Ultra-short-term power load forecasting is crucial for rapid response and real-time dispatch in a power system. Accurate load forecasting ensures the safety of the power system and improves electricity efficiency. To obtain accurate and reliable load forecasting results, a new ultra-short-term power load forecasting method based on CNN-BiLSTM-Attention (AC-BiLSTM) is proposed for the characteristics of nonlinear and time-series nature of grid load data. First, a convolutional neural network (CNN) and bidirectional long and short-term memory (BiLSTM) networks are used to extract the spatio-temporal features of the load data. The attention mechanism automatically assigns corresponding weights to BiLSTM to distinguish the importance of different time load sequences. These can effectively reduce the loss of historical information and highlight the information of key historical time points. Finally, the final load prediction results are output through the fully connected layer. Taking the real load data of a certain area as an example, the comparison between two experimental scenarios proves that the proposed method has high prediction accuracy and can provide a reliable basis for power system planning and stable operation. This work is supported by the Science and Technology Planning Project of Henan Province (No. 212102210226).