A Temporal Network Based on Characterizing and Extracting Time Series in Copper Smelting for Predicting Matte Grade

Area of Science:

• Metallurgical Engineering
• Data Science
• Process Control

Background:

• Traditional matte grade prediction models suffer from low accuracy and poor interpretability.
• These models often rely solely on pre-smelting input and assay data, neglecting valuable real-time process information.
• Copper smelting processes exhibit complex temporal characteristics and uncertain periodic behaviors that challenge existing prediction methods.

Purpose of the Study:

• To develop a novel temporal network for enhanced matte grade prediction in copper smelting.
• To improve prediction accuracy and interpretability by incorporating real-time process sensor data.
• To address the limitations of traditional models in capturing temporal dynamics and periodic information.

Main Methods:

• Utilized the Maximum Information Coefficient (MIC) to select highly correlated temporal process sensor data.
• Employed a Time to Vector (Time2Vec) module to extract periodic information and integrate time series processing.
• Implemented a temporal convolutional network combined with temporal multi-head attention (TCN-TMHA) for feature extraction and weighting.
• Applied specific weighting mechanisms to prioritize key time steps and relevant input features.

Main Results:

• The proposed TCN-TMHA model demonstrated superior performance in predicting copper content.
• Achieved significant improvements in the coefficient of determination (R²) compared to existing matte grade prediction models, ranging from 2.13% to 11.95%.
• Successfully integrated real-time sensor data and temporal analysis for more accurate metallurgical predictions.

Conclusions:

• The novel temporal network effectively addresses the limitations of traditional matte grade prediction models.
• Incorporating Time2Vec and TCN-TMHA modules enhances the model's ability to capture complex temporal dynamics in copper smelting.
• The proposed approach offers a more accurate and potentially more interpretable solution for real-time matte grade prediction.