Predictive control of blast furnace temperature in steelmaking with hybrid depth-infused quantum neural networks

TL;DR

The paper addresses the challenge of stabilizing blast furnace temperatures in steelmaking under complex, nonlinear dynamics. It introduces a hybrid quantum–classical neural network pipeline that combines a LSTM-based predictor with a quantum depth-infused (QDI) layer to forecast temperatures from multi-sensor data and to optimize Pulverized Coal Injection (PCI) accordingly. Key contributions include a data-driven feature selection pipeline, development of a temperature predictor and a all-sensors predictor (with quantum augmentation), and a gradient-based PCI optimization strategy that keeps the furnace temperature within $[1500^{\circ}C,1510^{\circ}C]$ and achieves substantial RMSE gains (e.g., RMSE $\approx 4.46$ for the quantum model vs $\approx 8.33$ for the classical model) and stabilization to $\pm 7.6^{\circ}C$. The work demonstrates the practical potential of quantum ML in industrial process control, offering a pathway to energy-efficient, stable steel production through data-driven, quantum-enhanced optimization.

Abstract

Accurate prediction and stabilization of blast furnace temperatures are crucial for optimizing the efficiency and productivity of steel production. Traditional methods often struggle with the complex and non-linear nature of the temperature fluctuations within blast furnaces. This paper proposes a novel approach that combines hybrid quantum machine learning with pulverized coal injection control to address these challenges. By integrating classical machine learning techniques with quantum computing algorithms, we aim to enhance predictive accuracy and achieve more stable temperature control. For this we utilized a unique prediction-based optimization method. Our method leverages quantum-enhanced feature space exploration and the robustness of classical regression models to forecast temperature variations and optimize pulverized coal injection values. Our results demonstrate a significant improvement in prediction accuracy over 25 percent and our solution improved temperature stability to +-7.6 degrees of target range from the earlier variance of +-50 degrees, highlighting the potential of hybrid quantum machine learning models in industrial steel production applications.

Figures (4)

• Figure 1: Optimization of PCI policy and its effect on blast furnace temperature stabilization: the left plot shows a comparison between the original (initial) and optimized PCI injection rates over time, the right plot presents the resulting temperature trends: the optimized policy (green) successfully maintains the temperature within the target operational window (shaded blue area between 1500°C and 1510°C), while the initial policy (blue) leads to significant deviations and instability. Results were obtained using the hybrid machine learning models $\text{M}_\text{all}$ and $\text{M}_\text{optim}$, following the procedure described in Algorithm 1 \ref{['algorithm:T-PCR']}.
• Figure 2: Architecture of the Hybrid Quantum Model for Temperature Prediction in a Blast Furnace: (a) The input consists of time-series sensor data from the blast furnace, including PCI. The model predicts future temperatures based on this data. (b) The core of the model is a Hybrid Quantum Neural Network combining a classical LSTM network with a QDI layer. The LSTM processes temporal sequences, then passes its output through a fully connected layer. The intermediate representation is encoded into quantum states using the QDI layer, which applies rotation and entanglement gates over 6 qubits. A final fully connected layer maps the quantum-encoded features to the predicted future temperature values.
• Figure 3: Schematic representation of the optimization algorithm (Alg. \ref{['algorithm:T-PCR']}). Step 1: predicting all features for 5 hours ahead using $\text{M}_\text{all}$. Step 2: concatenating predicted features with historical data and setting last $m$ PCI values to $0$. Step 3: predicting new PCI values for balancing temperature using $\text{M}_\text{optim}$. Step 4 and 5: predicting optimized temperature values. Step 6: calculating loss with respect to the optimized temperature and PCI values for training $\text{M}_\text{optim}$ weights.
• Figure 4: Comparison of the temperature predictions between the LSTM and Hybrid Quantum LSTM models. (a) Predictions are made for 10 minutes forward (1 timestamp). Hybrid model outperforms the classical one with $\text{RMSE}_\text{q}=7.59$ and $\text{RMSE}_\text{cl}=9.98$. (b). Predictions are made for 50 minutes forward (5 timestamps). Hybrid model outperforms the classical one with $\text{RMSE}_\text{q}=8.55$ and $\text{RMSE}_\text{cl}=9.91$.