Integrating HEC-RAS with machine learning and deep learning models as a rapid tool for river water level prediction

For flood management, accurate water surface elevation predictions are necessary. hydraulic models such as HEC-RAS produces detailed water surface elevation predictions but requires extensive input data and computational time. This study tested a hybrid approach on the Khorram Abad River in Lorestan Province, Iran, combining HEC-RAS with machine learning and deep learning models. HEC-RAS version 6.0 generated 856 records from 167 cross-sections and eight return periods (2 to 500 years). The dataset was split by cross-section (74 for training, 33 for testing) to prevent information leakage. Four models were trained: Random Forest (RF), Support Vector Machine (SVM), Deep Neural Network (DNN), and Recurrent Neural Network (RNN). RF had the highest accuracy in testing, with an R 2 of 0.9909, an RMSE of 0.619 meters, and an MAE of 0.446 meters. Channel bed elevation accounted for 32% of the prediction importance, and log-transformed discharge accounted for 24%. RF maintained its accuracy across all return periods. Repeated cross-validation gave a mean R 2 of 0.9933 with a standard deviation of 0.0029. The Wilcoxon test showed RF outperformed SVM (p = 0.000), DNN, and RNN (p < 0.001) with statistical significance. Bootstrap uncertainty analysis produced 95% confidence intervals for RMSE [0.5546, 0.8363] meters and R 2 [0.9834, 0.9927]. The framework provides a faster alternative to HEC-RAS for this river and may be useful for design flood estimation and hazard mapping. Testing on other rivers and under unsteady flow conditions is needed before operational flood warning use.