Abstract: Accurate short-term load forecasting can alleviate the contradiction between power supply and demand, coordinate load management and ensure power system security. Short-term power load had strong nonlinearity and time dependence, and was related to many external factors such as climate change and real-time electricity price, which made it difficult to accurately predict short-term power load. Therefore, we proposed a short-term power load forecasting method based on TPA（temporal pattern attention） mechanism of CNN（convolutional neural network）-BiGRU（bidirectional gated recurrent unit）. Among them, the convolutional neural network was used to mine the nonlinear spatial relation between different input variables and the current load, bidirectional gated recurrent unit was used to capture long-term dependencies from time series, and temporal pattern attention mechanism was introduced to assign feature weights, thus highlighting important information and achieving short-term load prediction. Taking the Australian public load data set as an example to verify the effect of the proposed method, the prediction accuracy reached 97.651%, and comparing with the CNN-LSTM and RNN prediction models, the accuracy was increased by 0.531% and 5.992% respectively. The experimental results show that the proposed TPA-CNN-BiGRU prediction method has higher prediction accuracy and universality.