[Objective] This study aims to develop an integrated method combining dynamic roughness calibration and post-simulation water level error correction to improve the accuracy of water level simulation in reservoir operation inversion. The core focus is water level error prediction based on measured data fitting, with the auxiliary of revealing the coupling mechanism between discharge error and water level error, and the ultimate goal is to provide reliable technical support for the post-event review and optimization of reservoir flood control and beneficial operation schemes. [Methods] The method consists of two key parts: discharge-classified dynamic roughness calibration and data-driven water level error correction. First, an adaptive iterative optimization method for roughness calibration was proposed. The Preissmann four-point eccentric implicit scheme was used to discretize and solve the Saint-Venant equations. Second, on the basis of roughness calibration, Linear Regression (LR) correction model and the Proportional Scaling (PS) correction model were employed to eliminate residual systematic errors. The n-fold cross-validation method was adopted to evaluate the correction effect, with Mean Absolute Error (MAE) and Nash-Sutcliffe Efficiency coefficient (NSE) as the core evaluation indicators. Hourly measured hydrological data from April to October in 2022-2024 at Panzhihua Station and Sanduizi Station in the lower Jinsha River were used for model verification. [Results] (1) The discharge-classified dynamic roughness calibration showed good convergence: the mean absolute error of water level at each station tended to be stable after about 8 iterations, and all roughness parameters converged after 14 iterations, with the calibrated roughness values able to accurately reflect the dynamic change characteristics of roughness with flow regime, and the simulated water level was in good agreement with the measured water level after calibration, which effectively reduced the systematic deviation caused by fixed parameters.(2) Both correction models significantly reduced the residual water level error, and the LR model performed better than the PS model with more stable results. After LR correction, the MAE of water level simulation at the two stations in 2022-2024 was reduced by about 60% on average compared with the benchmark model (only roughness calibration), with the maximum MAE reduction rate reaching 71.1% at Sanduizi Station in 2022, and the NSE was improved to more than 0.989 for all years, with the maximum NSE reaching 0.996. The PS model also achieved obvious error reduction, with MAE reduced by 30%-66% and NSE above 0.973, but its improvement amplitude was slightly lower than that of the LR model.(3) The box plot of water level error showed that both correction models significantly narrowed the error range and greatly reduced extreme errors, and the error distribution of the LR model was more concentrated. The regression coefficient of the LR model was extremely close to the scaling coefficient of the PS model, indicating that the proportional relationship through the origin dominated the discharge-water level error coupling after refined roughness calibration, and the intercept term of the LR model could absorb the tiny residual systematic deviation, making it more robust in prediction. The seven-fold cross-validation results verified the good generalization ability of the proposed method in practical engineering. [Conclusions] The proposed integrated method for water level simulation takes water level prediction as the core with clear physical interpretability and strong engineering practicability. Application results in the lower Jinsha River demonstrates its accuracy and stability. The innovative points are reflected in two aspects: 1) breaking the limitation of regarding roughness as a constant value by proposing a discharge-classified dynamic roughness calibration method with the objective of unbiased grouped water level error, and 2) constructing a data-driven water level error correction model based on the inherent error coupling relationship of the hydrodynamic model, which maintains the physical interpretability of the original model while realizing refined error correction.