Abstract: Under the dual influence of climate change and human activities, the rainfall—runoff relationship in river basins exhibits non-stationary characteristics, and hydrological models based on the traditional steady-state assumption struggle to capture the structural transitions of the system, leading to a decline in runoff simulation performance. In this paper, hydrological states are used as the representation of non-stationarity, and hidden Markov models (HMMs) with 1 to 3 states are constructed, the Viterbi algorithm is adopted to identify the hydrological states of the river basin and their evolution paths, and the decoded hydrological state results are introduced into the runoff simulation process to address the impact of non-stationarity. Experiments are conducted based on 240 typical river basins in the CAMELS database, and the results show that: approximately 22.5% of the river basins have significant multi-state hydrological characteristics; the multi-state model with hydrological state constraints shows significant advantages in both probabilistic simulation and deterministic simulation, with the uncertainty interval coverage rate increased by an average of about 30% compared with the single-state model; the Nash-Sutcliffe efficiency coefficient increases from 0.45 to 0.83 (an increase of about 84%), and the root mean square error decreases by an average of about 49%. Research shows that: the HMM-based hydrological state identification method can effectively capture hydrological state transitions, and incorporating state information into the simulation process helps improve runoff simulation performance under non-stationary conditions, providing a new path for hydrological simulation in a changing environment.