为实现高效短程生物脱氮及氨氮和亚硝酸盐氮的快速检测, 采用主成分分析结合BP神经网络的方法建立短程生物脱氮工艺中氨氮和亚硝酸盐氮的近红外光谱定量分析模型(BP神经网络模型)。 工艺运行结果表明: 原水经过好氧阶段氨氮从45.3 mg·L-1下降到2.7 mg·L-1, 亚硝酸盐氮从0.01 mg·L-1上升到19.6 mg·L-1, 硝酸盐氮受到抑制; 在缺氧段亚硝酸盐氮从19.6 mg·L-1下降至1.2 mg·L-1, 系统实现了良好的短程生物脱氮效果。 水样原始光谱主成分分析表明: 前13个主成分代表了原始光谱数据的信息, 其累计贡献率达到95.04%, 排除了冗余信息且大大降低了模型的维数, 光谱数据矩阵从192×2 203减少到192×13, 大大降低了运算量并提高了模型的精度。 BP神经网络模型校正结果显示: BP神经网络模型对氨氮、 亚硝酸盐氮校正时的决定系数(R2)分别达到0.950 4和0.976 2, 校正均方根误差(RMSECV)分别为0.016 6和0.010 9。 BP神经网络模型预测结果显示: BP神经网络模型对氨氮、 亚硝酸盐氮预测输出与期望输出之间的决定系数(R2)分别为0.974 0和0.981 4, 预测均方根误差(RMSEP)分别为0.033 7和0.028 7, 模型预测效果良好。 研究表明, BP神经网络模型可以通过快速测定水样的近红外光谱数据预测短程生物脱氮工艺中氨氮和亚硝酸盐氮浓度, 并根据氨氮和亚硝酸盐氮浓度变化及时、 灵活地控制工艺的运行, 为生物脱氮提供快速有效的检测技术和科学依据。

To achieve efficient nitrogen removal and rapid detection of ammonia nitrogen and nitrite nitrogen, principal component analysis and neural networks were used to establish quantitative analysis model of ammonia nitrogen and nitrite nitrogen in shortcut nitrification and denitrification based on near infrared spectroscopy—BP neural networks model. The results showed that ammonia nitrogen concentration decreased from 45.3 to 2.7 mg·L-1 after aerobic, and nitrite nitrogen concentration increased from 0.01 to 19.6 mg·L-1, while nitrite nitrogen concentration decreased from 19.6 to 1.2 mg·L-1 after anoxic, which means that rapid nitrification and denitrification are successfully achieved. The principal component analysis of the original near infrared spectra for water samples showed the first 13 principal components represented the information of the original spectrum data, with cumulative contribution rate being 95.04%. In this way, redundant information can be eliminated to reduce the number of dimensions in the model. The spectral data matrix is accordingly reduced from 192×2203 to 192×13, which contributes greatly to easier calculations and improves the accuracy of the model. The correction results of BP neural networks model showed the coefficient of determination for ammonia nitrogen and nitrite nitrogen concentration was 0.950 4 and 0.976 2 respectively, with the root mean square error of calibration being 0.016 6 and 0.010 9. BP neural networks model yields predicted values fitting well with the expected values for ammonia nitrogen and nitrite nitrogen concentration, with R2 being 0.974 0 and 0.981 4 respectively, with the root mean square error of prediction being 0.033 7 and 0.028 7, suggesting that BP neural networks model had a good prediction results for ammonia nitrogen and nitrite nitrogen concentration. The study demonstrated that ammonia nitrogen and nitrite nitrogen concentration can be rapidly predicted with BP neural networks based analysis of the near infrared spectroscopy of the water sample in shortcut nitrification and denitrification, which may provide timely and flexible control to shortcut nitrification and denitrification operation according to the ammonia nitrogen and nitrite nitrogen concentration changes, and makes a quick and effective detection technique for denitrification.