To improve the specific absorption rate (SAR) compression model capability in parallel transmission (pTx) MRI systems. A k-means clustering method is proposed to group voxels with similar SAR behaviors in the scanned object, providing a controlled upper-bounded estimation of peak local SARs. This k-means compression model and the conventional virtual observation point (VOP) model were tested in a pTx MRI framework. The pTx pulse design with different SAR controlling schemes was simulated using a numerical human head model and an eight-channel 7T coil array. Multiple criteria (including RF power, global and peak local SARs, and excitation accuracy) were compared for the performance testing. The k-means compression model generated a narrower overestimation bound, leading to a more accurate local SAR estimation. Among different pTx pulse design approaches, the k-means compression model showed the best trade-off between the SAR and excitation accuracy. The developed SAR compression model is advantageous for pTx framework given the narrower overestimation bound and control over the compression ratio. Results also illustrate that a moderate increase of maximum RF power can be useful for reducing the maximum local SAR deposition.