EVALUATION OF PERFORMANCE OF ADVANCED MACHINE LEARNING MODEL AND PHYSICAL MODEL IN STREAMFLOW MODELING OF BEAS RIVER BASIN

Abstract

The need for more accurate hydrological models is fundamental in addressing water resource management challenges, especially in understanding and predicting streamflow dynamics. This study aims to assess and enhance the predictive performance of hydrological models by integrating a traditional hydrological model, HEC-HMS, with a machine learning-based approach using a Long Short-Term Memory (LSTM) network. The general objective is to evaluate the performance of the LSTM and HEC-HMS models in streamflow modeling, with specific objectives focused on developing both models, comparing their performance, and exploring the integration of these models to improve the physical representation in the LSTM model. Initially, the HEC-HMS model was developed and calibrated to simulate the hydrological processes of a designated catchment area. Performance of the HEC-HMS model alone showed a Nash-Sutcliffe Efficiency (NSE) of 0.723, Root Mean Square Error (RMSE) Standard Deviation of 0.5, Percent Bias (PBIAS) of 6.41%, and a Coefficient of Determination (R²) of 0.74, indicating good model accuracy and reliability in reproducing the observed data. Parallelly, an LSTM model was tailored for the same dataset, focusing on capturing temporal dependencies and sequence patterns in streamflow data. In its standalone application, the LSTM model demonstrated superior performance metrics during both calibration and validation phases with an NSE as high as 0.98, a Kling-Gupta Efficiency (KGE) of 0.84, and a consistent R² of 0.82. Although it showed a higher Percent Bias at 26.15% during calibration, it reduced significantly to 11.55% in the validation phase, illustrating its capacity to adjust and learn from the dataset effectively. To bridge the gap between empirical and data-driven approaches, the study explored a hybrid model where outputs from the HEC-HMS model were used as inputs to the LSTM. This integration aimed at refining the predictive accuracy by combining the physical process representation of HEC-HMS with the dynamic learning capabilities of LSTM. The hybrid model markedly enhanced the model performance with an NSE of 0.99, KGE of 0.92, R² of 0.92, and a reduced MAE to 69.27 cumecs. The Percent Bias improved to 5.79%, highlighting the effectiveness of this integrative approach in streamflow prediction.