A hybrid physics-artificial intelligence approach for accurate prediction of reference evapotranspiration

In a context marked by water scarcity as is the case in Morocco - particularly in semi-arid regions most exposed to water challenges such as the Souss Massa region - cultivation under cover exerts enormous pressure on water resources, making increasingly precise irrigation management essential. This work proposes a hybrid approach to improve the accuracy, generalization and stability of reference evapotranspiration predictions, integrating physical laws into the neural network architecture, which makes it possible to create a model that respects both the observed data and the physical knowledge governing reference evapotranspiration. The proposed methodology is based firstly on the evaluation of three deep learning architectures with advanced attention mechanism (Attention-based LSTM, Attention-based bidirectional-LSTM, Attention-based CNN-LSTM), secondly the evaluation of the best architecture before and after the integration of the physical component (Physics-Informed Neural Networks) using a convex combination integrating the Priestley-Taylor physical model. The results show the superiority of the hybrid architectures outperforming the others, the Attention-based CNN-LSTM architecture already obtaining interesting performances (R2 = 0.934).

However, the PINNs architecture with a balance coefficient set at λ = 0.1 outperforms all other architectures with less error and better data explanation (R² = 0.945). This combination allows a reduction of the average absolute error of 7.5% compared to the Attention-based CNN-LSTM model also ensuring better stability of predictions against extreme values. The validation is carried out in a prototype connected greenhouse equipped with IoT sensors and a monitoring dashboard.

This hybrid physico-learned approach offers a scalable and interpretable solution for intelligent irrigation management in semi-arid conditions.