Traditional CNN, BIGRU and Attention models have significant advantages in feature extraction, time-dependent capture and key information focusing, however, these models often suffer from the disadvantage of insufficient prediction accuracy in short-term power load prediction. To this end, this study proposes a CNN-BIGRU-Multihead-Attention model optimization method based on the Improved Artificial Gorilla Optimization Algorithm (AIGTO). The global search capability and optimization performance of AIGTO are improved by introducing SBOA search, Gaussian variation, non-uniform variation and differential variation strategies. By applying AIGTO to model parameter optimization, the AIGTO-CNN-BIGRU-Multihead-Attention model with high prediction accuracy is constructed, and based on the significant seasonal characteristics of electric load, representative spring and summer load data are selected to carry out simulation experiments, and the model's prediction performance is systematically evaluated under different load modes. The results show that the model proposed in this paper can be used to predict the load in different load modes. The results show that the proposed model has certain advantages in terms of prediction error and accuracy.