Abstract:

In order to effectively predict hydraulic support loads and evaluate the operational status of supports, a hydraulic support load prediction model based on MTAM-LSTM was proposed. The CEEMDAN algorithm was employed to decompose the load data of supports and extract intrinsic mode functions. Redundant components in the intrinsic mode functions were eliminated according to K-L divergence criterion, thereby forming the input sequence for load prediction. An MTAM was constructed to capture the variation characteristics of hydraulic support loads. Static attention generated attention weights for feature information of data, while dynamic attention optimized the focus on different sequence features. Residual learning was introduced to maintain the integrity of feature signals. LSTM networks were then utilized to establish deep dependencies between feature information and hydraulic support loads, enabling advanced prediction of support load data. Field data from the 402102 working face of a rockburst-prone coal mine in Shaanxi were used for empirical validation. RMSE, R², and MAE were used as evaluation metrics for comparison among different models. The results show that the RMSE and MAE of the MTAM-LSTM model are significantly lower than those of the comparison models, with RMSE reduced by 0.16-0.45 and MAE reduced by 0.16-0.45, while the coefficient of determination R² reaches 0.91 under different scenarios, thereby validating the prediction accuracy and generalization capability of MTAM-LSTM model.

Key words: multi-convolution temporal attention module (MTAM), long short-term memory (LSTM), mining working face, load prediction, hydraulic support, generalization ability