有色金属(冶炼部分)

电积是目前最为常用的铜电解液净化工艺，其出口铜离子浓度波动大、人工调控难度高，易造成后续硫化单元处理负荷剧增及铜砷共沉淀产废量增大，而传统预测模型存在不可解释、稳态限制、低泛化能力等缺陷。为此，构建了企业电积生产过程电压及出液铜离子浓度准确预测的多参数模型。通过对比研究10种机器学习模型，发现GBR在电压预测中表现最优(决定系数R²=0.79，均方误差MSE=1.25),XGBoost对出液铜离子浓度的预测准确度最高(R²=0.87,MSE=5.58)。SHAP解释性分析表明，电流和时间分别是影响电压和出液铜离子浓度变化的主控因素。模型决策机制与电化学原理及质量守恒定律一致，突破了传统模型对非线性关系的表征局限，为异常工况的预警诊断、关键参数动态优化控制及减少污染物产生提供依据。

In recent years, the growing demand for copper concentrate raw materials and the tight supply in the raw material procurement market have led to increasing complexity in the raw material structure. Higher impurity contents in anode plates result in greater dissolution amounts in the electrolyte. Therefore, according to the composition of the electrolyte, a certain amount of electrolyte must be extracted from the electrolyte circulation system daily for purification to maintain the concentration stability and volume balance of copper and other components in the electrolyte. Electrowinning is currently the most commonly used purification method. However, the copper electrolytic refining and electrowinning process currently faces issues such as significant fluctuations in effluent copper concentration, increased processing loads on subsequent sulfidation units due to difficult manual regulation, and substantial generation of copper-arsenic coprecipitation waste. Meanwhile, traditional prediction models for copper electrowinning suffer from defects such as lack of interpretability, steady-state limitations, and low generalization ability. To address the above problems, this study collects data from the enterprise's secondstage electrowinning production process. After pretreatment(cleaning, filling, and normalization), time, copper concentration in the electrowinning inlet solution, electrolyte temperature, current intensity, and flow rate were used as input variables to establish models for predicting electrowinning voltage and copper concentration in the electrowinning effluent solution. Ten machine learning models were constructed based on default hyperparameters, and their performances were systematically evaluated through multi-dimensional indicators including the coefficient of determination(R²), mean squared error(MSE), and weighted mean absolute percentage error(wMAPE). Through time-series characteristic analysis of key parameters, it is found that each parameter exhibits significantly different time-series fluctuation characteristics, and correlation analysis is used to validate the rationality of the model feature screening results. Subsequently, the Gradient Boosting Regression(GBR) model(R²=0.72, MSE=1.51, wMAPE=2.78%) and eXtreme Gradient Boosting(XGBoost) model(R²=0.73, MSE=8.91), which performs excellently in the initial screening, are further optimized using Bayesian optimization and grid search. After hyperparameter optimization, GBR is found to be suitable for voltage prediction with its stable linear fitting advantages(R²=0.79, MSE=1.25), while XGBoost demonstrates excellent performance in concentration prediction by strengthening nonlinear feature capture(R²=0.87, MSE=5.58). Shapley Additive exPlanations(SHAP) interpretability analysis shows that current intensity is the main factor affecting voltage, with a strong positive correlation between it and voltage. Time is identified as the key factor influencing effluent copper concentration, where high-value time exhibits statistically significant positive contributions to the model's predicted effluent copper concentration. The SHAP values of inlet copper concentration, temperature, and flow rate are distributed close to zero with positive and negative fluctuations, indicating their low impact and dependence on feature interactions. The decision-making mechanism of the models is consistent with electrochemical principles and the law of mass conservation. The established models outperform traditional mechanistic and soft-sensing models in capturing the nonlinear relationships and complex couplings in the electrowinning process, providing a basis for the early warning and diagnosis of abnormal working conditions, the optimized control of dynamic working conditions and the reduction of pollutants generation.