Historical data is indispensable in power load forecasting

however

many problems exist in the selected historical data

such as a large amount of data but little characteristic dimension

many invalid data

and unclear characteristic relationship among data

which significantly affect the accuracy of power load forecasting. In order to improve the accuracy of ultra-short-term power load forecasting

an ultra-short-term power load forecasting method based on the double-layer XGBoost（eXtreme Gradient Boosting） algorithm is proposed. The first layer of the method

the data processing layer

is based on the XGBoost algorithm and feature engineering to construct multiple weak learners to train layer by layer

and to filter out the feature set that has a significant impact on the power load; the second layer is the load forecasting layer. The feature set and load selected in the first layer are input

the hyperparameters of the XGBoost algorithm are optimized and the model is trained to obtain the load forecasting model with the highest accuracy and the smallest root mean square error. The built load forecasting model can avoid standardization of data features

and can reduce the impact of missing data fields

regardless of whether the features are interdependent

and the model learning effect is good. In the analysis of the calculation example

the load forecasting models based on single layer XGBoost

BP neural network and ARIMA are compared. The proposed method has higher prediction accuracy and has good generalization ability under different time periods of data sets.