基于机器学习的微细矿物沉降上清液浊度预测研究

    Study on prediction of supernatant turbidity in fine mineral settling based on machine learning

    • 摘要: 为了提升煤泥水沉降效率,解决人工加药依赖经验导致的药剂浪费及环境成本高的问题,实现上清液浊度的精准预测,提出了基于机器学习的微细矿物沉降上清液浊度预测方法。研究以石英和蒙脱石沉降试验后的上清液浊度为研究对象,选用氯化镁、氯化钙、明矾、聚合氯化铝和石膏作为凝聚剂,构建支持向量回归、决策树、K近邻、随机森林、线性回归和多层感知机6种机器学习预测模型,并结合MSERMSEMAEMAPER2指标来评估模型性能,探究了凝聚剂种类和矿浆浓度对上清液浊度的影响,并对6种模型的拟合效果进行了评价。结果表明:石英和蒙脱石矿浆的沉降效果受不同凝聚剂影响存在差异,相同凝聚剂对不同浓度矿浆沉降趋势亦存在不同影响,二者的最优凝聚剂均为明矾;将机器学习算法的预测结果与试验结果进行对比,模型拟合精度由高到低的顺序为多层感知机、线性回归、随机森林、K近邻、决策树、支持向量回归模型;多层感知机模型在预测石英、蒙脱石等微细矿物的上清液浊度方面具有高精度,该模型的R2值达到0.99,且相较于线性回归模型,RMSEMAE分别降低了89.02%和87.25%。研究结果可为煤泥水处理中的智能加药系统开发提供理论指导和技术支持。

       

      Abstract: In order to improve the settling efficiency of coal slurry water, tackle the problems of agent waste and high environmental costs caused by manual dosage relying on experience, and realize the accurate prediction of supernatant turbidity, a machine learning based method for predicting the turbidity of fine mineral settling supernatant is proposed. This study focuses on the supernatant turbidity after settling tests of quartz and montmorillonite, selecting magnesium chloride, calcium chloride, alum, polyaluminum chloride, and gypsum as coagulants, six machine learning prediction models including support vector regression, decision tree, k-nearest neighbors, random forest, linear regression, and multi-layer perceptron are constructed, the model performance is evaluated by MSE, RMSE, MAE, MAPE, R2 indexes, the effects of coagulant types and pulp concentration on the turbidity of supernatant are studied and the fitting effect of these models is evaluated. The results show that the settling effects of quartz and montmorillonite pulps vary with different coagulants, and the same coagulant also has different effects on the settling trends of pulps with different concentrations, the optimal coagulant for both is alum; comparing the prediction results of machine learning algorithm with the experimental results, the sequence of model fitting accuracy from high to low is multilayer perceptron, linear regression, random forest, k-nearest neighbor, decision tree and support vector regression model; multi-layer perceptron model has high accuracy in predicting the supernatant turbidity of quartz, montmorillonite and other fine minerals, the R2 value of the model reaches 0.99, and compare with the linear regression model, RMSE and MAE are reduced by 89.02% and 87.25%, respectively. The results can provide theoretical guidance and technical support for the development of intelligent dosing system in coal slurry water treatment.

       

    /

    返回文章
    返回