Forecasting streamflow reliably is a key component in planning water resources, particularly in areas sensitive to climate variability such as northern Algeria. This research explores the performance of four recurrent neural network models—GRU, BiGRU, LSTM, and BiLSTM—in predicting daily streamflow at two hydrological stations (Azzeffoun RN24 and Baraki). Each model was assessed in its conventional form and in a hybrid configuration enhanced through Variational Mode Decomposition (VMD), a signal processing method designed to extract intrinsic oscillatory components from complex hydrological series. The predictive accuracy was evaluated using four statistical measures: The R, NSE, MAE, and RMSE. The findings demonstrate that all models incorporating VMD performed better than their standard versions. At Azzeffoun, the VMD-BiLSTM model yielded an R of 0.990, an NSE of 0.981, and an RMSE of 0.158. In Baraki, the same model achieved an R of 0.989, an NSE of 0.977, and an RMSE of 0.879, confirming notable performance gains. This study is one of the earliest efforts to apply VMD-enhanced deep learning frameworks for hydrological prediction in Algeria. The results validate the utility of VMD in managing non-linear and non-stationary dynamics, thereby enhancing forecast precision. The study underscores the promise of such hybrid models for supporting water resource strategies and early warning systems in semi-arid regions.