Abstract: Accurate short-term load forecasting is the key to ensure the smooth operation of power system. After the popularity of machine learning algorithm, it provides algorithm support for short-term and ultra short-term load forecasting which is difficult to solve before. In view of the good effect of the Catboost, the convolutional neural network-long short term memory, extratrees and other integrated models in processing nonlinear correlation data, this paper combines the above three methods to build an integrated prediction model, uses BP neural network to determine the weight coefficient, and combines the advantages of various single prediction models through the weight, so as to achieve a better prediction effect. In order to better illustrate the advantages of the method used in this paper, Mean Absolute Percentage Error (MAPE), root mean square error, mean square error, goodness of fit are used as measurement indexes. Compared with each single prediction model, the integrated model decreased by 1.01 percentage points, 0.94 percentage points and 1.19 percentage points respectively compared with Catboost, CNN-LSTM and Extratrees models in MAPE standard.

Key words: ultra short-term load forecasting； integrated model, Catboost model, CNN-LSTM model, Extratrees model