Short-term coalbed methane concentration prediction and early warning based on the STL-ENN-GRU hybrid model

Abstract

Coalbed methane (CBM) disasters are a major safety problem in coal mining, it is very important to accurately predict the concentration of CBM. Traditional prediction methods have shortcomings such as low prediction accuracy, inability to multi-step prediction, and insufficient data relationship mining. In this study, a hybrid model based on STL-ENN-GRU is proposed to predict short-term CBM concentration by analyzing a large amount of CBM measured data. First, the model establishes multi-dimensional data relationships through time-series decomposition. Then the multi-feature gated recurrent unit (GRU) prediction model is constructed to mine the implicit relationship between data and suppress the noise. Meanwhile, a hybrid strategy combining recursion and convolution is introduced to enhance the prediction model’s ability to memorize long-term relationships. On this basis, a CBM concentration warning platform based on edge computing is proposed by combining intelligent hardware. The proposed model was experimentally validated using four datasets collected from a mine in Shaanxi Coal Industry Company Limited (China). The results demonstrate that the proposed model exhibits the superior prediction accuracy and robust generalization capabilities compared to the baseline model. The root-mean-square error, mean absolute error, and R$_{}^2$ for the 3-hour prediction of CBM concentration are 0.018%, 0.0255%, and 0.965, respectively, the prediction accuracies are notably superior to those of other hybrid models. The early warning platform established on this model achieves a millisecond-level early warning response, demonstrating high efficiency and safety. It can offer auxiliary decision support for coal mine safety operations and CBM disaster prevention and control.