精细全光谱结合GS-SVR的复杂水体硝酸盐分析方法研究

硝酸盐是水质健康状态评价的一个关键要素。水体中高浓度的硝酸盐会导致生物多样性剧减以及生态系统的退化，同时对人类的健康产生不可逆转的伤害。基于光学测量的水质在线监测是当前及未来水环境动态监测的发展趋势。相较于传统硝酸盐现场采样加实验室分析的测定方法，具有操作便捷，无需前处理，检测效率高，可靠性好且无污染等显著优点。由于实际水体组分的复杂性与多样性，水体参数和吸光度二者并非呈现线性相关，传统的单波长法，双波长法，偏最小二乘法等线性回归预测模型已不适用。基于此，提出一种精细全光谱结合可变步长网格搜索，优化支持向量回归(GS-SVR)的水体硝酸盐分析方法。同时与陕西科技大学化学与化工学院合作，采用标准的硝酸盐溶液，铂-钴标准溶液，福尔马肼标准混悬液根据实验要求配制了不同浓度梯度94组溶液样本。首先将采集到的透射率光谱数据完成吸光度转换，并使用Kennard-Stone方法将94个溶液样本划分为80个训练集和14个测试集。其次使用改进的GS算法结合交叉验证，通过多次迭代，减小搜索范围、改变搜索步长对SVR进行参数寻优，并将最优惩罚参数C和核函数宽度σ用于训练集中进行模型建立，最后用所建立的模型对测试集进行浓度预测。并将预测效果与反向传播神经网络(BPNN)，SVR，GS-SVR，粒子群算法优化SVR(PSO-SVR)，遗传算法优化SVR(GA-SVR)的模型预测结果比较，结果显示，提出的算法模型相关系数R2=0.993 5，预测均方根误差RMSEP=0.043 5，最优参数C和σ组合为(512，0.044 2)，平均训练时间为13 s。相较于上述五种预测模型，R2分别提高了1.22%，11.66%，0.78%，0.74%和0.77%，训练效率分别提升4.15倍(BPNN)，8.30倍(GS-SVR)，21.38倍(PSO-SVR)，10.23倍(GA-SVR)。模型的预测精度以及训练效率方面都取得了很大的提升，为复杂水体硝酸盐浓度的快速实时在线监测提供了一种新的方法。同时，该方法具备一定的普适性，也适用于其他水质参数预测模型的建立。

Research on the Quantitative Analysis Method of Nitrate in Complex Water by Full Scale Spectrum With GS-SVR

Nitrate is an important index of water quality monitoring.The high concentration of nitrate in water results in the decrease of biodiversity and the degradation of the ecosystem.Meanwhile,it will cause irreversible harm to human health.Water quality monitoring technology based on the spectrum is the trend of modern water environment monitoring.Compared with the traditional method,nitrate field sampling and laboratory analysis,it has the advantages of simple operation,no pretreatment,fast detection,good repeatability and no pollution.Due to the complexity and diversity of water components,there is a high degree of nonlinearity between water parameters and absorbance.Traditional linear regression prediction models are not applicable,such as single wavelength method,dual wavelength method and partial least square method.Therefore,this paper proposes a new method for the determination of nitrate in water by fine full spectrum combined with the improved variable step grid search algorithm optimized support vector regression(GS-SVR).In cooperation with the college of chemistry and chemical engineering of Shaanxi University of Science and Technology,94 groups of solution samples with different concentrations were prepared according to different concentration gradients and the experimental requirements by using standard nitrate solution,platinum cobalt standard solution and formazine standard suspension.Firstly,the transmittance spectrum was converted to absorbance,and 94 solution samples were divided into 80 training sets and 14 test sets by Kennard stone algorithm.Secondly,the improved GS algorithm combined with 5-fold cross validation is used to optimize the parameters of SVR by reducing the search range and changing the search step for many times,and the optimal penalty parameters and kernel function width are used to build model,which is used to predict the test set.Meanwhile,the prediction results are compared with those of BPNN,SVR,GS-SVR,PSO-SVR and GA-SVR.The results show that the coefficient of determination R 2=0.9935,root means square of prediction RMSEP=0.0435.The optimal parameters are(512,0.0442),and the average training time is 13 s.Compared with the above five prediction models,R 2 increased by 1.22%,11.66%,0.78%,0.74%,0.77%,training efficiency increased by 4.15 times(BPNN),8.30 times(GS-SVR),21.38 times(PSO-SVR),10.23 times(GA-SVR).The prediction accuracy and training efficiency of the model has been greatly improved,which provides a novel approach basis for rapid and real-time online monitoring of nitrate concentration in the complex water body.This method is also suitable for the establishment prediction models of other water quality parameters.