The use of artificial neural network for modeling in vitro rumen methane production using the CNCPS carbohydrate fractions as dietary variables

Abstract

The objective of this trial was to investigate the suitability and accuracy of modeling the rumen methane production of mixed rations for cattle using artificial neural network (ANN). The three layer back propagation neural network (BP) which included the input, the hidden and the output layers, was used for modeling. Two datasets used in the trial were from Dong and Zhao (2013). The first dataset which contained the CH4, CO2 and total gas production and the Cornell Net Carbohydrate and Protein System (CNCPS) carbohydrate fractions of forty-five rations was for training the BP model and the second dataset which contained ten rations was for testing the BP model. The predicting performances of the BP models with different number of neurons in the hidden layer and different number of variables in the output layer were compared, and the effective BP models were established. Paired t-test showed that no difference was found between the observed and the predicted CH4, CO2 and total gas production based on the BP models (p>0.05). Model performance analysis based on the test data showed the root mean square prediction errors (RMSPE%) were 3.89%, 2.95% and 4.23%, and the determination coefficients (r2) between the observed and the predicted values were 0.95, 0.97 and 0.92 for CH4, CO2 and total gas, respectively. Testing of the BP models indicated that the in vitro CH4, CO2 and total gas production of mixed rations for cattle could be reliably and accurately predicted based on the CNCPS carbohydrate fractions using BP models. The BP models showed similar accuracy with the multiple regression model for predicting the CH4 production and better accuracy for predicting the CO2 and the total gas production than the multiple regression models.