Physics-Informed Neural Network-Based Inversion of Distributed Temperature Sensing Data for Injection Profile Interpretation

Distributed temperature sensing (DTS) provides continuous wellbore temperature measurements but does not directly quantify interval water-intake rates. This study develops a physics-informed neural network (PINN) framework for water-intake profile inversion in injection wells. A transient coupled wellbore–reservoir model accounts for variable-mass flow, pressure variations, heat transfer, reservoir seepage, and well trajectory. The governing equations, DTS observations, and initial and boundary conditions are incorporated into the loss function to jointly reconstruct the temperature field and identify interval intake rates. Softmax parameterization enforces non-negative intake rates and conservation of the total injection rate. Sensitivity analysis identifies injection temperature, injection rate, and injection duration as the principal factors affecting wellbore temperature. Synthetic cases with different interval thicknesses and permeabilities yield temperature root mean square errors (RMSEs) of 0.1061–0.1268 °C and intake-rate RMSEs of 0.979–1.854 m3/d, with maximum interval relative errors not exceeding 2.611%. A deviated-well case further demonstrates applicability to the tested trajectory variation. For a field injection well, the reconstructed temperature has a mean absolute error (MAE) of 0.0559 °C and an RMSE of 0.0901 °C. The inferred principal intake intervals and relative allocation agree with reference measurements acquired approximately three weeks apart, supporting the method’s applicability to DTS-based intake-profile interpretation.