<p>The electricity consumption behavior of urban residents is greatly influenced by temperature, especially during the high temperatures of summer and the cold of winter. Therefore, temperature is a crucial factor in short-term power load forecasting. However, the accuracy of power load forecasting is affected not only by external factors but also by the construction of prediction models and the length of the forecasting window. In this paper, we propose an incorporating temperature with parallel short-term power load forecasting method. Firstly, the dataset is divided into three intervals based on temperature factors, representing different seasonal cyclic variations. Secondly, the LSTM-Attention model is employed to transform daily predictions into parallel moment predictions, and finally, the predicted results are merged temporally. The proposed method exhibits strong performance, as evidenced by experimental results, with a maximum coefficient of determination (<i>R</i>²) of 0.91 achieved for short forecasting steps. In comparison with alternative models, the proposed method demonstrates notable reductions in root mean square error by 786.787&#xa0;MW, mean absolute percentage error by 3.105%, and mean absolute error by 478.745&#xa0;MW. These results indicate that setting different prediction models based on the seasonal variations of the dataset facilitates the exploration of data characteristics. Additionally, predicting data at the moment level effectively reduces cumulative prediction errors. The comprehensive performance of the proposed method outperforms direct prediction models such as LSTM-Attention.</p>