Abstract: The compressive strength of oxidized pellets is a key indicator for evaluating pellet quality and stability. Accurate prediction of its variation trend is essential for improving production efficiency and optimizing process parameters. However, due to the high dimensionality and strong nonlinearity of compressive strength prediction, existing models still face limitations in terms of reliability, applicability, and generalization. This study proposes the metallurgical-random forest-based Bayesian optimized bidirectional gated recurrent unit (BiGRU) attention prediction model (MRF-BBAPM) model, which employs feature selection guided by metallurgical mechanisms and random forest to enhance model efficiency and relevance. The BiGRU network parameters are optimized using Bayesian optimization, and an attention mechanism is incorporated to focus on critical features, further improving model performance. The SHapley Additive exPlanations (SHAP) method is introduced to quantify the contribution of each feature to the prediction results, revealing the model’s decision-making process and enhancing its interpretability and reliability. The model also incorporates a self-learning mechanism that automatically updates and optimizes itself based on weekly prediction errors. Experimental results show that the proposed model achieves a mean absolute error of 80.58 N (2.77% of the mean) and a root mean square error of 95.75 N (3.29% of the mean) in predicting pellet compressive strength, demonstrating strong stability and reliability in real-world applications. This method provides effective data support for accurate prediction of pellet compressive strength and informed decision-making in production.