Background and Objectives: Accurate daily river flow forecasting particularly in humid catchments characterized by dynamic hydrological regimes plays a vital role in sustainable water resources management, hydrological planning, and flood risk mitigation. In recent years, soft computing–based models, including machine learning and deep learning approaches, have garnered significant attention in water engineering due to their inherent capacity to capture complex, nonlinear, nonstationary, and multivariate relationships. However, empirical evidence indicates that a systematic and comparative evaluation of advanced artificial intelligence models especially under humid catchment conditions and at daily temporal resolution remains insufficient. This research gap not only limits a comprehensive understanding of the strengths and limitations of each method but also poses serious challenges to the transferability, robustness, and operational reliability of these models in practical decision-making. To address this gap, the present study evaluates and compares the performance of machine learning and deep learning models for daily river flow prediction, with a specific focus on a humid basin subject to climate variability, thereby clarifying methodological innovations and operational applicability of these AI-driven approaches.
Materials and Methods: In this study, daily meteorological and hydrometric data spanning the period from 1969 to 2018 were first collected. Precipitation (P_t), evaporation (E_t), and river flow discharge with one- to three-day lags (Q_t₋₁ to Q_t-3) were used as input variables. Pearson correlation coefficients were calculated between input and output variables, based on which five input scenarios and model configurations were selected. Data preprocessing (including homogenization, outlier removal, reconstruction of missing values, and normalization) was performed, and the dataset was split into training and testing subsets in a 70:30 ratio. For modeling, one machine learning algorithm Support Vector Regression (SVR) and two deep learning models Convolutional Neural Network (CNN) and Generative Adversarial Network (GAN) were employed. Model performance was evaluated using multiple metrics: Coefficient of determination (R²), Root Mean Square Error (RMSE), Percent Bias (PBIAS), and Kling–Gupta Efficiency (KGE). The results were compared in tabular form, and final outcomes were visually presented using scatter, time series, violin plot, and Taylor diagrams.
Results: The results indicated that SVR model achieved the best performance in Scenario 5 (SN5), yielding R² of 0.85, RMSE of 5.675 m³/s, PBIAS close to zero (0.475%), and KGE of 0.877 during the testing phase. In this scenario, SVR outperformed the CNN and GAN models by 3.8% and 3.4% in terms of R², and by 14.1% and 14.0% in terms of KGE, respectively. Moreover, the CNN and GAN models demonstrated acceptable performance only in the more complex scenarios (SN₃ to SN₅), while exhibiting poor results in the simpler scenarios (SN₁ and SN₂), with R² values below 0.04. Overall, SVR with the optimal input combination (SN₅) provided the most stable and accurate predictions, showing statistically significant superiority over the other two models across all effective scenarios.
Conclusion: In daily river discharge forecasting within highly variable basins, model input structure exerts a far greater influence than algorithmic complexity so much that excluding lagged discharge values led to complete predictive failure (R² < 0.04), even when employing deep learning architectures. Among the evaluated models, Support Vector Regression (SVR) under the optimal scenario (SN5) achieved the most accurate and unbiased performance in the testing phase, with R² = 0.850, RMSE = 5.675 m³/s, and PBIAS = 0.475%, significantly outperforming both CNN and GAN. These findings substantiate a fundamental principle in hydroinformatics: Under conditions of limited data and high climatic stress, model selection should be guided by alignment with the hydrological character of the data, rather than by algorithmic novelty alone, a conclusion with direct implications for designing operational forecasting systems and adaptive water governance in vulnerable regions.