Prediction of gas drainage changes from nitrogen replacement: A study of a TCN deep learning model with integrated attention mechanism

Abstract

Accurate prediction of pure gas volume changes in nitrogen injection replacement coal seams is of great significance to increase coalbed methane (CBM) production and prevent underground gas disasters in coal mines. However, the mechanism of coal seam gas flow during nitrogen injection replacement is complex, and it is extremely difficult to achieve quantitative prediction of the pure volume of gas in nitrogen injection replacement coal seams. In this study, we propose a new model for predicting the change of pure amount of gas in coal seam of nitrogen injection replacement in coal mines (Attention-TCN). Meanwhile, we believe that a stable and reliable predictive model should have the ability to capture the underlying patterns and relationships of regular and noisy datasets during the learning process, rather than memorizing training data. Therefore, in this paper, the Attention-TCN model is trained and tested using the results of nitrogen injection and replacement methane double diffusion test (regularity dataset) and the results of nitrogen injection and replacement experiment in coal mine underground engineering (noisy dataset), respectively. The difference between the Attention-TCN model and the baseline models of TCN, gated recurrent neural network (GRU) and long and short-term memory neural network (LSTM) was also compared. The validity of the model was demonstrated using ten-fold cross-validation results. The results show that the proposed Attention-TCN prediction model achieves good prediction performance on both regular and noisy datasets, with prediction R2 greater than 0.99, and is highly flexible and robust. The model is better able to extract the basic features and structures in different types of datasets, enabling it to perform well on unseen data. The ten-fold cross-validation results clearly show that the Attention-TCN model is more accurate and all outperforms the TCN, GRU and LSTM prediction models. In addition, the prediction results of all models for different mine underground nitrogen injection replacement coal seam pure gas volume datasets further validate that the Attention-TCN model has better ability to generalize to new datasets and is more effective and practical for predicting the variation of pure gas volume in underground nitrogen injection replacement coal seams in coal mines.