Document Type : Original Article
Authors
1
PhD student in Civil Engineering, Department of Civil Engineering, Faculty of Engineering, Estahban Branch, Islamic Azad University, Estahban, Iran.
2
Assistant Professor, Department of Civil Engineering, Faculty of Engineering, Babol Branch, Islamic Azad University, Babol, Iran
3
Assistant Professor, Department of Civil Engineering, Faculty of Engineering Estahban Branch, Islamic Azad University, Estahban, Iran
4
Professor, Department of Civil Engineering, Faculty of Engineering, Estahban Branch, Islamic Azad University, Estahban, Iran.
5
Assistant Professor, Department of Civil Engineering, Faculty of Engineering, Estahban Branch, Islamic Azad University, Estahban, Iran.
10.22125/iwe.2026.601094.1946
Abstract
The rainfall–runoff process is a complex hydrological process, and its accurate simulation depends on the appropriate selection of input variables and temporal data structure. This study aimed to investigate the effect of precipitation and discharge temporal memory on machine learning model performance and to identify suitable input lag structures for rainfall–runoff simulation. Daily precipitation and discharge data from the Eghbalabad station in Fars Province, Iran, were used. After data preprocessing, including missing-data assessment, outlier detection, and normalization, discharge lags Q1–Q6 and precipitation lags P1–P7 were generated. Subsequently, 49 scenarios with different combinations of precipitation and discharge lags were designed, and the performance of four models, including CatBoost, Random Forest, XGBoost, and LSTM, was compared. Results showed that the temporal memory structure of the inputs affected model performance, and increasing the number of lags alone did not necessarily improve performance. Among the evaluated scenarios, the Q6+P4 combination in scenario S4 achieved the highest NSE of 0.7525 using CatBoost. The Q3+P7 and Q6+P7 combinations also achieved high NSE values of 0.7513 and 0.7498, respectively. The findings indicate that simultaneously considering precipitation and discharge lags and selecting an appropriate temporal memory length can improve rainfall–runoff simulation.
Keywords
Subjects
