Author(s): Santiago Polti Albisu; Rafael Gonzalez Perea; Emilio Camacho Poyato

Linked Author(s):

Keywords: Water distribution networks; Deep learning; Time series forecasting; Temporal fusion transformer

Abstract: In modern irrigation districts, on-demand pressurized water distribution networks introduce substantial uncertainty for system managers, who must balance reliable water delivery with operational and energy costs. Reliable middle-term irrigation demand forecasts are therefore critical for operational planning. This work develops and evaluates Transformer-based models for 7-day forecasting of irrigation demand occurrence in an on-demand irrigation district characterized by intermittent, climate-driven consumption. The problem is formulated as a two-stage approach separating irrigation occurrence from irrigation demand magnitude. This paper focuses on the classification stage, implementing a Temporal Fusion Transformer (TFT) combined with Focal Loss to emphasize hard-to-classify irrigation events and support a recall-oriented design The main methodological contribution lies in adapting a multi horizon TFT framework to the specific . challenges of intermittent irrigation demand through a recall-oriented classification design. Model stability was assessed through stochastic validation using 10 independent random initializations. Across runs, F1-scores ranged from 0.945 to 0.952, MCC from 0.814 to 0.837, and PR-AUC exceeded 0.99. Recall consistently exceeded 98%, with precision around 90–93%. Results demonstrate that the TFT architecture effectively fuses static structural data with dynamic meteorological inputs, enabling reliable detection of irrigation events without performance degradation across the 7-day forecast horizon, supporting robust operational planning.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.305

Year: 2026