Next-Day Prediction of 0–20 cm Volumetric Soil Water Content in a Cotton Field Under Mulched Drip Irrigation in Southern Xinjiang, China, Using an ET0-Informed Trend-Constrained Long Short-Term Memory Model

Surface soil water content is an important indicator for daily irrigation management in cotton fields under mulched drip irrigation in southern Xinjiang, China. This study evaluated a long short-term memory (LSTM) model incorporating reference crop evapotranspiration (ET0) and a trend-consistency (TC) constraint (ET0-TC-LSTM) for next-day prediction of 0–20 cm volumetric soil water content using the preceding 7 d of observations. Four observation periods during 2022–2025 yielded 468 supervised samples, divided chronologically into 169 training, 153 validation, and 146 test samples. ET0 was included as a feature, and a one-sided training penalty restricted predicted increases beyond a tolerance under recent low-water-input and high vapor pressure deficit (VPD) conditions. Persistence, ridge regression, and a one-dimensional convolutional neural network–LSTM (CNN1D-LSTM) model were included as benchmarks. In the original seed-42 experiment, ET0-TC-LSTM achieved a test coefficient of determination (R2) of 0.9574 and a root mean square error (RMSE) of 0.0081 cm3 cm−3, a 29.7% RMSE reduction relative to LSTM. Across ten seeds, the mean RMSE reduction was 2.84%; TC-LSTM had the lowest mean RMSE among the four LSTM variants. Persistence was more accurate, with an RMSE of 0.0046 cm3 cm−3. Sensitivity and conditional analyses showed that the effects of process-informed inputs and penalties depended on parameter settings and evaluation conditions. These results identify improvements within the LSTM framework, but do not establish superiority over persistence for this site and forecast horizon.