Availability-Aware Historical Feedback Learning for Electric Arc Furnace Temperature and Chemical Composition Prediction
Organizations: Department of Industrial and Management Systems Engineering, West Virginia University, Morgantown, 26506, West Virginia, USA
Abstract
Electric arc furnace (EAF) steelmaking plays an increasingly important role in recycled and lower-carbon steel production. However, variability in charged materials, energy input, and operating conditions makes accurate prediction of molten steel temperature and final chemical composition challenging. Intermittent or delayed measurements limit the information available for timely prediction. This study develops an availability-aware historical-feedback learning framework for predicting EAF endpoint temperature and the final concentrations of twelve chemical elements. Eight industrial data sources are integrated into three progressively enriched predictor representations comprising 19 static, 39 temporal, and 47 history-informed features. Historical outcomes are incorporated only when eligible under an assumed 60-minute measurement-release delay. Using 19,738 temperature heats and 3,341 chemistry heats, twelve tuned regression algorithms are evaluated through chronological forward validation. The proposed historical-feedback representation reduces root mean squared error (RMSE) in 147 of 156 matched algorithm-target comparisons. For temperature prediction, ExtraTrees achieves an RMSE of 12.290 degrees C, representing a 5.54 percent reduction compared with its temporal-only counterpart. Historical-feature ablation and feedback-delay sensitivity analyses demonstrate the importance of recent outcome information and its availability. In a separate retrospective later-period evaluation, lightweight adaptive residual correction reduces temperature RMSE from 12.502 to 12.275 degrees C without retraining the underlying model. Exploratory deep learning comparisons and target-specific feature-attribution analyses provide additional insights into predictive behavior. The results establish historical outcome availability and sequential feedback as important considerations for EAF endpoint prediction.
Figures & tables
| Source file / log | Raw records | Role in analysis |
|---|---|---|
| EAF temperature and oxidation | 85,104 | Temperature outcome timestamp and value; earlier within-heat temperature and oxidation readings available before the task-specific cutoff. |
| EAF transformer | 271,712 | Completed operating intervals; transformer-stage and duration-based descriptors. |
| Carbon injector usage | 4,011,646 | Pre-cutoff carbon-flow summaries and screened cumulative injection information. |
| Basket charged | 137,685 | Pre-cutoff net basket mass, charging-event times, and material-code diversity. |
| EAF added materials | 79,944 | Pre-cutoff net additions, addition-event times, and material-code diversity. |
| Gas/O 2 lance usage | 5,748,194 | Pre-cutoff oxygen and gas-flow summaries, including recent and within-heat temporal descriptors. |
| Group | No. | Feature definitions |
|---|---|---|
| Pre-cutoff temperature / oxidation (STATIC) | 2 | Latest earlier temperature; latest earlier oxidation measurement. |
| Transformer (STATIC) | 4 | Completed interval count; maximum recorded stage; number of distinct stages; union duration of completed intervals. |
| Basket charging (STATIC) | 3 | Net charged mass; number of distinct charging timestamps; number of distinct material codes. |
| Additional materials (STATIC) | 3 | Net added mass; number of distinct addition timestamps; number of distinct material codes. |
| O 2 , gas, and carbon usage (STATIC) | 7 | Time-weighted means and peaks of three flow streams (six); screened carbon counter total (one). |
| STATIC total | 19 | Operational information available before the task-specific cutoff. |
| Element | Available heats | Missing heats | Available (%) |
| C | 3,340 | 1 | 99.97 |
| Si | 3,341 | 0 | 100.00 |
| Mn | 3,341 | 0 | 100.00 |
| P | 3,340 | 1 | 99.97 |
| S | 3,341 | 0 | 100.00 |
| Cu | 3,341 | 0 | 100.00 |
| Algorithm | Tuned hyperparameters and candidate values | Tests |
|---|---|---|
| Ridge | 6 | |
| Lasso | 7 | |
| Elastic Net | ; ratio | 20 |
| Bayesian Ridge | 9 | |
| Decision Tree | Maximum depth ; minimum split ; minimum leaf | 24 |
| KNN | Neighbors ; weights | 12 |
| Experimental component | Implemented configuration |
|---|---|
| Prediction tasks | Tapping temperature (one target) and final steel composition (twelve independently modeled elemental targets). |
| Feature representations | Static (19 predictors), Temporal (39 predictors), and History (47 predictors), constructed using cutoff-eligible information. |
| Chronological data partitioning | First 80% of heats for model development; remaining 20% for retrospective later-period evaluation, independently partitioned for each prediction task. |
| Model validation | Three expanding-window outer chronological folds and three inner chronological folds for hyperparameter optimization; training labels purged according to assumed outcome-release times. |
| Hyperparameter optimization | Twelve regression algorithms; up to 24 candidate configurations per algorithm, target, feature representation, and outer fold; random seed 2026. |
| Baselines and feature comparisons | Outer-training median and last-20-released-outcome median baselines; matched Static , Temporal , and History comparisons; focused historical-feature-group ablation on the third development fold. |
| Algorithm | STATIC RMSE | TEMPORAL RMSE | HISTORY RMSE | HISTORY MAE |
|---|---|---|---|---|
| ExtraTrees | 13.007 | 13.011 | 12.290 | 9.130 |
| Random Forest | 13.103 | 13.079 | 12.295 | 9.186 |
| Gradient Boosting | 13.091 | 13.028 | 12.322 | 9.161 |
| ElasticNet | 13.113 | 13.025 | 12.340 | 9.135 |
| Lasso | 13.106 | 13.012 | 12.340 | 9.135 |
| AdaBoost | 13.301 | 13.272 | 12.349 | 9.245 |
| Element | HISTORY model | RMSE (wt.%) | MAE (wt.%) | RMSE reduction (%) | |
|---|---|---|---|---|---|
| C | RandomForest | 2004 | 0.02464 | 0.01550 | 2.65 |
| Si | ExtraTrees | 2004 | 0.007579 | 0.002867 | 0.31 |
| Mn | XGBoost | 2004 | 0.03403 | 0.02203 | 7.06 |
| P | RandomForest | 2003 | 0.002893 | 0.001837 | 17.10 |
| S | Ridge | 2004 | 0.009383 | 0.006714 | 12.92 |
| Cu | ExtraTrees | 2004 | 0.02320 | 0.01717 | 12.56 |
| Target | Model | A0 | A1 | A2 | A3 | A4 |
|---|---|---|---|---|---|---|
| P | Random Forest | 0.004097 | 0.003802 | 0.003764 | 0.003766 | 0.003716 |
| Mo | Random Forest | 0.038747 | 0.027899 | 0.027888 | 0.028437 | 0.026694 |
| Si | ExtraTrees | 0.010743 | 0.010719 | 0.010695 | 0.010677 | 0.010585 |
| N | Random Forest | 0.029175 | 0.028890 | 0.028847 | 0.028800 | 0.028733 |
| Element | 1 h | 4 h | 24 h | 72 h |
|---|---|---|---|---|
| C | 0.024640 | 0.024860 | 0.024931 | 0.025248 |
| Si | 0.007579 | 0.007572 | 0.007620 | 0.007639 |
| Mn | 0.034028 | 0.035049 | 0.035616 | 0.036105 |
| P | 0.002893 | 0.003055 | 0.003101 | 0.003107 |
| S | 0.009383 | 0.010622 | 0.011749 | 0.010635 |
| Cu | 0.023199 | 0.024782 | 0.025209 | 0.025616 |
| Target | Model | Frozen RMSE | Adaptive RMSE | Frozen MAE | Adaptive MAE | RMSE gain (%) | |
|---|---|---|---|---|---|---|---|
| Temperature | ExtraTrees | 3948 | 12.502 | 12.275 | 9.755 | 9.608 | +1.8 |
| C | RandomForest | 668 | 0.02496 | 0.02496 | 0.01551 | 0.01551 | +0.0 |
| Si | ExtraTrees | 669 | 0.004852 | 0.004843 | 0.003456 | 0.003331 | +0.2 |
| Mn | XGBoost | 669 | 0.03122 | 0.03072 | 0.02016 | 0.01928 | +1.6 |
| P | RandomForest | 669 | 0.002880 | 0.002843 | 0.002205 | 0.002146 | +1.3 |
| S | Ridge | 669 | 0.009620 | 0.009620 | 0.006915 | 0.006915 | +0.0 |
| Target | MLP | LSTM | GRU | 1D CNN |
|---|---|---|---|---|
| Temp. | 12.589 | 12.957 | 12.470 | 12.682 |
| C | 0.02855 | 0.02721 | 0.02657 | 0.02775 |
| Si | 0.007704 | 0.007727 | 0.007783 | 0.007638 |
| Mn | 0.04122 | 0.04015 | 0.03822 | 0.03802 |
| P | 0.003191 | 0.003057 | 0.003188 | 0.003056 |
| S | 0.01058 | 0.01111 | 0.01057 | 0.01016 |
Manufacturing
Manufacturing research develops methods and systems to design, plan, automate, monitor, and improve industrial production.
See all 87 papers