Self-attention transformer model for pan evaporation prediction: a case study in Australia

Abstract

ABSTRACT In drought-prone regions like Australia, accurately assessing evaporation rates is essential for effectively managing and maximising the use of precious water resources and reservoirs. Current estimates show that evaporation reduces Australia's open water lake capacity by about 40% annually. With climate change, this water loss is expected to become an even greater concern. This study investigates a transformer-based neural network (TNN) to estimate monthly evaporation in three Australian locations. The models were trained and tested using monthly weather data spanning from 2009 to 2022. Input parameters were chosen based on Pearson's correlation coefficient values to identify the most impactful combinations. The developed TNN model was compared with two widely used empirical methods, namely Thornthwaite and Stephens and Stewart. The TNN model's impressive accuracy in evaporation prediction, attributed to its unique self-attention mechanism, suggests its promising potential for future use in evaporation forecasting. Additionally, the study revealed an intriguing result: Despite using the same input datasets, the TNN model surpassed traditional methods, achieving an average improvement of 18% in prediction accuracy. The TNN prediction model accurately predicts water loss (average R² = 0.970), supports irrigation management and agricultural planning and offers financial benefits to farming and related industries.