在农业水资源管理领域，参考作物蒸腾量的精确预测对灌溉水高效利用至关重要。当前逐日预测方法未能充分利用日内动态变化信息，限制了预测准确性。为解决该问题，研究提出了一种基于外部注意力机制（EA）的双向长短时记忆网络（BiLSTM）模型，使用沙猫群算法（SCSO）优化模型超参数，实现逐小时参考作物蒸腾量预测。首先利用SCSO方法对EA-BiLSTM模型进行优化，优化后的算法在70个epoch后收敛，平均 $R^{\mathrm{2}}$升至0.750；进而通过特征相关性分析，将模型输入的特征数据由10个减少为历史ET ₀、太阳辐射、空气温度、空气湿度和最大风速5个。以北京市昌平区的国家精准农业研究示范基地大田种植区ET ₀预测为例进行了方法验证，在对未来第7小时的预测中，R ²从0.619提高到0.644，获得了更好的预测效果；最后通过对模型可解释性分析证实，历史ET ₀对预测的贡献最高，贡献率达到了0.043，其次是空气湿度和总辐射。与DT（决策树）、Lasso（最小绝对收缩和选择算法）、LMP（多层感知机）、CNN（卷积神经网络）等预测方法的对比结果表明，采用EA-BiLSTM-SCSO的预测结果在MAE和MSE指标上均获得了最低的误差值，在R ²指标上，EA-BiLSTM-SCSO模型平均达到0.722较CNN模型提升了12.6%。研究验证了深度学习与特征工程在提高作物参考蒸腾量逐小时预测精度方面的优势。该方法在智慧灌溉中用于估算作物的水分需求，能够实现对未来灌溉的精准预测，从而制定合理的灌溉计划，提高灌溉水利用效率，进行有效的灌溉用水调度。

In the field of agricultural water resource management, accurate prediction of crop transpiration is crucial for efficient utilization of irrigation water. The current daily prediction method fails to fully utilize the dynamic changes within the day, which limits the accuracy of the prediction. To address the issue, this study proposes a bidirectional long short term memory network (BiLSTM) based on external attention mechanism (EA), which optimizes model hyperparameters using the Sand Cat Swarm Optimization (SCSO) algorithm for hourly reference crop transpiration prediction. Firstly, the SCSO method was used to optimize the EA-BiLSTM model. The optimized algorithm converged after 70 epochs, and the average coefficient of determination increased to 0.750. Furthermore, five characteristic parameters, historical ET ₀, solar radiation, air temperature, air humidity, and maximum wind speed, were selected from the 10 features. In this paper, the method was verified in the field planting area of the national precision Agriculture research Demonstration base in Changping, Beijing. In the prediction of the next 7 hours, the coefficient of determination was increased from 0.619 to 0.644, achieving better prediction results. Finally, through the interpretability analysis of the model, it was confirmed that historical ET ₀ had the highest contribution to the prediction, reaching 0.043, followed by air humidity and total radiation. The comparison results with DT, Lasso, LMP, CNN and other forecasting methods show that the prediction results using EA -BiLSTM-SCSO have achieved the lowest error values in both MAE and MSE indicators. In terms of coefficient of determination, the EA-BiLSTM-SCSO model achieved an average of 0.722, which was 12.6% higher than the CNN model. The study validated the advantages of deep learning and feature engineering in improving the hourly prediction accuracy of reference crop transpiration. The method is used to estimate the water demand of crops in intelligent irrigation, which can realize the accurate prediction of future irrigation, so as to formulate reasonable irrigation plans, improve the efficiency of irrigation water utilization, and carry out effective irrigation water scheduling.