Abstract: In response to the engineering challenges of complex top coal drawing laws and difficulties in determining a reasonable sublevel height during sublevel top coal caving mining of steeply inclined extra-thick coal seams, an integrated approach combining granular dynamics theory, discrete element numerical simulation, physical experiments, and deep learning algorithms was employed. This study revealed the "loosening-compaction" cycle mechanism dominated by the interlocking effect between top coal blocks in top coal caving mining. A multi-parameter numerical simulation scheme was developed, physical experiments were conducted for validation, and a prediction model for recovery rate and reasonable sublevel height based on sequence data was established. The influence patterns of sublevel height, face length, and coal seam dip angle on the interlocking effect and top coal recovery rate were investigated in this paper. Results indicate that a larger coal seam dip angle and greater sublevel height lead to a more pronounced interlocking effect, which is detrimental to the smooth drawing of top coal. In contrast, the face length has little influence on the interlocking effect. A smaller sublevel height helps weaken the interlocking effect and promotes smooth top coal drawing. A smaller face length coupled with a higher sublevel height results in more uniform coal drawing between the middle sections of the face. A larger dip angle and higher sublevel height cause a gradual increase in the amount of coal drawn from the openings near the roof side of the face. When the face length is larger and the coal seam dip angle is smaller, adopting a smaller sublevel height can effectively reduce top coal arching and improve the recovery rate of top coal. However, when the coal seam dip angle exceeds 70°, selecting a larger sublevel height can mitigate the interlocking effect and enhance the top coal recovery rate. A prediction model combining LSTM and Transformer was constructed for recovery rate and reasonable sublevel height, achieving accurate predictions (R²= 0.946 9). This research provides a dynamic optimization method for sublevel height in sublevel top coal caving faces, such as those in the Wudong Coal Mine, contributing to the high recovery rate mining of steeply inclined coal seams.