Abstract:

In order to improve the efficiency of agricultural water use in southern Xinjiang, a TCN-Attention-GRU-based water demand prediction model for jujube trees was constructed to characterize the agricultural irrigation system in southern Xinjiang. The model took jujube trees as the research object, took meteorological data as the input parameters of the model, and the output quantity was the water demand of jujube trees. Firstly, the attention mechanism (Attention) was used to extract the importance features from the data, and then the processed data were sent into the temporal convolutional network (TCN) to grab the temporal features to fuse them into a new feature vector, the final prediction was made using a gating unit (GRU), according to the characteristics of the combined model with the validation of the multi-step prediction, the proposed TCN-Attention-GRU combination model predicted well. The tested TCN-Attention-GRU based jujube trees water demand prediction model determination coefficient (R²) reached 94.4%, mean absolute percentage error (MAPE) and mean square error (MSE) were 7.9% and 28.8%, respectively, and the relative average error between the actual value and the predicted value was 12.23%, which had higher prediction accuracy compared with other models, and it could effectively improve the rate of water resources, and the proposed prediction model provided some references to the agricultural water conservation and stabilization. The proposed prediction model provides a certain reference for agricultural water conservation and stable production.

Key words: time convolutional network, combination model, crop water demand prediction, jujube