Utilization of agricultural residues as compacted fuels (in pellet or briquette form) for both domestic furnace and industrial boiler is more and more promising. Gross calorific value (GCV) is an important performance for biomass as a solid fuel, indicating the useful energy content of biomass. Measurement of gross calorific value using oxygen bomb calorimeter is time consuming. Thus, it is necessary to develop a fast and accurate method to evaluate the GCV of raw biomass residues. This is conducive to control the quality of feedstock for biomass pellets production. In this study, different GCV predicting models for multiple agricultural residues were proposed and analyzed, and optimal modeling and statistical methods for the GCV prediction of 5 crop residues viz. rice straw, wheat straw, corn stalk, rape stalk and cotton stalk were developed. Multiple linear regression (MLR), stepwise regression analysis (SWR), back propagation artificial neural networks (BPNN) models were proposed to predict the GCV from proximate and/or ultimate analysis of 5 crop feedstock. The best coefficients of determination of R2, root mean square error of predict (RMSEP), and the ratio of standard error of prediction to standard deviation of the reference data (RPD) of 0.921 1, 0.135 1 and 3.49 were obtained, respectively, when corresponding variables were introduced to MLR models. Additionally, GCV models developed based on visible-near infrared spectroscopy (Vis-NIR) also showed the highest R2 and RMSEP values of 0.881 2 and 0.412 9, respectively, when partial least squares regression (PLR) was used. This study demonstrates that MLR model and PLR model can be used to estimate the GCV of agricultural feedstock from proximate analysis, ultimate analysis and Vis-NIR technology.