With the increasing complexity of integrated load structure of power grid, the diversity and volatility of the load increase significantly. Higher generalization ability and accuracy are required for the prediction model. However, the traditional algorithms have inherent defects such as easy overfitting and low accuracy, making it difficult to achieve accurate peak load forecasting under complex grids. To solve above problems, Bayesian optimized XGBoost model for short-term peak load forecasting is proposed. Firstly, the important features are screened through feature importance score to ensure better mapping relationship between input and output. Then, Bayesian optimization algorithm is introduced to determine the hyper-parameters to ensure the best state performance of XGBoost. Effectiveness of the proposed model is verified using power load data of a certain city in China. The results show that Bayesian optimized XGBoost has higher prediction accuracy compared with other machine learning methods.