The Weirong Gas Field is China’s first deep shale gas field with over 100 billion m3 of proven reserve. Since its development in 2019, production forecast and analysis has encountered great challenges. The production is very unstable at the early stage, with high decline rate and water production rate. In addition, complex geological conditions often induce downhole problems, which fluctuate or even suspend production. Therefore, it is hard to accurately predict the production potential for new wells and analyze the key factors of production. This study aims to apply artificial intelligence methods to capture the complex relationship between static and dynamic parameters and early-stage shale gas production, and quantitatively analyze the key factors affecting long-term production.
First, a deep learning algorithm named temporal fusion transformer (TFT) is implemented to predict daily shale gas production using static and dynamic variables. This algorithm synthetically combines parameters in different dimensions using encoding mechanisms, gated networks, LSTM network, and multi-head attention. These mechanisms ensure full information integration while prevent overly complex models. After TFT synthesizes a stable production profile, long-term production can be predicted using decline curve analysis. Then, automated machine learning (AutoML) is used to model the relationship between geological and engineering factors and long-term production. Controlling factors that affect long-term production are identified through quantitative model interpretation with Shapley values.
By validating the method using field data from the Weirong shale gas field, TFT accurately predicts daily gas production over multiple horizons using static geological and engineering factors, historical pressure and production, and pressure over the forecast horizon. The importance of these parameters can be ranked, and dependence of production on dynamic parameters can be determined. AutoML automatically constructs an ensemble ML model, which improves the prediction accuracy of long-term production compared to single ML models when the sample size is very small, and requires no manual model selection and hyperparameter tuning. Finally, Shapley values coupled with the AutoML model quantitively analyze the impact of geological and engineering factors on long-term production.
This study proposes a data-driven method that predicts short-term and long-term shale gas production for new wells at early production stage and analyzes key geological and engineering factors. The method is especially useful when development time is short, production is highly volatile, and when it is hard to maintain a stable production profile to predict long-term production. The workflow comprehensively integrates multi-modal data in different dimensions and enable more efficient production forecast and key factor identification. The results provide suggestions and insights to fracturing optimization of new wells in the future.
