Abstract: In the light of the problems of strong randomness of current power load data, complex influencing factors and low accuracy of traditional single forecasting model, combining the different advantages of convolutional neural network(CNN), bi-directional gated recurrent unit(BiGRU) and attention mechanism(Attention) in short-term load forecasting, a hybrid CNN-BiGRU-Attention based prediction model is proposed. First the initial features of historical load and meteorological data is extracted through CNN, then time series association of feature data is further excavated by BiGRU. Then attention mechanism is introduced to give different weights to the output status of BiGRU, strengthen key features. Finally, load prediction is completed. The experimental results show that the mean absolute percentage error(MAPE), root mean square error(RMSE) and R-square(R²) of the model are 0.167%, 0.057% and 0.993, respectively, and the three indicators are significantly better than other models, with higher prediction accuracy and stability. The advantages of this proposed model in short-term load forecasting are verified.