A deep learning model for predicting the production of coalbed methane considering time, space, and geological features

Abstract

Coalbed methane (CBM) is high-quality clean energy and accurate prediction of daily gas production of CBM is critical for CBM engineering. However, the production process of CBM is a non-stable dynamic with significant fluctuation, and it is hard to predict by traditional statistical methods. This study processes a deep learning model T-DGCN considering time, space, and geological features for predicting complex long gas production sequences. T-DGCN innovatively measures the similarity of geological features between wells with Dynamic Time Warping (DTW), and merges geological and spatial features to dynamically correct the weight matrix in a multilayer neural network with multiple aggregations. Then, the model uses the Gated Recurrent Unit (GRU) to extract the temporal features of gas production and predict the daily gas production sequence. The experiments with the data set from Shanxi Province showed that T-DGCN achieves an accuracy of 0.9298 in short-term production prediction, which is higher than the baseline models. In addition, the geological similarity calculated by DTW in T-DGCN significantly improves the performance of the model. And T-DGCN can still have better performance in long-term prediction tasks with accuracy above 0.9. This study provides a new method for the theoretical guidance for adjusting development schemes of CBM and the prediction of long-time series in geoscience.