.png){kind=icon}
.png){kind=icon}
Return
Optimization of the energy consumption of an induction furnace for steel billets using a reinforcement learning framework applied to a finite element digital twin
.png){kind=icon}
OA
AI
S
L
J
D
C
E
M
DOI：10.1007/s10845-026-02933-0.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Efficient billet heating is a critical economic and environmental challenge in wire-rod steel rolling. This work presents an artificial-intelligence (AI) framework to support energy-consumption reduction in an industrial induction furnace by combining a calibrated finite-element (FE) digital twin, a deep-learning (DL) surrogate, and a deep reinforcement learning agent. The FE model, calibrated with real production data from Global Steel Wire (Spain), reproduces the outlet-temperature trends of the resulfurized billet family with RMSE < 5 °C and Pearson correlation > 0.95. To enable fast surrogate-based optimization, a convolutional-recurrent surrogate was trained using real plant time series as inputs and FE-simulated outlet-temperature profiles for those same real billets as target outputs, achieving RMSE values of 5.7 °C on the training billets and 9.85 °C on the test billets, both within the uncertainty range of the outlet pyrometer. This surrogate was then embedded as the environment of a Deep Q-Learning (DQL) framework that searches for lower-energy heating schedules under practical operational constraints and the prescribed outlet-temperature similarity criterion. Across the evaluated resulfurized billets, the learned policy reduced surrogate-predicted energy consumption in every case while keeping the predicted exit-temperature profiles close to the prescribed 5 °C RMSE similarity threshold to the target profile. In representative cases, the surrogate-predicted savings reached 24% and 27%. These results indicate the potential of reinforcement learning as a supervisory decision-support layer for induction-furnace energy optimization when coupled with a physically grounded and computationally efficient digital environment. However, FE re-evaluation and plant-level validation of the optimized schedules are required before industrial deployment.
Keywords:
Artificial intelligence
Reinforcement learning
Deep q-learning
Convolutional neural networks
Digital twin
Steel rolling
Journal
IF:
7.4
Papers:
3.4K
Citations:
1.1W
Organization
Cited Papers