Objective  Mathematic models for accurate real-time prediction on evapotranspiration of greenhouse cucumber plants during fruiting period were evaluated to optimize the irrigation operation.

  Method   Cucumber plants were cultivated in a greenhouse. During the fruiting period, microclimate conditions were automatically monitored by sensors and recorders, and plant evapotranspiration determined by weighing the fruits. Using transplanting time, air temperature, relative humidity, light intensity, and daily average irrigation amount of previous 5d as inputs, models including the Back Propagation Neural Network (BPNN), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) were evaluated according to the cucumber evapotranspiration prediction. A data collection interval of 20min was applied.

  Result  Of the tested models, GRU performed with the highest coefficient of determination (R²) of 0.8577, root mean square error (RMSE) of 2.3279 g, and mean absolute error (MAE) of 1.6744 g. It also yielded the lowest relative error fluctuation between the predicted and the measured data ranging from 0.11% to 10.01% when the daily real-time cumulative evapotranspiration of cucumbers exceeded 50 g.

  Conclusion   The GRU-based model could best predict the greenhouse cucumber evapotranspiration at fruiting stage. The information could aid better water management for cucumber cultivation.