Online accurate proximate analysis of coal is vitally important to the optimization of industrial production and reduction in coal consumption. However, due to the “matrix effect” caused by the complex and diverse coal species in China, the measurement accuracy needs to be improved by using laser-induced breakdown spectroscopy (LIBS). In our experiment, both the spectral pretreatment method and the calibration model for the conversion of laser induced coal plasma spectra to the coal proximate analysis results were optimized. Experimental results showed that, compared with the traditional method, the proposed single- or multiple-peak Lorentzian spectral fitting for spectral line intensity calculation reduced the mean RSD from 12.1% to 9.7%. For kernel function parameters optimization, the mean absolute error (MAE) of the particle swarm optimization (PSO) was smaller than that of the grid parameter (Grid) and the genetic algorithm (GA). The root mean square error (RMSEP) of support vector machine (SVM) regression model based on PSO parameter optimization was less than that of partial least squares regression (PLS). By combining the single- or multiple-peak Lorentzian spectral fitting method with the PSO based SVM for regression modeling, the average absolute errors (AAE) of predicted proximate analysis results were certified to be: 1.37% for coal ash content of 16%~30%, 1.77% for coal ash content of 30% or more, 0.65 MJ·kg-1 for calorific value of 9~24 MJ·kg-1, 1.09% for volatile matter of 20% or less, and 1.02% for volatile matter of 20% or more.