Author(s): Kyoungwon Min; Gyumin Lee; Doosun Kang

Linked Author(s):

Keywords: Feature importance; Machine learning; Temporal dependency; Urban water supply; Water demand forecasting

Abstract: Accurate water demand forecasting is essential for ensuring the reliable operation and planning of urban water supply systems. Previous research has primarily focused on improving forecast accuracy by applying machine learning techniques to historical demand and meteorological data. However, the performance gap between algorithms is often small, making it difficult to identify superior models based solely on accuracy. In this study, we analyze how different combinations of input variables influence forecasting performance using hourly water consumption data from Incheon, Republic of Korea. Multiple machine learning models are trained with varied feature groups, and the contribution of each variable is evaluated. The results indicate that the previous week's water demand at the same hour is the most influential factor for improving forecasting accuracy, reflecting the strong temporal dependency in consumption behavior. These findings offer practical guidance for selecting meaningful input variables and planning data acquisition strategies, and can support the development of more reliable operational decision-making frameworks for reservoir, water treatment plant, and intake facility operations in urban water supply systems.

DOI:

Year: 2026