Abstract: In the anode refining process, accurate prediction of precious metal production is critical for resource recovery and production optimization. To address the issues of multi-source coupling, noise interference and insufficient frequency-domain characteristics in process data, this paper proposes a multi-channel fusion prediction model based on time series correlation and frequency-domain enhancement, termed MTS-MCF. The Pearson correlation coefficient method was used to screen input variables, and a time series correlation network (CorNet) was constructed to achieve dynamic label propagation. The coupling weights of parameters were quantified using a frequency-domain enhanced channel attention mechanism (FECAM), and the network response was optimized with the ELU activation function. Experimental results show that compared with the benchmark models, the proposed model reduces MAE for gold and silver by more than 5.7%, reduces MSE by more than 6.3%, and increases 
 by more than 6.3%, outperforming mainstream models such as MTS-mixers and providing effective technical support for production control.

Key words:  precious metals, intelligent prediction, attention mechanism, multidimensional time series representation learning, correlation network, feature fusion, deep learning