Enhancing the accuracy and generalizability of reference evapotranspiration forecasting in California using deep global learning

This research focuses on the Central Valley of California, a climatically homogeneous region known for its significant agricultural productivity and reliance on extensive irrigation. Our study utilizes monthly reference evapotranspiration (ET O ) time series data from 55 standardized weather stations as part of the California Irrigation Management Information System (CIMIS). ET O is a critical component of regional water cycles, indicating atmospheric water demand. This study evaluates the potential of deep learning (DL) models for ET O forecasting, particularly emphasizing the efficacy of a global learning scheme compared to traditional local learning. Global learning involves training forecasting models on pooled data from multiple time series, tested over new instances. We compared the performance of statistical models and advanced DL models, demonstrating significant accuracy enhancements in global learning schemes. We also explored automatic hyperparameter optimization for these models to achieve state-of-the-art forecasting accuracy, yielding RMSE values below 10 mm/month for one-year-ahead forecasts on new, unseen stations. Applying global learning methodologies to DL models markedly improved forecasting performance, showcasing an ability to generalize findings to ungauged regions and even newly established weather stations. This suggests a promising avenue for enhancing water resource management efficiency in data-scarce areas. Our findings argue that such data-centric methodological shifts could play a critical role in better managing the irrigation demands of the Central Valley, thereby supporting sustainable water usage and agricultural productivity in the region. • Global learning enhances the accuracy of DL models for regional ET O forecasting. • High-capacity DL models outperform traditional methods with global learning. • Achieves RMSE below 10 mm/month for one-year-ahead ET O forecasts on unseen stations. • Models are generalizable to ungauged regions and new weather stations. • Automatic hyperparameter tuning boosts the performance of deep global forecasting.