<p>Accurate streamflow forecasting is critical for water-resources management and disaster early warning. Combining seasonal–trend decomposition (STD) preprocessing with hydrological prediction models has become a prevailing approach. However, mainstream models such as LSTM, CNN, and GRU still struggle to capture the nonlinear dynamics of streamflow, and the effect of STD techniques on their predictive performance remains insufficiently explored. To address this gap, we introduce three cutting-edge architectures—FITS, FGN and PatchTST—into hydrology for the first time and benchmark them against the traditional LSTM, CNN and GRU baselines. Each of these six model classes is then paired with four STD techniques (MOV, LD, EXP and DFT-MOV), producing 24 hybrid models. We systematically evaluate all models at four streamflow gauging stations along the Jialing River. Without STD preprocessing, the FITS model achieved the highest predictive accuracy across the four stations (mean Nash–Sutcliffe efficiency, NSE = 0.986), followed by FGN (mean NSE = 0.984) and PatchTST (mean NSE = 0.981); all three also exhibited markedly greater stability than traditional models. Notably, FITS delivered these advantages at the lowest computational cost. Under a controlled computational budget, STD substantially improved the performance of LSTM, CNN, and GRU models (mean NSE increases of 0.12%–1.76%), whereas the gains for FITS, FGN, and PatchTST were limited (mean NSE changes of − 0.13%–0.52% ). These findings provide valuable insights for selecting and optimizing STD techniques and streamflow forecasting models.</p>