Laser-induced breakdown spectroscopy (LIBS) is an emerging atomic spectroscopy technique with promising applications in coal analysis but is limited by its relatively low quantification performance. Various machine learning methods have been applied in coal analysis on LIBS to improve its quantitative performance in recent years. However, most of these machine-learning models were established purely based on statistics. They ignored the physical rules involved in the quantification, resulting in reduced robustness, application range, and a lack of model interpretability. This work proposed a physics-statistics combined regression method based on the dominant factor (DF) and multilayer perception (MLP), called DF-MLP, to incorporate spectral domain knowledge into machine learning. The new proposed method built a physical-based dominant model to predictelement concentration with the characteristic lines selected with spectral knowledge and correct the residual errors using MLP. DF-MLP combines the dominant factor model and residual error correction using the MLP method can utilize the domain knowledge to improve model robustness and interpretability without reducing complexity. DF-MLP was compared with normal MLP, dominant factor partial least squares regression (DF-PLSR), dominant factor support vector regression (DF-SVR), and other baseline methods, and optimal results were obtained. Compared with normal MLP, the proposed method reduces root mean squared error of prediction (RMSEP) by 13.21%, 14.54%, and 21.77% for carbon, ash, and volatile, respectively. Compared with DF-SVR, the proposed method reduces RMSEP by 14.75%, 23.13%, and 5.99%, respectively. We further discussed the impact of different modeling patterns in the dominant factor method. The experimental results showed that combining domain knowledge with machine learning methods was a feasible approach to improve the performance of LIBS quantification.