A VMD-KPCA-LSTM-attention model for multi-parameter prediction in coal mine goaf environments

Abstract

Coal spontaneous combustion and the emission of harmful gases in goafs pose significant threats to mine safety. Accurate prediction of environmental parameter variations in goaf areas is crucial for hazard identification and preventive control. This study develops a novel multi-parameter prediction model for goaf environments using variational mode decomposition (VMD), kernel principal component analysis (KPCA), and a long short-term memory neural network with an attention mechanism (LSTM-Attention). The environmental parameters analyzed include absolute pressure, temperature, O₂ concentration, and CO concentration. VMD is employed to perform adaptive time-frequency decomposition, revealing intrinsic features of parameter sequences at different time scales, while KPCA is used to reduce feature dimensionality, retaining 98.33 % of the data's information content and eliminating feature redundancy. Finally, the LSTM-Attention model captures nonlinear relationships and temporal dependencies in the environmental data, providing accurate predictions. The results demonstrate that goaf environmental parameters exhibit multi-scale, nonstationary characteristics with abrupt changes. O₂ concentration and absolute pressure show strong correlations, while CO concentration and temperature display significant nonlinear and sudden variation patterns. The proposed VMD-KPCA-LSTM-Attention model achieves superior prediction accuracy and robustness compared to traditional models, with significantly reduced errors and improved generalization capability. This study provides a new methodological framework for predicting complex environmental parameter variations in coal mine goafs, contributing to improved mine safety and intelligent hazard prevention.