Polygenic risk score (PRS) has been shown to be predictive of disease risk such as type 2 diabetes (T2D). However, the existing studies on genetic prediction for T2D only had limited predictive power. To further improve the predictive capability of the PRS model in identifying individuals at high T2D risk, we proposed a new three-step filtering procedure, which aimed to include truly predictive single-nucleotide polymorphisms (SNPs) and avoid unpredictive ones into PRS model. First, we filtered SNPs according to the marginal association p-values (p≤ 5× 10−2) from large-scale genome-wide association studies. Second, we set linkage disequilibrium (LD) pruning thresholds (r2) as 0.2, 0.4, 0.6, and 0.8. Third, we set p-value thresholds as 5× 10−2, 5× 10−4, 5× 10−6, and 5× 10−8. Then, we constructed and tested multiple candidate PRS models obtained by the PRSice-2 software among 182,422 individuals in the UK Biobank (UKB) testing dataset. We validated the predictive capability of the optimal PRS model that was chosen from the testing process in identifying individuals at high T2D risk based on the UKB validation dataset (n = 274,029). The prediction accuracy of the PRS model evaluated by the adjusted area under the receiver operating characteristics curve (AUC) showed that our PRS model had good prediction performance [AUC = 0.795, 95% confidence interval (CI): (0.790, 0.800)]. Specifically, our PRS model identified 30, 12, and 7% of the population at greater than five-, six-, and seven-fold risk for T2D, respectively. After adjusting for sex, age, physical measurements, and clinical factors, the AUC increased to 0.901 [95% CI: (0.897, 0.904)]. Therefore, our PRS model could be useful for population-level preventive T2D screening.