Physics-Guided and Horizon-Aware Dynamic Liquid-Level Identification and Forecasting in Oil Wells Using Distributed Temperature Sensing

Accurate dynamic liquid-level monitoring in rod-pumped oil wells requires not only reliable identification of the gas–liquid interface from distributed sensing data but also stable forecasting over extended horizons. This study develops a physics-guided and horizon-aware PatchTST framework (PGH-PatchTST) for dynamic liquid-level identification and forecasting based on distributed temperature sensing (DTS). The measured time–depth–temperature field is reconstructed and denoised, and the gas–liquid interface is localized using the maximum absolute spatial temperature gradient. Unlike the previous multiscale formulation that uses only the derived liquid-level sequence and fixed multiscale feature fusion, PGH-PatchTST incorporates liquid-level depth, temperature-gradient magnitude, peak-dominance index, and liquid-level variation rate as DTS-aware state features. A physics-guided routing mechanism adaptively weights patch representations at lengths of 8, 16, and 24 samples. In addition, a horizon-aware trend–residual forecasting branch adjusts the contributions of local fluctuations and long-term trends according to the forecasting distance, aiming to reduce systematic degradation at long horizons. The proposed framework demonstrated improved forecasting performance compared with representative time-series forecasting models, while the horizon-aware mechanism effectively alleviated systematic deviation during long-horizon prediction.