<p>Climate change and intensifying human activities pose increasing challenges to streamflow simulation and forecasting. Traditional single models have limitations in capturing the characteristics of different temporal scale components within runoff sequences. This study proposes a trend-based differentiated modeling framework that integrates Variational Mode Decomposition (VMD) with Mann–Kendall trend testing. The framework first decomposes runoff series into multiple subsequences using VMD, identifies subsequences with significant trends through Mann–Kendall testing, and then matches prediction models to subsequences’ characteristic: SARIMA models are applied to subsequences without significant trends to capture linear temporal dependencies and seasonal patterns, while hybrid LSTM models (CNN-LSTM and Attention-LSTM) are applied to subsequences with significant trends to extract nonlinear dynamic features. Application to monthly runoff data from Dongting Lake Basin (1958–2011) shows that the VMD-CNN-LSTM-SARIMA hybrid model outperforms single models at all stations. The model achieves NSE &gt; 0.91, R2 &gt; 0.91, and KGE &gt; 0.88. Compared to the worst-performing single model at each station, the hybrid model improves NSE by 54.9–261.2% (from 0.263 to 0.950 at Xiangtan station), while reducing RMSE and MAE by 63.76–88.03% and 63.83–87.04%, respectively. Robustness analysis shows that VMD hybrid models exhibit good stability, reduce information complexity, maintain learning capability under data-limited conditions, and narrow prediction uncertainty intervals. The proposed framework, based on trend-based differentiated modeling and integrating strengths of multiple approaches, demonstrates improved prediction accuracy and robustness for runoff simulation under changing hydroclimatic conditions. The framework can be extended to other watersheds with appropriate calibration of decomposition parameters and assessment of hydroclimatic regime compatibility.</p>