Optimizing Priestley–Taylor Model Based on Machine Learning Algorithms to Simulate Tomato Evapotranspiration in Chinese Greenhouse

To further improve the prediction accuracy for greenhouse crop evapotranspiration (ET) under different irrigation conditions and enhance irrigation water use efficiency, this study proposes three methods to revise the Priestley–Taylor (PT) model coefficient α for calculating ET at different growth stages: (1) considering the leaf senescence coefficient fS, plant temperature constraint parameter ft, and soil water stress index fsw to correct α (MPT model); (2) combining the Penman–Monteith (PM) model to inversely calculate α (PT-M model); (3) using the machine learning XGBoost algorithm to optimize α (PT-M(XGB) model). Accordingly, this study observed the cumulative evaporation (Ep) of a 20 cm standard evaporation pan and set two different irrigation treatments (K0.9: 0.9Ep and K0.5: 0.5Ep). We conducted field measurements of meteorological data inside the greenhouse, tomato physiological and ecological indices, and ET during 2020 and 2021. The above three methods were then used to dynamically simulate greenhouse tomato ET. Results showed the following: (1) In 2020 and 2021, under K0.9 and K0.5 irrigation treatments, the MPT model mean coefficient α for the entire growth stage was 1.27 and 1.26, respectively, while the PT-M model mean coefficient α was 1.31 and 1.30. For both models, α was significantly lower than 1.26 (conventional value) during the seedling stage and the flowering and fruiting stage, rose rapidly during the fruit enlargement stage, and then gradually declined toward 1.26 during the harvest stage. (2) Predicted ET (ETe) using the PT-M model underestimated the observed ET (ETm) by 8.71~16.01% during the seedling stage and the harvest stage, and overestimated by 1.62~6.15% during the flowering and fruiting stage and the fruit enlargement stage; the errors compared to ETm under both irrigation treatments over two years was 0.1~3.3%, with an R2 of 0.92~0.96. (3) The PT-M(XGB) model achieved higher prediction accuracy, with errors compared to ETm under both irrigation treatments over two years of 0.35~0.65%, and R2 above 0.98. The PT-M(XGB) model combined with the XGBoost algorithm significantly improved prediction accuracy, providing a reference for the precise calculation of greenhouse tomato ET.