.png){kind=icon}
.png){kind=icon}
Return
Dimensionality reduction for steel rolling reheating furnace energy consumption prediction: feature importance, model performance, and industrial case studies
.png){kind=icon}
PRE
AI
Y
G
X
L
X
DOI：10.1007/s42243-026-01856-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As a core high-energy-consuming unit in the steel industry, the steel rolling reheating furnace (SRRF) presents a critical technical challenge in achieving energy savings and carbon emission reduction through accurate energy consumption prediction. The multivariable coupling and high-dimensional nonlinear characteristics of SRRF operational data were addressed by systematically evaluating four feature selection methods: principal component analysis (PCA), process-driven feature engineering, Spearman correlation filtering, and random forest (RF) feature importance. These methods were integrated with two machine learning algorithms, RF and gradient boosting regression tree (GBRT), to develop predictive models for specific energy consumption. Based on 2000 real production data samples collected from a steel plant, a high-quality dataset of 1251 samples was constructed using a combined boxplot–process threshold filtering strategy. To eliminate dimensional inconsistencies, all features were normalized using the min–max scaling method. Experimental results show that using the top nine features identified by RF importance ranking (accounting for approximately 50% of total features), the GBRT model achieved optimal prediction performance, with a test set root mean square error (RMSE) of 0.1032 GJ/t and a coefficient of determination (R2) of 0.8830, while reducing the training time by 33.1% compared to the model trained on the full feature set. In contrast, while PCA improved computational efficiency by 42.3%, it also increased the RMSE by 25.8%. The correlation-based method and process-driven feature construction exhibited context sensitivity and local optimization effects, respectively. The findings demonstrate that the model-driven feature selection strategy based on RF importance, combined with ensemble learning GBRT, effectively balances dimensionality reduction and information retention, offering an interpretable and efficient solution for intelligent prediction and control of the SRRF.
Keywords:
Steel rolling reheating furnace
Specific energy consumption
Dimensionality reduction
Random forest
Gradient boosting regression tree
Journal
IF:
3.6
Papers:
3.6K
Citations:
6.1K
Organization
Cited Papers