Scholarlink

<p>Gasification of biomass into gaseous fuels appears to be one of the most promising thermochemical conversion methods; however, predicting the net calorific value (NCV) of gases is rather complicated since it is a complex phenomenon, and there are very few studies with experimental data for such cases. The current paper presents an integrated approach based on the use of a combination of the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and stacked ensemble regression, which can be used to overcome the data problem and increase prediction accuracy. In particular, the authors used 1,200 physical constraints-based synthetic samples generated from a small initial set of nine samples using physics-informed post-processing and target labeling of synthetic data by the teacher. Data quality verification has been carried out using the Kolmogorov–Smirnov test (80% passed, α = 0.05) and correlation MAE of 0.071 at the best generator epoch (7,500 of 8,000). Four regressions have been trained only on synthetic data. The stacked ensemble model scored an R<sup>2</sup> value of 0.979, mean absolute error of 0.021&#xa0;MJ Nm⁻<sup>3</sup>, and root-mean-square error of 0.036&#xa0;MJ Nm⁻<sup>3</sup>, in contrast to the leave-one-out cross-validation baseline of R<sup>2</sup> =  − 0.30 on the original dataset alone. Through permutation feature importance, it was established that fixed carbon (0.371), gasification temperature (0.237), and air surplus ratio (0.235) were the most significant features, in accordance with existing gasification literature. The approach offers a repeatable template for predictive analytics in areas of energy production where data are limited.</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

Enhancing gasification net calorific value prediction using GAN-based synthetic data and stacked learning

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Farzad Zolfaghari,
  • Gyorgyi Kale-Halasz,
  • Arturas Kilikevicius

摘要

Gasification of biomass into gaseous fuels appears to be one of the most promising thermochemical conversion methods; however, predicting the net calorific value (NCV) of gases is rather complicated since it is a complex phenomenon, and there are very few studies with experimental data for such cases. The current paper presents an integrated approach based on the use of a combination of the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and stacked ensemble regression, which can be used to overcome the data problem and increase prediction accuracy. In particular, the authors used 1,200 physical constraints-based synthetic samples generated from a small initial set of nine samples using physics-informed post-processing and target labeling of synthetic data by the teacher. Data quality verification has been carried out using the Kolmogorov–Smirnov test (80% passed, α = 0.05) and correlation MAE of 0.071 at the best generator epoch (7,500 of 8,000). Four regressions have been trained only on synthetic data. The stacked ensemble model scored an R2 value of 0.979, mean absolute error of 0.021 MJ Nm⁻3, and root-mean-square error of 0.036 MJ Nm⁻3, in contrast to the leave-one-out cross-validation baseline of R2 =  − 0.30 on the original dataset alone. Through permutation feature importance, it was established that fixed carbon (0.371), gasification temperature (0.237), and air surplus ratio (0.235) were the most significant features, in accordance with existing gasification literature. The approach offers a repeatable template for predictive analytics in areas of energy production where data are limited.