In the operation and management of Underground Gas Storage (UGS), the accurate and efficient calculation of bottom hole pressure is crucial for the dynamic analysis and production optimization of gas wells. To enhance the operational and maintenance efficiency of UGS, this paper innovatively proposes a new method for calculating bottom hole pressure. The study begins by comprehensively analyzing the key factors affecting bottom hole pressure calculation during gas injection, withdrawal, and shut-in stages based on the wellbore flow theory. Subsequently, it delves into the characteristic variables closely related to bottom hole pressure and constructs a neural network model on this basis. Finally, by integrating the wellbore flow equations under different well conditions, using theoretical models to generate samples, and establishing a loss function guided by real samples, a theory and data-driven neural network model (TDDNN) has been successfully developed, achieving rapid and accurate calculation of bottom hole pressure. The novel method significantly outperforms traditional techniques across five precision metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and an enhanced R-squared (R2) value. Compared to traditional theoretical approaches, the method in this paper not only maintains high prediction accuracy but also significantly enhances computational efficiency, reducing the processing time from seconds to milliseconds. Furthermore, the method provides a valuable reference for the application of deep learning in environments with limited samples.
To address the issue of data distortion in traditional well log density reconstruction methods caused by harsh borehole environments in complex heterogeneous formations, and the challenge for general machine learning models to capture high-frequency abrupt change features due to the lack of geophysical constraints, this paper proposes a density curve reconstruction framework integrating geologically-inspired feature engineering and a mutation-aware mechanism. The method first constructs a multi-dimensional enhanced feature space covering sedimentary cycle characteristics and fluid properties through multi-scale geologically-inspired analysis, utilizing robust range normalization to eliminate interference from outliers. Subsequently, a mutation-aware driven algorithm based on the gradient change rate is innovatively designed. By dynamically adjusting loss weights, it resolves the drawback where traditional smoothing algorithms tend to erase formation interface information. Finally, geological prior constraint terms are embedded into the XGBoost architecture to achieve multi-objective optimization that balances data-driven accuracy with consistency in geological laws. Under leave-one-well-out cross-validation over ten wells, the proposed method achieves a mean R² of 0.98, an RMSE of 0.029, and an MAE of 0.022. Experimental results demonstrate that the proposed method not only improves reconstruction accuracy but also significantly enhances the physical consistency of the results. Furthermore, it precisely identifies thin interbedded boundaries, providing highly reliable basic data support for the refined evaluation of oil and gas reservoirs.
Predicting the injection-production capacity of depleted gas reservoir underground gas storage is schallenging due to the difficulties in combining complex inter-well dynamic coupling with physical consistency, alongside the poor computational efficiency of traditional mechanistic models. To address these issues, a physics-informed graph neural network (PI-GNN) model is proposed, integrating well pattern topological features with the Law of Conservation of Mass. In this approach, injection-production wells are abstracted as graph nodes and gathering pipeline connections as edges to construct a spatial graph structure reflecting the real production topology. Furthermore, hierarchical mass conservation constraints are incorporated into the network loss function to ensure that predictions strictly adhere to the total injection-production balance. Experimental results based on historical dynamic data from the Wei 11 gas storage in Zhongyuan Oilfield demonstrate that, compared to conventional numerical simulations and data-driven models, PI-GNN is about a thousand times faster in computation while maintaining high predictive accuracy. Moreover, it exhibits superior physical rationality and generalization across diverse injection-production scenarios. This research provides a novel pathway for developing high-fidelity, interpretable digital twins and rapid evaluation of gas storage facilities. Ultimately, this helps operators respond faster when gas demand is high.