Abstract: Tropical cyclone (TC) size forecasting is currently an important and challenging aspect of operational forecast. This study analyzes the regional typhoon numerical prediction system (CMA-TYM) of CMA Earth System Modeling and Prediction Center for TCs in the Northwest Pacific Ocean during 2020-2023, focusing on the impact of initial vortex size errors on the performance of later size and intensity predictions of TCs. The analysis indicates that, despite the accurate initial center position and intensity, CMA-TYM model’s initial vortices display substantial errors in both inner-core and outer-core sizes. In 47% of samples, errors in the radius of maximum wind (RMW) could be more than twice as large as observations, while R33 and R26 are also overestimated. In contrast, errors for the radius of gale-force wind (R17) and asymmetry are comparatively smaller. Initial size errors significantly influence subsequent size predictions: Larger initial vortex size errors require longer spin-up times for adjustment and are often associated with larger forecast errors in later stages. After initial adjustments, R17 exhibit a strong lagged correlation (greater than 0.6) with subsequent forecast sizes for up to 48 hours. It indicates a significant influence of initial size on subsequent size changes. It’s found that larger RMW size errors adversely affect subsequent intensity forecasts, especially during TC intensification stage. Specifically, a larger initial RMW results in a weaker rate of intensification. Additionally, large initial errors in RMW and R17 typically result in subsequent size forecasts. However, CMA-TYM model can effectively simulate and predict the relationship between size and intensity, indicating that the model can accurately describe TC changes, but the initial size error significantly impacts its predictions. Based on this relationship, a machine learning model, XGBoost, is used to develop an R17 size correction scheme that incorporate initial and forecast intensity, inner-core and outer-core sizes, and initial errors as predictors to estimate and correct model-predicted size errors. Results indicate that root mean square error for 24-hour R17 forecast decreases from 59.8 km to 31.8 km, representing a reduction of 46.8% and significantly enhancing forecast accuracy. The scheme also demonstrates strong performance for both 30-hour and 36-hour forecasts. Approximately 79% of TC forecasts shows an improvement in size error following adjustments, highlighting the practical effectiveness of correction scheme.