Abstract: Manganese is an important alloying element in iron and steel. Adding the appropriate amount of manganese can enhance the properties of steel. The manganese content directly influences steel quality in the converter steelmaking process. Too little manganese results in insufficient hardness and strength of steel products, whereas excessive manganese leads to increased embrittlement and production costs. Therefore, determining the appropriate amount of manganese is crucial for improving steel quality and reducing smelting costs. The quantity of manganese added during converter steelmaking primarily depends on the predicted final manganese content. However, this content is influenced by various factors, such as the oxidation reaction process and the addition of other alloying elements. These factors exhibit nonlinear effects on the manganese content, and the factors are highly interconnected, making accurate prediction of manganese content at the end point challenging. In response to the challenges posed by noise and strong coupling in predicting manganese content at the end point of converter steelmaking, a research framework was developed to address these issues and facilitate accurate predictions. Key influencing factors in the smelting process were identified through Pearson correlation coefficient analysis and mechanistic analysis. Subsequently, the relationship between these influencing factors and end-point manganese content was modeled using the long short-term memory network (LSTM). To mitigate the effects of high-frequency noise in nonlinear and nonstationary sequences, singular spectral analysis (SSA) was employed during the prediction process. This led to the development of a method known as SSA?LSTM for predicting end-point manganese content. The effects of different test sets and the number of neurons on the prediction results were investigated using converter steelmaking production data from Hebei Jingye Iron & Steel Co., Ltd. The proposed method achieved minimal prediction error when the test set comprised 10% of the data and the number of neurons was set to 85. At these parameters, the mean absolute error of the prediction method for end-point manganese was 1.19%, with a root-mean-square error of 1.48%. These results demonstrate that the proposed method effectively addresses issues related to large noise and nonlinear data. Moreover, compared with existing time series prediction methods, the proposed method, particularly after SSA treatment, showed reduced prediction errors. This validates the effectiveness of the method and provides a basis for accurate alloy addition in actual production processes.