Abstract: Precise prediction of alloy yield is essential for quality control and cost reduction in electric arc furnace (EAF) steelmaking. Chromium (Cr) alloy yield is not directly measured and is obtained from post calculations, which makes the target highly sensitive to measurement uncertainty and process disturbances. Industrial production data contain missing values, inconsistent records, extreme heats, and label noise, which reduce the reliability of conventional data-driven models. To address these challenges, this study proposes a robust and noise-aware deep learning framework for Cr yield prediction in EAF steelmaking. The proposed framework integrates consistency-based sample filtering with latent representation learning, and employs a latent-enhanced residual attention network (LERAN) for robust Cr yield prediction. A label-aware denoising autoencoder (LADAE) is developed to reconstruct process features while learning yield-informed latent representations. Based on the disagreement between extreme gradient boosting and LADAE predictions, a consistency filtering strategy is introduced to identify and remove unreliable records. The proposed LERAN captures high-dimensional nonlinear interactions to enhance prediction robustness. Experiments were conducted on 3788 industrial heats collected from an EAF steel plant. The proposed method achieved the best overall performance among all compared models. For Cr yield prediction, the method obtained a mean absolute error of 0.035, a root mean square error of 0.046, a mean absolute percentage error of 7.216%, and a determination coefficient of 0.624. Compared with the best competing model, the mean absolute error, root mean square error, and mean absolute percentage error were reduced by 35.2%, 42.5%, and 35.0%, respectively. The determination coefficient increases from 0.160 to 0.624. The hit rates of proposed model within the error ranges of ±0.03, ±0.04, and ±0.05 achieve 60.55%, 82.41%, and 87.94%, respectively. Ablation analyses further confirmed the contributions of consistency filtering, latent representation learning, and residual-attention regression. These results demonstrate that the proposed framework effectively improve the robustness, prediction accuracy, and engineering applicability of Cr yield prediction under noisy industrial conditions.