Abstract: Side-blown smelting is a key process in nickel production, and its operational data serves as the foundation for process modeling and optimization control. In the actual industrial process, there are common problems such as data loss, strong nonlinearity and noise interference, which destroy data integrity and timing continuity, and directly affect the reliability of downstream tasks. However, the production data of side-blown casting furnace has the coupling of time series correlation, and the global and local characteristic information are complex and intertwined, and the existing method of imputation has limitations. Therefore, TCN-SAITS, a hybrid imputation model that fuses time convolutional network (TCN) and self-attention-based imputation for time series (SAITS) model, is proposed to realize data imputation in the side-blowing casting process. The model captures local time series features through TCN, combines the global dependency modeling ability of SAITS, and introduces a joint loss with dynamic time warping to optimize the interpolation accuracy.In the experiments of public datasets and real data of nickel smelting, the proposed method achieves the best imputation effect compared with the conventional data imputation methods.The predictive model trained on data imputed by our method achieved an R² of 0.75, outperforming other imputation methods (0.71) and the un-imputed baseline (0.08), significantly enhancing the reliability of downstream predictions.