Abstract: As an important basic energy in China, the clean and efficient utilization of coal is of great significance to achieve the goal of “double carbon”. The rapid and accurate acquisition of coal quality parameters is the premise to realize the efficient utilization of coal. However, the traditional coal ash content determination method is complex and time-consuming, and it is difficult to meet the needs of real-time monitoring. Laser-induced breakdown spectroscopy ( LIBS ) technology provides a new idea for rapid detection of coal ash content due to its advantages of rapid, in-situ and multi-element simultaneous analysis. However, LIBS technology is affected by matrix effect and signal stability in practical applications, resulting in insufficient quantitative analysis accuracy. Therefore, this paper proposes a self-attention neural network ( DRAN ) model based on dominant factor-residual correction. The model combines physical-driven feature line analysis with data-driven global feature learning by using a dual-branch parallel architecture. On the one hand, the dominant factor branch constructs a sub-model based on the feature line of the target element to ensure the physical interpretability of the model prediction. On the other hand, the residual correction branch introduces a self-attention mechanism to dynamically weight and fuse the full-spectrum information, thereby improving the adaptability and generalization performance of the model to complex samples. In order to verify the performance of the model, two sets of LIBS data sets of coal ash from different sources were used to compare the proposed model with the commonly used baseline methods. The results show that DRAN shows the best performance on the test sets of the two sets of data sets, and its root mean square error of prediction (RMSE_P) is 0.996% and 2.448%, respectively.Compared with the baseline model with the best performance, the errors are reduced by 9.2% and 13.3%, respectively. The results confirm that the model still maintains stable prediction performance and shows good extrapolation ability when the distribution of test samples exceeds the training range. In addition, through ablation experiments and contribution analysis, it is found that the synergy between the dominant factor branch and the self-attention mechanism plays a key role in improving the performance and extrapolation ability of the model, and the prediction results of the model are highly correlated with the ash-related characteristic lines, which has clear physical significance. Compared with the traditional method, the proposed DRAN model effectively improves the intelligent quantitative analysis ability of LIBS technology under complex matrix conditions by integrating physical mechanism and deep learning method, and provides a new solution for the accurate detection of coal quality.