Performance Evaluation of Machine Learning Models (Random Forest and M5 Model Tree) for Runoff Simulation Under Climate Change and Reservoir Operation Scenarios

Shabnam Vakili

doi:10.1007/s11269-026-04829-3

Performance Evaluation of Machine Learning Models (Random Forest and M5 Model Tree) for Runoff Simulation Under Climate Change and Reservoir Operation Scenarios

2026-08-10

0

PRE

AI

S

Shabnam Vakili *

S

S. Morteza Mousavi

DOI：10.1007/s11269-026-04829-3

Original

Original request for help

Save

Abstract

En 中文

Rainfall–runoff (R–R) modeling is essential for effective water resources management, particularly under changing climatic conditions and increasing anthropogenic pressures on hydrological systems. This study evaluates the capability of machine learning models for runoff simulation in the Kharroud River basin, Iran, under climate variability and reservoir operation scenarios. Two machine learning approaches, Random Forest (RF) and the M5 Model Tree (M5), were applied using monthly hydro-meteorological data from 1984 to 2025. Precipitation (Pt−1 and Pt−2), two-month cumulative precipitation (Cum2 = Pt−1 + Pt−2), mean temperature, evaporation, and lagged runoff variables (Qt−1 and Qt−2) were used as input predictors. The dataset was divided into 70% training and 30% testing subsets, and model performance was assessed using R², RMSE, MAE, and NSE. Results showed that RF consistently outperformed M5 across all scenarios, with R² and NSE values ranging from 0.81 to 0.94, while M5 ranged from 0.65 to 0.74. Scenario analysis indicated that reduced precipitation combined with increased evaporation intensified runoff decline and increased predictive uncertainty. Under these conditions, RF demonstrated higher robustness and accuracy than M5, particularly in simulating extreme flows. In contrast, reservoir regulation reduced flow variability and improved the predictive performance of both models. Overall, RF showed superior ability in capturing nonlinear relationships and hydrological extremes, confirming its suitability for runoff prediction under non-stationary conditions. The proposed framework supports adaptive reservoir operation, irrigation planning, and sustainable water allocation under climate uncertainty. Future studies should investigate hybrid and deep learning models, along with additional hydro-climatic predictors and uncertainty analysis, to further improve predictive reliability.

Keywords:

Rainfall-runoff

Climatic conditions

M5 model tree

Random Forest (RF)

Performance Evaluation of Machine Learning Models (Random Forest and M5 Model Tree) for Runoff Simulation Under Climate Change and Reservoir Operation Scenarios

Abstract

Journal

Organization

Cited Papers

Citing Papers