Author(s): Mariana Akemi Ikegawa Bernabe; Rafael Gonzalez Perea; Juan Antonio Rodriguez Diaz; Jorge Garcia Morillo

Linked Author(s):

Keywords: Artificial intelligence; Deep learning; Energy resource management; Renewable energy; Time series forecasting; Transformer neural network

Abstract: Accurate energy demand (ED) forecasting is essential for optimizing the use of water and energy resources across diverse sectors. This study evaluates a Transformer-based encoder–decoder model for hourly, medium-term ED forecasting in four end-users: fisheries (Ireland), an irrigation district (Spain), a port (Spain), and a local energy community (Portugal). Input variables were selected through fuzzy logic and correlation analysis, resulting in 7 inputs for the Fisheries, Port, and Community cases and 12 inputs for the Irrigation District. The proposed model achieved consistently high predictive performance across all case studies, with R² values between 97.9% and 99.62% and low error metrics. The highest accuracy was obtained for the Irrigation District (R² = 99.62%, MSE = 0.038), while the Community and Port cases showed the lowest MAE values (0.44 kWh and 6.64 kWh, respectively). Differences in absolute error levels are mainly associated with variations in demand magnitude and operational variability. Overall, the results confirm the robustness and generalization capability of Transformer-based models under diverse input configurations and data conditions, supporting their applicability for reliable ED forecasting and operational planning in multi-sector environments.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.76

Year: 2026