Abstract: Accurate short-term load forecasting can ensure the safe,reliable,economic and efficient operation of power system. Traditional forecasting methods cannot meet the requirements of high-precision load forecasting,while machine learning algorithm is widely used as the accurate scheme in short-term load forecasting. This paper presents a short-term load forecasting method based on attention mechanism（ i. e. Convolutional Neural Network and Long Short-Term Memory model）. The proposed method not only helps reduce the loss of historical information,but also realizes short-term load forecasting. Considering the seasonal characteristics of power load value,the forecasting method is designed to carry out short-term power load forecasting in different seasons. Finally,with load of a certain region in China as data sources,this forecasting method is compared with other commonly used forecasting models. The experimental results show that attention mechanism-based CNN-LSTM model has higher forecasting accuracy in different seasons.