基于机器学习的地下水水质预测研究

基于实测的地下水水质数据（pH、总硬度、溶解性总固体、硫酸盐、氯化物、Fe、Mn 7种）和气象数据（平均气温、最低气温、最高气温、平均最低气温、平均最高气温、20:00—20:00降水量、日降水量≥0.1 mm的时间、最大日降水量8种），分别使用BP神经网络、随机森林（RF）和支持向量机（SVM）构建了地下水水质参数的机器学习预测模型．对于每一种水质参数，分别使用不同的机器学习算法基于不同滞后期的数据进行模拟，将结果与实测水质进行对比，选择精度最高的机器学习模型及其对应的滞后期作为该水质参数的最优模型和最佳滞后期．结果表明，不同机器学习方法和滞后期的选择对预测精度影响很大，BP神经网络对pH（R2=0.225，RMSE为2.411）、总硬度（R2=0.503，RMSE为47.973 mg·L?1）、氯化物（R2=0.994，RMSE为0.544 mg·L?1）和Fe（R2=0.302，RMSE为7.772 mg·L?1）的预测精度最高，RF对硫酸盐（R2=0.908，RMSE为3.788 mg·L?1）和Mn（R2=0.522，RMSE为0.429 mg·L?1）的预测精度最高，BP神经网络、RF和SVM对溶解性总固体的预测性能均较好（R2=0.994~0.996，RMSE为674.660~950.470 mg·L?1）．此外，硫酸盐和Mn预测模型对应的最佳滞后期为0个月，溶解性总固体和氯化物预测模型对应的最佳滞后期为1个月，pH、总硬度和Fe预测模型对应的最佳滞后期为2个月． 

Machine learning to predict groundwater quality

Groundwater quality data (pH, total hardness, total dissolved solids, sulfate, chloride, iron and manganese) and meteorological data (average temperature, minimum temperature, maximum temperature, average minimum temperature, average maximum temperature, daily (20:00-20:00) precipitation, daily precipitation ≥ 0.1 mm days, maximum daily precipitation) were subject to analysis by machine learning models, using BP neural network, random forest and support vector mechanism．For each groundwater quality parameter, different machine learning algorithms were used to simulate data in different lag phases, results were then compared with measured groundwater quality parameters．Machine learning model with highest accuracy and corresponding lag phase were selected as the optimal model．Different machine learning methods and choice of lag phase were found to have great influence on prediction accuracy．BP neural network showed the highest prediction accuracy for pH (R2 = 0.225, RMSE is 2.411), total hardness (R2 = 0.503, RMSE is 47.973 mg·L?1), chloride (R2 = 0.994, RMSE is 0.544 mg·L?1) and iron (R2 = 0.302, RMSE is 7.772 mg·L?1)．RF showed the highest prediction accuracy for sulfate (R2 = 0.908, RMSE is 3.788 mg·L?1) and Manganese (R2 = 0.522, RMSE is 0.429 mg·L?1)．All methods used showed good predictive performance for total dissolved solids (R2 = 0.994-0.996, RMSE is 674.660-950.470 mg·L?1)．The best lag phase of sulfate and Manganese monitoring model was 0 month, the best lag phase of chloride monitoring model was 1 month, the best lag phase of pH, dissolved total solids and total hardness monitoring model was 2 months．