Abstract:NH3 is a major harmful gas that affects the growth of broilers in a chicken house. The accurate measurement and prediction of its emissions will help establish environmental regulation model and improve the welfare in the chicken house. The electrochemical sensors are commonly used in the real practice for measuring NH3 concentration, which shows a low accuracy and short life time and makes it difficult to measure NH3 emissions directly. Combined with the mechanism process of NH3 released from manure, CO2 and H2O emissions that are relatively easier and cheaper obtained are selected to predict NH3 emissions. The gaseous emissions from chicken manure housed in a deep litter system were experimentally simulated. The same amount of chicken manure was injected into the experimental setup for multiple days to simulate the daily manure generated in a chicken house, and to monitor the temperature, relative humidity and the emission of CO2, H2O and NH3 from manure. The prediction model for the NH3 emission was developed based on a variety of machine learning methods and environmental parameters. The importance of features and permutation were analyzed to explore the importance of parameters, and the partial dependence graph as well as the individual condition expectation graph were analyzed to explore the dependence of the model on the parameters. Water pressure difference (VPD) was calculated using the temperature and relative humidity and introduced in modeling according to the knowledge of mechanism process of ammonia emissions. Comparisons were made to investigate the influence of different parameters on the optimal model after the introduction of VPD. The model based on extreme random tree showed the best performance in predicting NH3 emissions, with R2 of 0.9167, RMSE of 0.2897mg/(kg·h), and MAPE of 10.82%. The most important parameter in the model was the H2O emission, and the extreme random tree model had the greatest dependence on H2O emission. The introduction of VPD did not improve the prediction ability of extreme random trees. Therefore, the optimal model was the extreme random tree model established based on T,H,EH2O, ECO2 to predict the NH3 emission from broiler manure in a deep litter system.
