TY  - EJOU
AU  - Qu, Yaguang 
AU  - Zhang, Can 
AU  - Tao, Zhen 
AU  - Ren, Lixia 
AU  - Wang, Bin 
AU  - Hou, Pengjun 

TI  - Research on Interwell Connectivity Based on Machine Learning
T2  - Energy Engineering

PY  - 
VL  - 
IS  - 
SN  - 1546-0118

AB  - The study of interwell connectivity is crucial for understanding and optimizing the transmission and exchange of fluids, gases, and other substances within subsurface reservoirs. Accurate assessment of reservoir connectivity is key to developing effective development strategies and enhancing the recovery factor of water-flooding reservoirs. This research conducts a quantitative study of dynamic interwell connectivity using production dynamic data from injection and production wells simulated through reservoir numerical modeling. It employs a combination support vector regression (SVR) baseline models, long short-term memory (LSTM) models, mutual information (MI) methods, and SHapley Additive exPlanations (SHAP) values. In the inverse dynamic-response design, the daily oil-production rates of the four production wells are used as model input sequences, and the corresponding daily injection volume of the central injection well is used as the regression target. This input-output setting is adopted to infer injector-producer dynamic communication patterns from production responses and does not imply that production physically controls injection. Two analytical methods are used to provide reasonable interwell connectivity relationships and connectivity coefficients. The study analyzes the interwell connectivity of three typical characteristic reservoirs: homogeneous reservoirs, reservoirs with an impermeable fault, and high-permeability-streak reservoirs. The results show that the combination of LSTM models and SHAP values achieves a high degree of consistency with simulated seepage patterns and high model accuracy, verifying the practicality of the model in determining interwell connectivity coefficients. In addition, the capacitance-resistance model (CRM) is discussed as a classical physically interpretable baseline rather than ignored. Since all validation samples are generated from numerical simulation, the present results are framed as controlled synthetic verification of methodological feasibility and physical consistency, not as independent field-scale validation for real reservoirs.
KW  - Interwell connectivity; machine learning; support vector regression; long short-term memory; mutual information; SHAP; reservoir numerical simulation

DO  - 10.32604/ee.2026.086721