China Nonferrous Metallurgy

Accurate prediction of the outlet temperature of nickel matte during nickel smelting is crucial for process optimization, equipment safety assurance, and product quality stability. However, the process involves multi-variable nonlinear coupling and dynamic parameter fluctuations, which restrict the application effect of traditional methods. To address this issue, this paper proposes a nickel matte temperature prediction method integrating Temporal Convolutional Network (TCN) with XGBoost dynamic error compensation. First, a fusion strategy of Pearson correlation coefficient and F-score is adopted to select key features from multi-dimensional industrial parameters such as furnace feed composition, blast volume, and oxygen concentration, reducing redundant interference. Second, the dilated causal convolution of TCN is used to capture long-and short-term temporal dependencies and nonlinear relationships among features, generating initial temperature predictions. Furthermore, XGBoost is introduced and first-order/second-order time difference features are constructed to learn the evolution law of prediction residuals for dynamic error compensation. Relying on XGBoost quantile regression, interval prediction of nickel matte temperature can be realized. Experimental results show that in terms of point prediction performance, the TCN-XGBoost model significantly improves accuracy compared with the baseline TCN model: MAE decreases from 7.6277 to 7.4941, MAPE reduces from 0.5928 to 0.5842, RMSE optimizes from 9.7913 to 9.5732, and R2 increases to 0.4516. It also outperforms comparative methods such as LSTM and AGCRN. In terms of interval prediction performance, the 90% prediction interval of TCN-XGBoost exhibits a balanced advantage of “reliable coverage and compact width”. Compared with TCN-LightGBM and TCN-KDE, it can not only tightly wrap the true values in stable temperature intervals but also reliably accommodate changes in true values in sharply fluctuating sections, avoiding problems of excessively wide intervals or missing coverage. This method possesses high-precision prediction capability for the outlet temperature of nickel matte and can effectively adapt to dynamic fluctuations in industrial production, providing scientific support for real-time monitoring and process regulation of the nickel smelting process.