基于机器学习的水质COD预测方法

Water COD prediction based on machine learning

运用紫外光谱进行水质有机污染物浓度(化学耗氧量(COD))的检测,必须建立紫外光谱数据与COD值之间的数学模型.运用机器学习方法中的LM-BP神经网络和支持向量机,建立了紫外多波段光谱数据与COD值的相关性模型,讨论了在LM-BP神经网络建模中网络结构选择、输入数据处理和训练程度控制,以及在支持向量机建模中核函数及其参数选择等问题.对某种水样的紫外多波段光谱,分别运用最小二乘法、LM-BP神经网络、支持向量机的相关性模型进行COD预测.结果表明,2种机器学习方法的预测能力明显优于最小二乘法,能够得到满意的预测精度,为运用物理方法解决化学量测量中普遍存在的相关性问题,提供了实际可行的解决方案.

In order to measure the content of organic pollutants in water, i.e. chemical oxygen demand （COD） by UV spectral analysis, a mathematical model must be established between UV spectral data and COD value. Two machine learning methods, i.e. Levenberg Marquardt-back propagation （LM-BP） neural network and support vector machine （SVM）, were used to build up a correlated model with UV multi- band absorbencies and actual COD values. The main courses of LM-BP neural network modeling, including model structure selecting, data processing and training controlling were discussed, also the selection method of SVM＇s parameters and model representations were studied. The COD value of the sample water was predicted by using three methods of LM-BP neural network, SVM and least-square procedure. The results indicate that the predictive ability of LM-BP neural network or SVM is better than that of least-square procedure. The two machine learning methods provide feasible solutions for the correlation problem widely existing in the chemical measurement using physical method.