基于机器学习的畜禽粪便资源化预测研究

Prediction of livestock and poultry manure utilization based on machine learning

为了比较不同来源的畜禽粪便理化性质差异,考察堆肥过程中与腐熟度有关的理化指标变化情况,采集5种畜禽粪便的原料(堆肥原样)和堆肥腐熟后的样品,对其进行性质测定分析;同时,为探索预测堆肥腐熟度的最优模型,选用XGBoost、随机森林、支持向量机(SVM)、多元非线性回归模型对训练集样本堆肥腐熟度进行预测,并利用测试集样本对比分析4种模型的预测精度。结果表明:羊粪、牛粪、猪粪、兔粪、鸡粪原样均为弱碱性,其浸提液中可溶性盐浓度(EC)较高,鸡粪原样总养分(N、P、K)含量最高,猪粪原样中金属元素铜、锌、铁、锰总含量和有机质含量最高;5种粪便原样经过升温腐熟过程后,铜、锌、铁、锰元素含量和发芽率均表现出明显的上升趋势,pH值、EC值和有机质含量呈显著下降趋势;堆肥有机质含量和含水率对堆肥腐熟度影响最大,%IncMSE值分别为14.92%和13.61%;通过构建XGBoost、随机森林、SVM机器学习模型和多元非线性回归模型,经特征选择模型优化后,仅选取含水率和有机质含量作为特征变量,即可准确地预测堆肥腐熟度,模型预测值与实测值间的拟合优度(R2)分别为0.994、0.871、0.908、0.800。其中,XGBoost模型表现出较高的预测性能,均方根误差和平均绝对误差分别为4.690%和4.042%。相比于模型优化前,XGBoost、随机森林、SVM的R2分别升高41.39%、5.83%和36.30%。由于堆肥腐熟前后pH值、含水率、有机质及发芽率变化显著,且与其他性质之间存在较高的相关性,据此选取其作为堆肥腐熟度的评价指标。综合分析结果认为,XGBoost模型对堆肥腐熟度的预测精度最高,特征选择是提高模型预测精度的有效方法。

In order to explore the utilization value of agricultural waste,investigate the changes of physicochemical indexes related to maturity degree during composting,the original and mature compost samples were collected and analyzed.Meanwhile,in order to explore the optimal models for composting maturity prediction,the XGBoost,RandomForest,Support Vector Machine(SVM)and multiple nonlinear regression models were used to predict the compost maturity of training set samples.The prediction accuracy of four models was compared by using test set samples.Results showed that:the sheep,cow,pig,rabbit,and chicken manures original composts were alkalescence and the soluble salt concentrations were high in leach liquor.The contents of total nutrients(N,P,K)in chicken manure compost were the highest while the contents of copper,zinc,iron,manganese and organic matter in pig manure compost were the highest.The concentrations of copper,zinc,iron,manganese and germination index(GI)showed an obvious uptrend,while the pH,EC and organic matter contents showed a downtrend after maturity process.Organic matter content and water content of compost exerted the greatest influence on compost maturity,and the%IncMSE values were 14.92%and 13.61%,respectively.GI was predicted by constructing the XGBoost,RandomForest,and SVM machine learning model and the multiple regression nonlinear regression model.The water and organic matter content were selected to be feature variables for constructing models to accurately predicted the compost maturity.After the feature selection optimized,the R2 values for XGBoost,RandomForest,SVM and multiple regression nonlinear surface models were 0.994,0.871,0.908,0.800,respectively.The XGBoost exhibited the highest performance with the root-mean-square error and mean absolute error were 4.690%and 4.042%.Compared with unoptimized model,R2 of XGBoost,RandomForest and SVM increased by 41.39%,5.83%and 36.30%,respectively.The pH,water content,organic matter and GI changed significantly before and after maturity process,which were selected to be the evaluation index of compost maturity.Comprehensive analysis results showed that XGBoost model had the highest prediction accuracy for compost maturity,and feature selection was an effective method to improve the prediction accuracy of the model.