厌氧发酵技术是最具发展前景的有机废弃物资源化利用技术之一, 其研发与利用在国内外都已广泛开展。 在有机废弃物厌氧发酵过程中, 通常采用生化甲烷潜力(BMP)表示物料的厌氧降解能力。 传统BMP的测定方法存在成本高、 耗时长等缺点, 因此提出了利用近红外光谱分析技术快速预测有机废弃物的生化甲烷潜力(BMP), 采用遗传算法(GA)结合支持向量机(SVM)建立函数模型, 对有机废弃物生化产甲烷潜力进行预测。 实验收集了64份水生植物和能源藻类生物质, 样品BMP原始数据通过自行搭建的产甲烷潜力实验平台获得, 同时, 利用傅里叶近红外光谱仪获取样品的近红外光谱数据。 首先, 对光谱数据进行预处理后在全谱区范围内分别建立主成分回归(PCR)、 偏最小二乘法(PLS)和递归指数偏最小二乘法(RPLS)模型, 将原始BMP数据与光谱数据建立关联, 从而实现水生植物和能源藻类BMP的快速预测。 结果表明, 在全谱区上, 递归指数偏最小二乘能够解决传统偏最小二乘法的抗粗差效果差, 易受不良数据影响等问题, 该方法可以提高模型的稳定性, 但响应速度慢、 计算效率低, 在此基础上提出遗传算法(GA)结合支持向量机(SVM)的机器学习方法, 该方法具有良好的全局搜索能力, 适用于小样本情况, 避开了从归纳到演绎的传统过程, 剔除了大量冗余样本信息, 算法简单且具有良好的鲁棒性。 结合近红外光谱频带分配可知, 利用遗传算法(GA)筛选出1 404个波长点, 大致可划分为3个代表性波段, 因此在所选取的波段利用支持向量机建立回归模型。 依据模型评价结果可知, 采用遗传算法和支持向量机所建立的预测模型不仅简化了数据规模, 同时还能提高模型预测精度, 其预测均方根误差(RMSEP)为10.32 mL, 相关决定系数( R²)为0.92, RPD为6.56, 与常规的PLS和RPLS算法建模相比, RMSEP分别减少了19.56和14.81 mL, R²分别提高了0.06和0.04, RPD分别提高了4.31和3.85。 结果表明, 采用GA-SVM算法建模预测有机废弃物生化甲烷潜力的模型准确度较高, 可以代替传统的BMP测定方法, 满足快速检测的需要。

Anaerobic fermentation technology is one of the most promising technologies for the utilization of organic waste resources. Its research and utilization have been widely carried out at home and abroad. Usually, biochemical methane potential (BMP) is used to represent the anaerobic degradation of the material in the anaerobic degradation technology of organic waste. The traditional measuring methods of BMP, are usually expensive and time-consuming. Therefore, near-infrared spectroscopy is proposed to rapid predict the biochemical methane potential (BMP) of organic waste in this paper. And genetic algorithm (GA) combined with support vector machine (SVM) is applied to establish a functional model to predict the biochemical methane potential of organic waste. 64 samples of aquatic plants and algae are collected from the south and east of China. The original BMP data of samples were obtained from the experimental scale digesbers. At the same time, near-infrared spectral data are obtained by Fourier transform near-infrared spectrometer. First of all, the prediction models were developed by the principal component regression, partial least squares, recursive exponential partial least squares (RPLS) on the pre-processed data, respectively. The aim is to connect the original BMP date with the spectral data and realize the rapid prediction of aquatic plants and algae BMP. The results show that the RPLS method on the full spectral can solve the problem of poor robustness and the poor data interference caused by the traditional PLS method. Although this method improves the robustness of the model, it has slow response speed and low computational efficiency. Therefore, we proposed a genetic algorithm (GA) combined with support vector machine (SVM) method, which is suitable for small sample cases, has good global search ability, and also avoids the traditional process from induction to deduction, and eliminates a lot of redundant sample information. In summary, the GA-SVM method is simple, and it has good stability. Combined with the band assignment of the near-infrared spectrum, it could know that the 1 404 characteristic wavelength points were selected, and roughly divided into 3 representative bands by genetic algorithm (GA), so we built the regression model by support vector machines on the selected characteristic bands. According to the results of model evaluation, it is known that the prediction model based on GA-SVM not only simplifies the date scale, but also improves the prediction accuracy. The root mean square error of prediction (RMSEP) is 10.32 mL, the coefficient of determination ( R²) is 0.92; the residual prediction deviation (RPD) is 6.56. Compared with the models PLS and RPLS, the RMSEP was decreased by 19.56 and 14.81 mL respectively; the R² increased by 0.06 and 0.04, the RPD increased by 4.31, 3.85 respectively. The results show that the NIRS model based on GA-SVM can predict the biochemical methane potential of organic waste rapidly and has higher accuracy, it can replace the traditional BMP determination method to meet the needs of rapid detection.