Continuous-variable quantum key distribution (CVQKD) is an important application of quantum technology, which enables long-distance communicating parties to establish a string of unconditionally secure keys in an insecure environment. However, in a practical CVQKD system, the finite sampling bandwidth of the analog-to-digital converter (ADC) at the receiver may create inaccurate sampling results, leading to errors in parameter estimation process and leaving a security loophole for eavesdroppers. In order to eliminate the finite sampling bandwidth effect, we propose a peak-compensation-based CVQKD scheme, which estimates the discrepancy between the maximum sampling value and the peak value of each pulse based on the characteristics of Gaussian pulse. The maximum sampling values are compensated by the estimated discrepancy, so that the legitimate parties can obtain correct sampling results. We analyze the influence of the finite sampling bandwidth on the security of the system, expounding the specific steps of peak-compensation, comparing the estimated excess noise before and after peak-compensation, and discussing the security of the system under Gaussian collective attacks. Simulation results show that this scheme can greatly improve the accuracy of pulse peak sampling and remove the finite sampling bandwidth effect. Moreover, the channel parameters estimated by the communicating parties are also corrected by using the compensated values. Compared with the scheme without peak-compensation, this scheme eliminates the limitation of the system repetition to the secret key bit rate, and has longer secure transmission distance and higher secret key bit rate. In addition, compared with other methods of solving the finite sampling bandwidth effect, the proposed scheme can be directly implemented in data processing stage after sampling without any additional devices, and thus increasing no complexity of the system.