Machine Learning-Based Flood Forecasting at Downstream Station (Case Study: Zarrinehrud River, Iran)

Flooding events, as a natural hazard, create challenges for decision-makers seeking to identify potential scenarios for flood forecasting. In recent years, Machine Learning (ML) has emerged as a trending decision-supporting tool to capture the complexity and nonlinear relationships within hydrological prediction. In this study, the monthly river peak discharges of the Zarrinehrud River, Iran, are used to investigate five ML algorithms, namely RF and XGBoost as a tree-based ensemble, SVM as a supervised learning algorithm, ELM as a feedforward neural network, and MLP as a deep learning algorithm at the downstream station named Nezamabad. The input data comprises 300 months of river peak discharges between 1994 and 2018. The study aims to compare the predictive performance of different ML algorithms for downstream discharge forecasting using an 80:20 train-test split. After training the original models, MLP outperformed with higher NSE (0.96) and lower RMSE (7.06 m³/s), and visually followed the overall trend. Meanwhile, the residual analysis highlighted remaining challenges in reproducing extreme peak discharges, representing the uncertainties and the heteroscedasticity in prediction. Hence, the hyperparameter tuning approach led to significant improvements in model performance, particularly for SVM and MLP, achieving higher NSE values from 0.74 to 0.98 for SVM and from 0.97 to 0.99, as confirmed by the improved alignment of peaks for optimised versions of the algorithms. Additionally, the residuals depicted a significant reduction in bias and improved fit to the data for the MLP algorithm. This study relies solely on training the downstream station’s historical river discharge data, and for further investigation, incorporating meteorological inputs and upstream hydrological data is required to explore promising modelling behaviour.