative experiments indicate that the predictive performance of TFEN not only surpasses that of other algorithms, but also exceeds the results of the model employing static variables and dynamic variables at 8 h and 24 h as the input variables. Moreover, its accuracy is close to the model with other dynamic variables at 40 h. Also, the more comprehensive the historical data is, the more accurate TFEN is. The proposed TFEN achieves the best prediction accuracy with an RMSE of 0.0526 (normalized), corresponding to 0.339 g/100 mL in physical units.