Objective Mine pressure prediction represents an important means for early warning and management of disasters in coal seam roofs, serving as the prerequisite and foundation for safe production in intelligent mines. However, the complex and variable conditions of fully mechanized mining face lead to significant variations in the distribution of support pressure data acquired using electro-hydraulic control systems, complicating mine pressure prediction.

Methods This study developed a Transformer-based mine pressure prediction model. Specifically, missing mine pressure data were filled using linear interpolation, and the data structure of mine pressure was adjusted using a sliding window algorithm. Based on the time-series features of mine pressure data, an input sequence incorporating these features was constructed. The dynamic weight computation of the model was performed using the multi-head attention mechanism, while adaptive focus on critical time steps was achieved based on the data itself. Consequently, complex nonlinear time-series dependencies can be effectively captured, thus significantly enhancing feature representation and prediction performance. Finally, the support resistance prediction for the upper, middle, and lower sections of the mining face was conducted using the transfer learning method, and a big-data-driven intelligent platform for mine pressure analysis and prediction was established.

Results and Conclusions  By capturing global features of mine pressure data using a multi-head attention mechanism rather than neural networks, the Transformer model demonstrated stronger long-sequence dependencies and feature learning capabilities compared to the recurrent neural network (RNN) and long short-term memory (LSTM). Therefore, this model effectively reduced loss and was more suitable for mine pressure prediction, with a mean squared error (MSE) of 0.34% and a mean absolute error (MAE) of 2.57% on the test set. Furthermore, the Transformer model exhibited robust generalization performance. This model exhibited effectively reduced loss after fine-tuning with transfer learning and yielded superior generalization performance when transferred to adjacent supports within the same mining face. These findings further validated the applicability and feasibility of this model in mine pressure prediction. The visualization system of the intelligent platform presented the precise analysis results of key parameters before and after prediction, including roof weighting frequency, advancing distance of the mining face, weighting criterion, and mine pressure contour maps. The results of this study provide a new philosophy for the early warning of disasters, including those in coal seam roofs, while also laying a solid foundation for safe, efficient mining and intelligent construction of mines.