In the current context, accurate prediction of meteorological elements plays an increasingly important role in agricultural production, social life, and transportation. Therefore, an improved SCSO (sand cat swarm optimization) algorithm gets proposed to optimize the CatBoost model, addressing the issue of inaccurate traditional temperature and wind speed predictions. Meteorological data from the Nanjing area covering the period from January 1, 2012, to December 31, 2014, was utilized, with ERA5 reanalysis data serving as ground truth. The data was divided into training and validation sets. The SCSO optimizes the CatBoost model for correcting temperature and wind speed forecasts at 24, 48, and 72-hour intervals. To overcome issues of SCSO easily falling into local optima and slow convergence, the Halton Sequence search algorithm initializes the sand cat swarm positions. Levy flight and triangle walk strategies are introduced to optimize the search process. During iterations, the LOBL strategy and boundary mutation operator were employed to ensure avoidance of local optima. Results indicate that the improved SCSO-CatBoost model exhibits higher accuracy and superiority compared to XGBoost, LightGBM, traditional GBDT, Random Forest, Support Vector Machine, and Linear Regression models. Root mean square errors for temperature and wind speed predictions at 24 hours stand at 0.514 5 and 0.174 9, respectively. Significant improvements occur at 48 and 72 hours as well. This study provides a scientific basis and technical support for enhancing the accuracy of meteorological element forecasts.