Interpretable Deep Learning Approach for Production Forecasting of Fractured Horizontal Wells

Timely and accurate forecasting of well production in tight gas reservoirs is of paramount importance for comprehensive field development and financial decision-making. Despite the application of deep learning (DL) models in constructing predictive frameworks, their intricate nature and limited interpretability present challenges for petroleum engineers, hindering their understanding of learned inference mechanisms and trust in predictions. This study advocates for the adoption of interpretable DL solutions, incorporating SHapley Additive exPlanations (SHAP), to provide explicit elucidations of a prediction model by implemented by gated recurrent networks. The method’s efficacy is substantiated using data from the Tao2 gas field in the Ordos Basin, China. The outcomes underscore the model’s exceptional predictive capabilities. Leveraging the SHAP method for global interpretation provides valuable insights into the collective impact of various factors. Simultaneously, employing SHAP for local interpretation furnishes personalized explanations for well productivity predictions. The findings gleaned from this research are poised to enhance well operations and field development planning.