基于GSA的厌氧发酵原料碳氮比NIRS快速检测

在以预处理后玉米秸秆、秸秆粪便混合物为原料进行厌氧发酵生产沼气时，为了对厌氧发酵原料碳氮比进行快速检测，将近红外光谱(NIRS)与偏最小二乘(PLS)回归相结合构建快速检测模型，并基于遗传模拟退火算法(GSA)构建遗传模拟退火区间偏最小二乘算法(GSA-iPLS)和双重遗传模拟退火偏最小二乘算法(DGSA-PLS)分别用于特征谱区优选和特征波长点优选，以提高回归模型的检测精度和效率。全谱1844个波长点经GSA-iPLS进行谱区优选后，得到641个波长变量，再经DGSA-PLS进行特征波长点优选后，得到628个波长变量。DGSA-PLS回归模型验证集的决定系数(R2p)为0.920，预测均方根误差为7.178，相对分析误差为3.805。与全谱建模相比，DGSA-PLS模型的RMSEP减小了15.87%。通过波长优选，参与建模的波长点数量显著减少，有效降低了变量维度和模型复杂度，提升了预测精度和预测能力。本文通过优选碳氮比的敏感波长变量，有效提高了预测模型的鲁棒性，为直接、快速、准确测量厌氧发酵原料的碳氮比提供了新途径。

Rapid Determination of C/N Ratio for Anaerobic Fermentation Feedstocks Using Near Infrared Spectroscopy Based on GSA

The essence of anaerobic fermentation (AF) is the cultivation process of microorganisms, and the carbon-nitrogen ratio (C/N) is an important factor that affects the production of biogas. To quickly detect the C/N for the AF feedstocks, such as the pretreated corn stover, the mixture of corn stover and feces, a rapid detection model was constructed based on near infrared spectroscopy (NIRS) combined with partial least squares (PLS) regression. To further improve the detection accuracy and efficiency of the model, the genetic simulated annealing interval partial least squares algorithm (GSA-iPLS) and double genetic simulated annealing partial least squares algorithm (DGSA-PLS) based on genetic simulated annealing algorithm (GSA) were proposed for selecting the efficient spectral regions and characteristic wavelength points for NIRS, respectively. Totally 1844 wavelength points of the whole spectrum were selected by GSA-iPLS, and 641 wavelength variables were obtained, and 628 wavelength variables were obtained after the characteristic wavelength points were optimized by DGSA-PLS. The coefficients of multiple determination for prediction (R2p), root mean squared error of prediction (RMSEP) and residual predictive deviation (RPD) in DGSA-PLS regressive model were 0.920, 7.178 and 3.805, respectively. Compared with the whole spectrum model, the RMSEP was decreased by 15.87% in the DGSA-PLS model. It was shown that the number of wavelengths was significantly decreased after the optimization, and the performance of regressive model was obviously higher than that of the whole wavelengths. The research improved the adaptability of the prediction model based on optimizing sensitive wavelength variables for C/N, which provided a new way for directly rapid and accurate measurement of the C/N of AF feedstock.