<p>Precise runoff prediction is essential for the effective management and utilization of water resources. However, effectively modelling highly nonlinear and nonstationary runoff processes remains challenging. To address this issue, this study develops advanced frameworks that incorporate a Kolmogorov–Arnold Network (KAN) alongside conventional deep learning (DL) models, including long short-term memory (LSTM) and gated recurrent unit (GRU), together with a dynamic postprocessing (DP) module for peak correction. The proposed framework captures temporal features through recurrent layers, enhances nonlinear representation through the KAN, and addresses the underestimation of runoff peaks through the DP module. With respect to the Chengbi River Basin, the model performance was evaluated across prediction horizons ranging from 1 to 7 days. The results show that both the KAN-GRU and KAN-LSTM models exhibit considerable accuracy, markedly surpassing their respective standalone models, with the NSE improving from 0.82 to 0.87 at the 1-day prediction horizon. Moreover, the DP module reduces the runoff peak underestimation, reflected in a reduction in the PPSD from more than 26% to approximately 16%, and an improvement in the NSE of 0.25 for the KAN-LSTM model and 0.20 for the KAN-GRU model under high-runoff conditions. Notably, the KAN-LSTM-DP model achieves the highest accuracy, with an RMSE of 24.55 m³/s, an NSE of 0.96, and a KGE of 0.88. This research presents an effective modelling strategy for short- to medium-term hydrological prediction, highlighting the role of incorporating a KAN and a DP module in improving the predictive performance of DL models.</p>