Production Forecast of Deep-Coalbed-Methane Wells Based on Long Short-Term Memory and Bayesian Optimization

Abstract

Summary This study analyzes the production behaviors of six deep coalbed-methane (CBM) wells (>1980 m) completed in the Ordos Basin and presents a machine-learning method to predict gas production for six target wells. The production behaviors of target wells are characterized with several months of rapidly declining pressure, following by several years of stabilized gas rate and pressure. Production data analysis suggests a relatively large amount of free gas (but limited free water) in coal seams under in-situ condition. The production mechanisms generally transit from free-gas expansion and fracture/cleat closure at early stage to gas desorption at later stage. We treated the target wells’ production data as time-series data and applied the Long Short-Term Memory (LSTM) model on the target wells for gas-rate predictions. We also employed a Bayesian-probabilistic method to optimize the LSTM model (BO-LSTM). Our results demonstrate the BO-LSTM model’s robustness in gas-rate predictions for target wells. Also, treating casing pressure and liquid level as inputs is sufficient for the BO-LSTM model to reach a reliable production forecast. This study provides a promising tool to forecast the gas production of deep-CBM wells using surface rates and pressure data. The findings of this study may guide the reservoir management and development-strategy optimizations of deep-CBM reservoirs.