Author(s): Santiago Polti Albisu; Rafael Gonzalez Perea; Emilio Camacho Poyato
Linked Author(s):
Keywords: Water distribution networks; Deep learning; Time series forecasting; Transformer neural network
Abstract: In modern Irrigation Districts (IDs), where distribution networks operate under pressurized conditions and often provide continuous, on-demand access, the availability and management of water has become a critical concern. Allowing farmers to draw water at any time introduces substantial uncertainty for ID managers, who must balance energy use but, above all, ensure the efficient allocation of water resources and control the associated operational costs. Reliable forecasts of irrigation water demand several days in advance would enable more precise system management, contributing to improved water-use efficiency and reduced energy cost. This work evaluates Transformer-based models for 7-day forecasting of irrigation water demand in an on-demand ID, where consumption is highly variable due to climate and farmer decisions. The prediction of irrigation water demand was divided into two distinct tasks: determining when irrigation occurs at ID scale (a classification problem) and estimating how much water is actually required (a regression problem). Thus, two Transformer-based models were implemented: a classification model forecasting irrigation event occurrence, and a regression model estimating continuous demand. Preliminary classification results (F1 = 0.93, PR-AUC = 0.98, MCC = 0.86) indicate that attention mechanisms effectively capture behavioural and climatic patterns. The regression model achieved an R² value above 0.93 with an error below 12%. These findings highlight the potential of Transformer models to support medium-term irrigation forecasting and improve planning in on-demand water distribution networks for irrigation.
Year: 2026