Torpedo ladle cars play a role in transporting hot metal within steel plants, and optimizing their operation is crucial for reducing energy consumption and CO2 emissions.This study investigates the feasibility of developing a digital twin for simulating the thermal management of torpedo ladle cars at Tata Steel by combining reduced-order models with machine learning techniques. Four distinct digital twin models were developed and evaluated: Random forest combined with singular value decomposition (SVD) (RF-SVDmodel), neural network combined with SVD (NN-SVDmodel), neural network without SVD (NN-model), and recurrent neural network without SVD (RNN-model). These models generalization abilities were evaluated using a learning curve or K-fold cross validation. The models accuracy was
tested using validation datasets of a full cycle prediction of the torpedo ladle car process. The RMSE and R2 of the validation prediction of each model were obtained and compared. Results show that RF-SVDmodel, NN-SVDmodel and RNN-model exhibit some challenges such as potential overfitting and inconsistency in performance. The NN-model is the most promising option with robust performance, high generalization abilities, and competitive accuracy to the AnsysTwinbuilder.