Abstract: Precisely determining the orbit of near-Earth asteroid (NEA) is one of the important parts of the near-Earth objects monitoring and early warning. However, due to the constraints of the observation arc length, observation accuracy and force model, different objects have different orbit errors. Based on the orbit data of more than 30000 near-Earth asteroids in JPL (Jet Propulsion Laboratory) small body database, the orbital elements error can be analyzed statistically. It is found that the semi-major axis error of the NEAs orbit has a bimodal distribution, and the reason is related to the distribution of the observation arc. The relationship between semi-major axis error and observation arc length is studied, and a regression equation with a goodness of fit of 0.90 is obtained. Furthermore, taking the absolute magnitude parameter into account, BP (Back Propagation) neural network training method is used to build a parameter training network for the beginning and end time span of observation, orbit period, absolute magnitude and semi-major axis error. The proposed method further improves the goodness of fit to 0.96, which can be used to quickly and reasonably evaluate the semi-major axis error of near-Earth asteroids. In addition, the influence of the observation arc length on the semi-major axis error, eccentricity error and inclination error is compared, and it is found that the improvement degree of the three is different with the increase of the observation arc length. Finally, the variation of the distribution of semi-major axis and orbit inclination error with orbit inclination is compared and analyzed, and it is found that the distribution characteristics of inclination error are related to the selection effect of observation accuracy. These statistical analyses contribute to a better understanding of the distribution of NEA orbit errors, and provide reference for further improvements in the orbital accuracy.