Sintering is an important pretreatment process in blast furnace ironmaking. In the sintering process, accurate prediction of the output of iron ore sinter production is of great significance for optimizing production parameters, improving production efficiency and ensuring product quality. In order to accurately capture the complex correlation between characteristic variables and prevent the error accumulation and propagation in the training process, this paper proposes a multi-step prediction model of finished ore in sintering process based on pre-training framework. Firstly, the Spearman correlation coefficient is used to select the relevant variables to reduce the interference of redundant information. Secondly, the encoder-decoder structure is introduced into the pre-training framework, and the bidirectional long short-term memory network (Bi-LSTM) is used to extract the time series features. Decoder is used to decode and restore the extracted features to reduce the amount of calculation and training time. Finally, in the formal training part, the output of the pre-training stage is merged with the embedding of the stage, and the time and space features are integrated to learn the relationship between the feature variables, so as to realize the multi-step prediction of the large and small sinter output. The experimental results show that the model can effectively reduce the prediction error, provide reliable yield prediction for iron ore sinter production, and promote the optimization of sintering process.