Comparative analysis of deep learning and machine learning models for one-day-ahead streamflow forecasting in the Krishna River basin

Karad, Keesara, Sarati and T.Ramapuram catchments located in the Krishna River basin, India This study focused on 1-day ahead streamflow forecasting in four distinct catchments using a wide array of Deep Learning (DL) and Machine Learning (ML) models. A comprehensive evaluation of eleven models was conducted to assess their strengths and limitations across different datasets. The study implemented Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), Bidirectional GRU, Convolutional Neural Network, WaveNet, K-Nearest Neighbours, Random Forest (RF), Support Vector Regression, Adaptive Boosting, and Extreme Gradient Boosting (XGBoost) to forecast streamflow at each site. Lagged precipitation and antecedent streamflow emerged as key predictors. Model performance was assessed using multiple evaluation metrics and visualization techniques. Bi-LSTM achieved the highest performance in three catchments with Nash-Sutcliffe efficiency (NSE) of 0.864 in Karad, 0.708 in Keesara, and 0.702 in T. Ramapuram, while GRU performed best in Sarati with NSE close to 0.7. The best model achieved "very good" accuracy in one catchment and "good" in three, as indicated by performance metrics. However, even the best-performing DL models struggled to capture peak flow events, revealing limitations in extrapolation. The study also highlights the potential of ML models based on ensemble techniques, such as RF and XGBoost, which demonstrated performance comparable to that of complex DL architectures. • A rigorous comparison of six DL and five ML models across four unique catchments. • Variants of LSTM and GRU emerged as the best models for streamflow forecasting. • WaveNet outperformed Convolutional Neural Networks in capturing streamflow patterns. • RF and XGBoost delivered results nearly on par with advanced DL models.