Research on interval prediction of coal mine gas concentration and multi-index early warning method based on spatiotemporal attention mechanism

Aiming at the characteristics of deep coal mine gas emission, such as strong nonlinearity, strong non-stationarity, and scarcity of sudden abnormal samples, as well as the problem that traditional deterministic point prediction methods are difficult to characterize the uncertainty of prediction, thus restricting their application in safety decision-making, a proactive prevention, control and early warning method for gas disasters is proposed, which integrates spatiotemporal correlation modeling, prediction interval estimation and risk quantitative decision-making. Taking the 3272 working face of a coal mine in Sichuan Province as the research object, an empirical study was carried out based on multi-source sensor online monitoring time series data with a sampling interval of 1 minute over a period of nearly 8 months. A multi-attention spatiotemporal prediction interval network (MAST-PI Net) fusing multiple attention mechanisms and bidirectional long short-term memory networks is proposed. The network first adaptively extracts spatial heterogeneous features through a sensor attention module to dynamically characterize the contribution weights of different monitoring points to the prediction target. Then, combining the temporal attention mechanism and bidirectional long short-term memory network, it mines the nonlinear temporal features in the dynamic evolution process of gas emission. On this basis, quantile regression and pinball loss function are introduced to simultaneously output prediction intervals corresponding to 10%, 50% and 90% quantiles. Finally, for the future gas over-limit risk, physically meaningful features are extracted from three dimensions: concentration intensity, dynamic volatility and prediction interval width, and a physically-driven linearly weighted multi-index early warning index system is constructed. The experimental results show that under the test set, the MAST-PI Net achieves a root mean square error of 0.022 9, a prediction interval coverage probability of 0.835 0, and a normalized average interval width of 0.010 1. Compared with benchmark models such as long short-term memory network, temporal convolutional network and TimesNet, the proposed model exhibits better comprehensive performance in both prediction accuracy and interval quality. In the task of sudden gas anomaly identification, discriminant analysis based on the multi-index early warning index for the feature distribution of safe samples and high-risk samples achieves a recall rate of 97.89%, which is significantly higher than the static fixed threshold rule and traditional point prediction discriminant methods. It effectively alleviates the problem of risk omission caused by the regression smoothing effect. The proposed interval prediction and multi-dimensional early warning method breaks through the limitations of traditional deterministic prediction in risk characterization. While maintaining prediction accuracy, it can provide prediction interval information for quantitative decision-making. Moreover, by fusing multi-dimensional information such as intensity, volatility and uncertainty, the advance perception ability for concealed and sudden gas anomaly precursors is enhanced.