In recent decades, climate change has altered the patterns and regimes of river flow, increased uncertainty in the hydrological behavior of basins, and posed many challenges for river flow prediction models. Therefore, this study develops and evaluates two deep learning models, LSTM and TCN, for river flow prediction and uncertainty estimation in the Shahroud basin. The data included river discharge, temperature, precipitation at the Lowshan station, and snow depth in the Shah Alborz Mountains. For estimating and decomposing uncertainty, the Gaussian Negative Log-Likelihood loss function was used to quantify aleatoric uncertainty, while the Monte Carlo Dropout method was used to estimate epistemic uncertainty. During the testing phase, the TCN model achieved an RMSE, R², and KGE of 8.7238 m3/sec, 0.9370, and 0.9495, respectively, whereas the LSTM model achieved corresponding values of 9.4268 m3/sec, 0.9264, and 0.9359. The evaluation of extreme event detection showed that the TCN model outperformed the LSTM model at high streamflow thresholds by producing fewer false negatives (FN), thereby enhancing flood detection capability, although this improvement came at the cost of more false positives (FP). The uncertainty analysis indicated that the LSTM model achieved PICP and MPIW values of 0.9055 and 15.0558, respectively, providing coverage closer to the target PI90, and the TCN model achieved corresponding values of 0.9329 and 20.4160, resulting in higher-than-target coverage for PI90. Overall, the TCN model outperformed the LSTM model in river flow prediction. In terms of uncertainty estimation, the LSTM model exhibited better calibration, whereas the TCN model provided more conservative prediction intervals with coverage exceeding the target PI90, making the TCN model potentially more suitable for risk-sensitive decision-making applications.