Development of dam inflow prediction technique based on explainable artificial intelligence (XAI) and combined optimizer for efficient use of water resources

Abstract

Accurate inflow forecasts are crucial for managing water resources, particularly in regions experiencing both floods and droughts. This study proposes a combined optimizer (CO) that combines adaptive moment and vision correction algorithms to improve the shortcomings of deep learning optimizers, thereby enhancing deep learning accuracy. CO improves the shortcomings of deep learning optimizers, such as storage space and local optimal solution convergence potential. Additionally, explainable artificial intelligence (XAI) was applied to CO, creating a model termed Dual-AI, which enhances interpretability and accuracy. As a result of application to Daecheong Dam in Korea, Dual-AI showed a maximum reduction of root mean squared error (RMSE) by approximately 3.68 (R2 increased by about 0.0628) in verification and approximately 678.4922 (R2 increased by about 0.0664) in prediction compared to the existing optimizer. Dual-AI shows potential for various hydrological applications, providing accurate forecasts to support effective water management.