<p>Accurate flood forecasting is of essential importance for preventing floods and mitigating disasters. Flood forecasting models based on neural networks have gained increasing popularity among hydrological researchers. Nevertheless, these models usually have three main disadvantages. Firstly, these models often lack targeted feature extraction mechanisms. Secondly, loss functions adopted in these models focus on the average level of prediction errors across all samples and assign equal weights to each sample, failing to highlight the importance of peak flows. Finally, the majority of existing flood forecasting models produce deterministic outputs, making it difficult to quantify forecast uncertainty. To overcome these limitations, this paper proposes a Synergistic Sequence-to-Sequence (Seq2Seq) and Peak-enhanced Probabilistic model for flood forecasting (simplified as SSPP). Within the Seq2Seq framework, the Encoder’s synergistic mechanism enables complementary extraction of hydrological features. The diffusion model bridges the Encoder and Decoder, quantifying hydrological uncertainties. The peak-enhanced loss function dynamically assigns greater weight to peak errors. Comparative experiments are evaluated with various advanced models that are commonly applied in flood forecasting. The results demonstrate that, for deterministic forecasting, SSPP attains the highest predictive precision and the most precise peak flow fitting among the chosen models. For probabilistic forecasting, SSPP achieves a narrow interval while holding high interval coverage within a 90% prediction probability interval, exhibiting more reliable forecasting performance.</p>