A novel prediction method for coalbed methane production capacity combined extreme gradient boosting with bayesian optimization

Coalbed methane plays a significant role for the sustainable utilizing of resources and ecological environment. Production capacity forecasting of coalbed methane wells can effectively guide the optimization of development schemes directly affecting the economic benefits. To overcome the inefficiency of traditional theory-based numerical simulators and their weak adaptability to observational data, we explore a potential and efficient alternative for modeling of production capacity in a data-driven approach. This study makes full use of dynamic production data and geological static data from 530 CBM wells. We develop a production capacity prediction model utilizing the extreme gradient boosting algorithm and incorporated bayesian optimization to implement an automated search for hyperparameters. The results demonstrate that the prediction model developed by extreme gradient boosting has a more powerful prediction performance with an R-2 close to 0.9 compared to other machine learning or even deep learning. Moreover, the coupled framework of extreme gradient boosting and bayesian optimization can notably upgrade the prediction power of the production capacity model by about 8%. The analysis of influencing factors also illustrates that dynamic production data during the first three years of development can well characterize the coalbed methane adsorption-desorption-seepage features, which contribute to the construction of the production capacity model.