Flood forecasting has been a major challenge in hydrology for decades. A variety of approaches have been developed, including numerical models and data-driven models, such as Deep Learning (DL), for disaster early-warning and monitoring systems. This study proposed two combination models, exploring different feature separation techniques to improve the 12-hour multi-step discharge prediction. Specifically, two different approaches were implemented, namely the separation of wet and dry periods (Combination Model No. 1; CM-1) and the separation of flood and non-flood events (Combination Model No. 2; CM-2) for comparison with the baseline model trained on an entire dataset. Multivariate time series discharge data were used, measured at upstream and downstream stations within the same river basins. Experiments were conducted in three case study areas in Thailand: the Lower Loei River Basin, the Upper Nan River Basin, and the Upper Chi River Basin.In the proposed model combination approach, three deep-learning models were investigated: vanilla recurrent neural networks (RNNs) and their variants, long short-term memory networks (LSTMs), and gated recurrent units (GRUs). In general, CM-2 exhibited superior performance compared to the other models in most of the experimented neural networks, while the CM-1 approach outperformed the baseline model. This emphasized the advantages of our proposed feature separation framework, particularly in improving model performance, with respect to flood events. However, some variations were observed in the Upper Chi River Basin because the more complex models, such as LSTMs and GRUs, possibly overfitted the training set with less fluctuating data. Our study highlighted the significant contribution of combining two models through feature separation, tailored to specific periods, thereby enhancing the accuracy of flood predictions.