Availability-Aware Historical Feedback Learning for Electric Arc Furnace Temperature and Chemical Composition Prediction

Ahmed Shoyeb Raihan, Farzana Islam, Ashish Nimbarte, Imtiaz AhmedarXiv 2026 · arXiv

In the high-stakes world of steelmaking, the Electric Arc Furnace (EAF) is a cornerstone of sustainable production. However, managing an EAF is a bit like cooking a complex meal where the ingredients change every time, and you can only check the temperature every hour. Accurate, real-time predictions of molten steel temperature and chemical composition are vital for quality, yet the data needed for these predictions is often delayed by laboratory processing times.

A new paper, "Availability-Aware Historical Feedback Learning for Electric Arc Furnace Temperature and Chemical Composition Prediction," introduces a framework that bridges this "information-availability gap." By systematically incorporating historical outcomes only when they would realistically be available to a furnace operator, the researchers have achieved significant gains in predictive accuracy.

TL;DR

The researchers developed an "availability-aware" machine learning framework that uses historical heat outcomes—accounting for real-world measurement delays—to predict EAF temperature and chemistry. This approach reduced prediction errors across 147 out of 156 tested scenarios, outperforming both standard temporal models and complex deep learning architectures.

What Problem Does This Paper Address?

EAF steelmaking involves melting scrap metal using high-power electric arcs. Because the "recipe" (the scrap mix) varies wildly and process conditions change rapidly, predicting the final state of the melt is incredibly difficult.

While machine learning (ML) has shown promise in this area, most models suffer from a subtle but critical flaw: they assume that the results of the immediately preceding heat are available to predict the current heat. In reality, chemical assays and temperature readings take time to process and release. Using "future" data that wouldn't be available in a real plant leads to overly optimistic results in the lab that fail in production. This paper tackles the challenge of building models that respect these operational constraints while still extracting maximum value from historical data.

Methods

The authors propose a unified, heat-indexed predictive framework. The core innovation lies in how they structured their data and features:

  1. Cutoff-Aware Engineering: They established a strict "prediction cutoff" five minutes before the actual measurement. Only data recorded before this cutoff is used for prediction.
  2. Nested Feature Representations: They compared three levels of information:
  3. STATIC: 19 basic operational descriptors (e.g., total scrap mass).
  4. TEMPORAL: Adds 20 within-heat summaries (e.g., oxygen flow in the last 10 minutes).
  5. HISTORY: Adds 8 descriptors of previously released outcomes (e.g., the average phosphorus level of the last five heats that have actually been reported by the lab).
  6. Availability Logic: Historical data is only "eligible" if it passed an assumed 60-minute measurement-release delay.
  7. Adaptive Residual Correction: A lightweight "online" layer that adjusts the base model's predictions based on the errors of the most recently released results, allowing the system to adapt to process drift without needing a full model retrain.
Figure 1: Cutoff-aware preprocessing and construction of the two modeling cohorts. Heat-indexed source streams are filtered using task-specific outcome cutoffs, quality-screened, and summarized as operational and within-heat temporal descriptors. Historical outcome features are assembled separately from earlier, assumed-released labels. Final eligible cohorts comprise 19,738 temperature heats and 3,341 chemistry heats; the two cohorts are not required to contain identical heats.

Datasets

The study utilized a massive industrial dataset from a DSP-120 EAF installation, covering three and a half years (January 2015 to July 2018).

  • Temperature Cohort: 19,738 heats.
  • Chemistry Cohort: 3,341 heats (covering 12 elements including Carbon, Phosphorus, Sulfur, and Manganese).
  • Data Sources: Eight heterogeneous logs, including transformer records, basket charging, gas/O2 lance usage, and final chemical assays.
Table 1: Industrial EAF source records and their roles in the cutoff-aware prediction protocol. Source-row counts are those documented in the original data inventory; they are not the numbers of eligible modeling heats.

Key Insights and Ideas

  • The "Freshness" Factor: The most recent released measurement is the most powerful predictor. As the delay in receiving lab results increases, the predictive power of historical feedback decays significantly.
  • Domain Knowledge > Model Complexity: Interestingly, classical ML models (like ExtraTrees and Random Forest) equipped with these carefully engineered historical features generally outperformed deep learning models like LSTMs and GRUs. This suggests that in industrial tabular data, what you feed the model is often more important than the architecture of the model itself.
  • Target-Specific Sensitivity: Not all elements behave the same way. Elements like Molybdenum and Phosphorus showed massive improvements (up to 18.8%) from historical feedback, likely due to persistent "heat-to-heat" patterns in scrap quality or furnace lining state.
  • Real-Time Correction: The adaptive residual correction proved that you don't need to retrain a model to fix its biases. By simply looking at how wrong the model was on the last few reported heats, you can significantly sharpen the next prediction.
Figure 2: Prediction-time and historical-outcome availability assumptions. The task-specific prediction cutoff occurs five minutes before the corresponding recorded outcome measurement. Earlier temperature or composition outcomes become eligible as historical inputs only after the assumed release delay; physical tapping and laboratory reporting times were not independently verified.

Key Results

  • Broad Success: The historical-feedback representation (HISTORY) reduced the Root Mean Squared Error (RMSE) in 147 of 156 matched algorithm-target comparisons.
  • Temperature Accuracy: For tapping temperature, the ExtraTrees algorithm achieved an RMSE of 12.290 °C, a 5.54% reduction over the temporal-only model.
  • Chemistry Gains: Significant RMSE reductions were seen for chemical elements:
  • Molybdenum: 18.80% reduction.
  • Phosphorus: 17.10% reduction.
  • Copper: 12.56% reduction.
  • Adaptive Boost: The residual correction layer further reduced temperature RMSE by 1.82% in a retrospective evaluation, proving its value for handling process drift.

Limitations and Future Work

While the study is robust, the authors note that the 60-minute release delay was an assumed constant rather than an observed variable from the plant's IT system. In a real-world deployment, this delay might fluctuate.

Future research directions include:

  • Integrating the framework with actual, real-time measurement-release systems.
  • Testing the "transferability" of these models across different furnaces and production sites.
  • Exploring how these predictions can be directly integrated into automated control loops for energy and material optimization.

By respecting the clock as much as the chemistry, this research provides a practical blueprint for more reliable, data-driven decision support in the heart of the steel mill.

Summary updated

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.
BibTeX
@misc{raihan2026availability,
  author = {Ahmed Shoyeb Raihan and Farzana Islam and Ashish Nimbarte and Imtiaz Ahmed},
  title = {{Availability-Aware Historical Feedback Learning for Electric Arc Furnace Temperature and Chemical Composition Prediction}},
  year = {2026},
  eprint = {2611.17463},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17463},
}
Figure