Abstract: To evaluate the applicability of WRF-Solar driven by different reanalysis datasets for irradiance simulation at a photovoltaic power station in complex terrain, and to account for the uncertainty associated with different driving data sources, this study used the mesoscale model WRF-Solar v4.3 to conduct dynamical downscaling simulations for a photovoltaic power station in Yunnan Province. ECMWF-ERA5, NCEP-FNL, and CMA-RA were used as the initial fields, respectively. The simulation performance of the three driving datasets was systematically compared at different temporal scales, under different weather types, and across irradiance intensity bins. On this basis, a station-scale irradiance correction model based on LimiX was constructed. The results show that: (1) all three datasets can reproduce the basic annual variation in irradiance. The correlation coefficients of ERA5, FNL, and CMA-RA are 0.73, 0.71, and 0.65, respectively, and their RMSE values are 225.86, 226.60, and 245.00 W·m-2, respectively. ERA5 shows the best overall performance, whereas CMA-RA is relatively weaker in terms of correlation and comprehensive error metrics. (2) The three schemes show obvious seasonal differences. The correlation is higher and the error is lower in winter, whereas the error is most significant in spring. (3) In terms of weather types, the simulation performance is best under clear-sky conditions, followed by cloudy conditions, while the error increases significantly under rainy conditions. This indicates that cloud-related processes have an important influence on irradiance simulation accuracy. (4) The irradiance binning results show that the RMSE values of the three schemes peak in the range of 200-300 W·m-2, indicating that the errors are mainly concentrated in the low-to-medium irradiance range. The simulations also show a systematic bias characterized by positive bias under low irradiance and negative bias under high irradiance. (5) The structured LimiX correction model shows excellent station-scale irradiance bias correction performance. During the test period, ERMSE decreases from 36.58% to 29.27%, and R and IOA increase to 0.90 and 0.92, respectively. Overall, ERA5 shows the best stability and overall applicability as the WRF-Solar driving dataset for the photovoltaic power station. Combining ERA5-driven WRF-Solar simulations with LimiX post-processing correction can further reduce station-scale systematic bias and improve irradiance simulation accuracy. This study provides a reference for numerical simulation of photovoltaic irradiance, selection of driving datasets, and station-scale correction in complex terrain regions.