Detecting the ash content of biomass raw materials was the basis for efficient energy conversion. However, the traditional high-temperature calcination method was time-consuming and costly, while the near-infrared spectroscopy analysis technology could achieve non-destructive, rapid and low-cost qualitative, and quantitative analysis of unknown samples. This study used 1465 biomass raw material samples of 5 locations and 10 types as the research object. The sample set was divided into 9 sample sets by the “screening classification set method” to construct the ash content model of biomass samples by near-infrared spectroscopy. The main results were as follows: the best principal components of corn straw (M), wheat straw + corn straw + cotton straw (WCM), and wheat straw+weeds+garden leaves (WWL) were 5, 6, and 6, respectively. The R2cv of corn straw (M) was 0.975, the R2p of WCM was 0.983, and the model fitting degree was the highest. The RMSE of the set of Changbai+cotton straw (WC) was 0.588 7 and 0.486 4, respectively. The highest ratio of prediction to deviation (RPDcv) of M was 6.3, and the highest ratio of prediction to deviation (RPDp) of WCM was 7.8. The minimum average relative deviation (ARDcv) of maize straw (M) collection was 6%, the minimum average relative deviation (ARDp) of maize straw and WCM collection was 8%, and the RMSECV/RMSEP of wood (W) collection was 1.01. The R2 range of the set model of ash content of 9 biomass samples was 0.753 8～0.979 4, and there was a good linear relationship between the predicted value and the measured value. Among them, H set (R2=0.942 5), M set (R2=0.979 4) and the WCM set (R2=0.978 7) had the best fitting degree and linear relationship. The R2 of the L set (wood scrap) was the lowest, and its value was 0.753 8. The main factor in judge the influence was that the sample contained impurities such as sediment, adhesive, and paint. In order to solve the problem of raw material detection and evaluation of common biomass power plants, 9 biomass ash collection models were used to predict and evaluate the average relative deviation (ARD) of 11 biomass samples. The grass sample model had a good prediction effect (ARD range was 3.7%～16.5%). The “screening classification set method” was used to divide the sample set to establish the near-infrared spectrum biomass ash content model, and its fitting degree, robustness, and accuracy were higher than those of the full sample set model.