Belief Propagation(BP) decoders of Polar Codes can run in parallel and have a remarkable throughput achievement at a cost of performance degradation, when compared with other decoding algorithms of polar codes. The existing improved algorithms still have a gap to be optimized in performance, complexity or latency. This paper proposes a modified BP flipping algorithm by introducing random penalty items into the reliability metrics of bits to correct un-converged errors under the type of noise with Gaussian distributions considered. Simulation results illustrate that the proposed penalizing method can exhibit significant performance gains over the original BP and other three existing BP flipping decoding algorithms with different blocks in the additive white Gaussian noise channel(AWGN) under a lower decoding latency. This shows the proposed scheme identifies the error-prone bits more efficiently.