Data-driven multidimensional static-dynamic models for the prediction of coalbed methane performance based on numerical models

The accurate prediction of coalbed methane (CBM) is often challenged by geological conditions, engineering technologies, and data gaps. However, with an interdisciplinary field application of intelligent algorithms, data-driven models can effectively predict its productivity characteristics. Here, this work develops data-driven models to predict CBM static and dynamic productivity. Based on the regression relationships of static datasets, models are developed based on intelligent algorithms including multiple linear regression, random forest, support vector regression, and gradient boosting regression (GBR). Due to the temporal variations and nonlinearity of dynamic datasets, models based on recurrent neural network, long short-term memory (LSTM), and gated recurrent unit (GRU) have been developed. The results show that GBR demonstrates the best performance according to the evaluation metrics. In importance permutation, GBR prioritizes matrix porosity, hydraulic fracture intrinsic permeability and fracture permeability, and these three parameters account for nearly 90% of its contributions. Both LSTM and GRU demonstrate strong prediction capabilities in dynamic productivity. Moreover, data-driven models require substantially less time computing. LSTM possesses stronger prediction performance than GRU in the gas adsorption mass and reservoir pressure field evolution. In conclusion, this work provides a multidimensional prediction and evaluation system for the productivity prediction of unconventional resources.