Climate change poses significant challenges to water resource management in arid catchments, where reliable projections of hydrological extremes remain difficult to obtain. This study applies a hybrid deep learning framework for downscaling and a hybrid HEC-HMS-Long Short-Term Memory (LSTM) approach for streamflow simulation to assess climate change impacts in an arid region. Specifically, it evaluates projected streamflow changes at Oman's largest dam using advanced deep learning models-LSTM, Temporal Convolutional Network (TCN), and Gated Recurrent Unit (GRU)-driven by CMIP6 climate projections under Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, and SSP5-8.5). The HEC-HMS model demonstrated strong performance (NSE = 0.842 for calibration, 0.811 for validation) but failed to simulate peak events, such as Cyclone Gonu. Hybridization of HEC-HMS with LSTM significantly improved accuracy, with a peak flow error decrease to 3.60%. Downscaling with a hybrid LSTM-Transformer model outperformed the other models and improved precipitation projections. Prpjection of future climate showed decreasing precipitation under SSP1-2.6 (annual peak: 23.82 mm between 2080 and 2099), as opposed to increasing extremes (36.65 mm) in SSP5-8.5. Temperature rose across all scenarios, with summer temperatures reaching 37.4 degrees C under SSP5-8.5. Streamflow projections saw an 80% decline under SSP1-2.6 but a near-baseline return under SSP5-8.5 (243.37 m3/s).