A new runoff forecasting method based on dynamic identification of similar processes is proposed, based on similarity theory, aiming to improve the accuracy and adaptability of real-time runoff forecasting. By constructing a spatiotemporal feature index system for rainfall-runoff processes and applying a multi-timescale similarity search algorithm with sliding windows, the method dynamically identifies historical events most similar to the ongoing process. The future runoff trajectories of these identified events are then combined through weighted averaging to generate a composite forecast, which is further corrected and extended by dynamically aligning it with the current observed discharge. Using historical daily rainfall and runoff data from the basin upstream of Xiashan Hydrological Station in the upper Ganjiang River as a case study, a leave-one-out cross-validation approach is employed to evaluate model performance. The results show that the multi-timescale similarity search algorithm with sliding windows can fully utilize the key features of adjacent time periods. Even in the absence of rainfall forecasts, the method yields high prediction accuracy and reliability for 1-2 day lead times, effectively capturing both peak discharge and peak timing of the runoff process. Although forecast accuracy decreases for a 3-day lead time, the inclusion of rainfall spatial distribution indicators significantly enhances the predictive performance. Moreover, the method requires relatively low effort for parameter calibration, is insensitive to variations in feature weighting, and exhibits robustness. This approach offers an effective and innovative solution for short-term runoff forecasting and flood warning.