Abstract: The FeO content is an important indicator affecting the strength and metallurgical performance of sintered ore. Therefore, accurately predicting the FeO content of sintered ore is helpful for stabilizing the sintering quality and ensuring the smooth operation of the blast furnace ironmaking process. As a typical sequential production, the FeO content prediction method based on recurrent neural networks has been widely applied. However, traditional recurrent neural networks and their variants face the problem of information dilution when dealing with long sequences, and the information from distant sequence ends has a relatively weak impact on the final output. Therefore, this paper constructed the BiGRU-Attention model. By analyzing the characteristics of on-site data, it used the isolation forest to monitor outliers and eliminate them, the missing forest interpolation model to fill in the missing values, and finally performed data transformation to improve the quality of the input data of the model. In addition, the recursive feature elimination algorithm was used for feature reduction, reducing the model load. By constructing a bidirectional gated recurrent unit and processing the forward and backward time series in parallel, the long-term temporal dependencies in the sintering process are captured, achieving the complete modeling of the global context information of the time series. An attention mechanism was introduced to dynamically focus on key stages. Through the Query matrix,Key matrix and Value matrix of the attention layer, the importance of each part was calculated, adaptively allocating weights for the input data, generating a more relevant context matrix, optimizing resource allocation, and improving the prediction accuracy of the model on complex time series data. The results show that compared with GRU and BiGRU models, the BiGRU-Attention model can better understand the intrinsic correlations between sintering features. This model shows higher prediction accuracy and flexibility on the test set, with M_AE = 0.033 7, R_MSE = 0.103 6, M_APE=4.054%, R²=0.893 5. The prediction results of the BiGRU-Attention model are highly consistent with the true values. Furthermore, in order to test the industrial application value of the model, by incorporating sliding windows and masks, and introducing different levels of random noise during the training and prediction of the model, the robustness and application potential of this method in accurately controlling the trend changes of FeO content in sintered ore were verified.