Comparative Performance of SWAT and Machine Learning Models for Runoff Prediction During Extreme Climatic Events: A Case Study of the Hunza River Basin, Pakistan

Abstract This research evaluates the prediction capabilities of the Soil and Water Assessment Tool (SWAT) and machine learning (ML) models. Climatic data from 2007 to 2022 and five machine learning models: XGBoost, Long Short-Term Memory (LSTM), K-Nearest Neighbours (KNN), Support Vector Regression (SVR), and Artificial Neural Networks (ANN) were used for predicting runoff in the Hunza River Basin (HRB), Pakistan. Models were assessed using R², RMSE, NSE, RSE, PBIAS, and MAE, and all underwent hyperparameter tuning to improve performance. XGBoost outperformed other models with an R² of 0.88, RMSE of 54.56, and PBIAS of 10.41%. The ANN model indicated competitive results with R² of 0.85 and RMSE of 61.64. In contrast, the LSTM model provided effective performance at seasonal prediction, yielding a higher R2 of 0.86 and a lower RMSE of 58.83. KNN and SVR models showed moderate performance and underestimated at peak flows. KNN yielded R² at 0.84 and RMSE of 64.97, whereas SVR evaluated R² at 0.83 and RMSE at 41.76. The SWAT model showed lower prediction, with R² at 0.77, RMSE at 77.52, PBIAS at 10.74%, and RSE at 0.28. The runoff simulation model faced difficulties, particularly under conditions of elevated runoff. Five-fold cross-validation, however, affirmed the overall robustness of the machine learning models. Residual analyses employing Partial Autocorrelation Function (PACF) and Autocorrelation Function (ACF) plots stressed the models constrained ability to capture immediate dependencies and extreme runoff events accurately. The results show that ML models perform better than the SWAT model in glacier basins and provide more accurate runoff predictions. But, in the future, Additional factors need to be incorporated to enhance the model’s accuracy during extreme runoff events.