Addressing flowing bottom-hole pressure data scarcity in underground gas storage wells: A simulation-augmented surrogate modeling framework for VLP prediction

Reliable Vertical Lift Performance (VLP) relationships are essential for operating underground gas storage (UGS) wells, yet measured flowing bottom-hole pressure (FBHP), the standard training signal for data-driven VLP models, is unavailable at many UGS facilities. This study presents a three-phase information-transfer framework that removes this requirement. A history-matched reservoir simulator converts sparse static pressure surveys, routinely collected for inventory verification, into synthetic FBHP data for surrogate modeling. In Phase 1, the simulator generates FBHP data under historical operating conditions. In Phase 2, seven surrogate models are trained and evaluated using nested, cycle-grouped cross-validation (5 outer × 5 inner folds). In Phase 3, the selected model is converted into a vertical flow performance (VFP) table, the discretized form of the VLP read by the simulator, and evaluated inside the simulator against measured flowing tubing head pressure (FTHP) over a nine-year history. The framework is demonstrated on a single UGS well with 475 withdrawal-phase and 527 injection-phase daily observations in 9 withdrawal and 10 injection cycles (2011–2020), without any measured FBHP. No candidate was significantly more accurate than ordinary least-squares regression in cross-validation (RMSE: 1.252 bar withdrawal, 1.718 bar injection). The linear surrogate was selected for its small generalization gap, exact VFP representation and two- or three-coefficient expression. End-to-end, the linear VFP table reproduced measured FTHP with RMSE = 1.377 bar (95% cycle-bootstrap interval: 1.18–1.64 bar), MAPE = 3.11 % , R 2 = 0.967 , and a mean residual below 0.01 bar. A nonlinear GMDH candidate was 0.06 bar more accurate end-to-end but less accurate and less stable in cross-validation. This shows operational consistency within the historical operating envelope, not a verification of the unmeasured FBHP. Out-of-time tests showed that a change in storage operating scheme requires retraining. Support Vector Regression achieved the lowest training RMSE (0.851 bar) but failed end-to-end (withdrawal RMSE of 5.91 bar), so training accuracy alone is insufficient for model selection. Decomposing the deployment gap between cross-validation and end-to-end accuracy showed zero pure tabularization error for linear surrogates and at most 1.0 bar for curved candidates.