To effectively solve the problems of the accuracy loss of prediction models for traffic accident prediction of open-pit mines, the decrease of the algorithm's capability of anti-noise tolerance and convergence rate caused by small sample data and outliers, we propose a modified support vector regression model based on penalized trimmed squares (PTS). According to the data distribution characteristics of the training samples, the penalized trimmed squares submitted to the application scenario of open-pit mine is studied to improve the anti-noise tolerance capability of the regression model. In consideration of the difficulties of the nonlinear prediction model impact factor selection, the method of principal component analysis is introduced into the preprocessing algorithm to reduce the data dimension and ensure that the algorithm can get ideal input data. In view of the problems of premature and slow convergence speed caused by the nuclear parameter selection, the inertial factor and the learning factor of particle swarm are studied and an improved particle swarm algorithm to optimize nuclear parameters regression of the model is proposed. The prediction and comparison experiments are carried out in the case of the accident frequency prediction of open-pit mine. The experimental results show that the test set prediction results of the PTS model are better than those without the PTS policy model.This indicates that the modified penalized trimmed squares strategy and parameter optimization algorithm proposed in this paper is feasible and effective for the study of accident prediction of complex systems.