Objective: In an era of big data, machine learning algorithms are more widely used in diabetes. Based on the application and evaluation of GM (1,1) model in blood glucose prediction, the difference of the predictive ability of this model in different blood glucose ranges was evaluated to guide the more reasonable application. Methods: GM (1,1) was modeled based on the metabolism algorithm and predicted blood glucose after 5 min, 15 min, and 30 min. Standard filtering based on smooth prediction resulted in a glycemic range (denoted as a ~ b) and a range of LAGE (denoted as LAGE ≤ c), by which blood glucose was divided into different ranges. Group A was defined as the group less a value, group B as the group in a ~b, and group C as the group greater than $b$ value. Group D was defined as less than and equal to $c$ value, and group E was defined as greater than $c$ value. They were verified by statistical methods. Results: The blood glucose range at smooth prediction, predicted 5 min, 15 min, 30 min in advance, were 3.4,~11.5 mmol/L, 3.3~11.4 mmol/L and 3.2~11.4 mmol/L, respectively. Similarly, the rang of LAGE were less than and equal to 3.5 mmol/L, 3.8 mmol/L and 3.3 mmol/L, respectively. After grouping by ranges above, the $R$ values (correlation coefficient) of the predicted and measured values of group B were higher than those of group A and group C in different predicted time lengths (P<0.01). Nonparametric tests showed that there was a statistical difference in absolute error between each two groups (P<0.01). The $R$ value of group D and group E did not much differently in different prediction time (P<0.01). Nonparametric tests of two independent samples showed that the difference of prediction error between groups was statistically significant (P<0.01). Conclusion: GM (1,1) model had better predictive performance for blood glucose data at 3.4~11.4 mmol/L.