Predictive capabilities of novel deep neural networks learning for long-term streamflow prediction: insights from the barandouz chay river

Dams and reservoirs are essential for public health, food security, economic growth, and flood protection, especially as climate change challenges sustainable water management. This study explores regression-based deep learning models for long-term prediction of Barandouz River flow. Four models were employed: Continuous Weighted Residual Method (CWERM), Physics-Informed Koopman Networks (PI-Koopman), Convolutional Neural Network (CNN), and Implicit Neural Representations (INRs). Forty years (1980–2020) of daily time-series data were used, including temperature, precipitation, relative humidity, wind speed, and discharge. To improve prediction accuracy, lagged discharge (up to three days) was incorporated. Data preprocessing included missing value reconstruction (KNN), outlier removal, and normalization to mitigate overfitting risks. Feature selection and scenario construction were guided by the Minimum Redundancy Maximum Relevance (MRMR) method, resulting in five predictive scenarios. Model evaluation employed multiple criteria (R², MAE, RMSE, MAD, NSE, and KGE). Results show that Scenario 4 excluding historical flow variables exhibited severe predictive failure across all models in both training and testing phases, with testing R² ranging from 0.087 (CWERM) to 0.189 (PI-Koopman) and NSE remaining below 0.23, highlighting the critical role of antecedent discharge in modeling hydrological memory. In contrast, CNN consistently outperformed all other models in both phases. During testing, CNN achieved R² values of 0.991 (SN1), 0.992 (SN3), and 0.981 (SN5), with corresponding NSE > 0.910 and RMSE as low as 4.157 m³/s (SN5). Its average testing R² (0.977) was 14.0–18.3% higher than those of ANN-OA (0.857), INRs (0.855), and PI-Koopman (0.826), while its average RMSE (7.25 m³/s) was 34.5–44.1% lower. PI-Koopman, despite moderate performance (best testing R² = 0.858 in SN3), maintained predictions within physically reasonable bounds, particularly in degraded scenarios like SN4. Overall, the results confirm CNN’s exceptional ability to capture complex, nonlinear streamflow dynamics under historical conditions, while also underscoring the limitations of purely data-driven approaches when hydrological stationarity is violated providing crucial insights for deploying deep learning models in semi-arid, human-influenced basins under climate uncertainty.