Objective This paper aimed to find a method to improve the convergence speed of nonparametric estimation for the probability density function of a univariate random variable. Methods By combining parameter estimation and nonparametric estimation with weighting a semiparametric estimation method based on regular penalty was proposed. In this method a parameter model was assumed to obtain its corresponding parameter estimation. Nonparametric estimation was computed using kernel methods. A weight was then obtained through empirical likelihood loss with a regular penalty or integral square loss. Based on this weight the proposed semiparametric estimation based on regular penalty was obtained by combining the parameter estimation with nonparametric estimation. Results Asymptotic theoretical properties show that the proposed semiparametric estimation method based on regular penalty combines the advantages of parameter estimation and nonparametric estimation. It converges in any case without relying on any model assumption and when the parameter model assumption is correct the weight tends toward the parameter estimation resulting in the same convergence speed as the parameter estimation otherwise it tends toward the nonparametric estimation with consistent convergence speed as nonparametric estimation. Numerical simulation experiments show that when the data satisfy the parameter model assumption the weight of the penalized semi-parametric density estimation method tends toward the parameter estimation consistent with the theoretical results when the data do not satisfy the parameter model assumption the weight of the proposed semi-parametric estimation tends toward the nonparametric estimation also consistent with the theoretical results. Finally this method was applied to precipitation data in Chongqing to study the distribution of monthly precipitation. Conclusion The results of the case study show that the proposed semiparametric density estimation method based on the regular penalty is smoother compared with nonparametric estimation and fits better compared with parameter estimation thus validating the rationality of this method.