Prediction of Crop Water Requirements in the Chushandian Irrigation District Based on TA-Bi-GRU

Accurate prediction of crop water requirements is key to efficient water resource scheduling and agricultural irrigation management in irrigation districts. Taking the Chushandian Reservoir Irrigation District as the study area, this research calculated crop water requirements using the Penman-Monteith formula and crop coefficient method, based on monthly meteorological data (1981–2020) from the Changtaiguan hydrological station and five surrounding national meteorological stations (Zhumadian, Xinyang, Queshan, Zhengyang, and Luoshan), along with crop coefficients for six major crops (wheat, rapeseed, corn, peanut, rice, and vegetables). The water requirement distribution across different growth stages was then analyzed, and XGBoost-based feature importance ranking combined with Pearson correlation analysis was used to identify the key meteorological factors influencing crop water requirements. Finally, a bidirectional gated recurrent unit incorporating a temporal attention mechanism (TA-Bi-GRU) was constructed and compared with MGMA-LSTM, GCRA-TCN, and LSTM-CNN in both joint prediction and crop-specific independent prediction experiments. The results showed that vegetables had the highest total water requirement over the full growth period (656.79 mm), followed by rice (475.04 mm), and rapeseed the lowest (260.05 mm); the net irrigation water requirement of each crop peaked during the mid-growth stage. Relative humidity, sunshine hours, and maximum temperature were identified as the top three meteorological factors affecting crop water requirements, with a cumulative importance exceeding 75%. In the joint prediction task, TA-Bi-GRU achieved an R2 of 0.9487 and an RMSE of 0.2257 mm, representing the best numerical performance among the four evaluated deep-learning models. In the crop-specific independent prediction task, TA-Bi-GRU achieved R2 values of no less than 0.9024 across the six crops. By integrating bidirectional temporal representation with temporal attention, TA-Bi-GRU shows potential as a data-driven tool for monthly crop-water-requirement prediction and irrigation decision support in the studied district.