Partial least squares (PLS) and artificial neural networks (ANN) prediction model for four components of feedstuff has been established with good veracity and recurrence. The spectra put into the model should be processed by second derivative and standard normal variate (SNV). Ten principal components compressed from original data by PLS and two peak values were taken as the inputs of Back-Propagation Network (BP), while four predictive targets as outputs, according to Kolmogorov theorem and experiment, and twenty three nerve cells were taken as hidden nodes. Its training iteration times was supposed to be 10,000. Prediction deciding coefficient of four components by the model are 0.9950, 0.9980, 0.9990 and 0.9670, while the standard deviation of an unknown sample scanned parallelly are 0.02774, 0.04853, 0.03292 and 0.02204.