Accurate short-term power load forecasting plays an important role in ensuring the safe and stable operation of the power grid, optimizing energy management, improving the utilization rate of power generation equipment and reducing operating costs. Aiming at the problem of regional short-term power load prediction in univariate scenarios, a prediction method based on MA (moving average) and CNN-LSTM (convolutional networks long short-term memory networks) was proposed, and self-attention mechanism was added. Firstly, multiple sliding average was used to decompose the raw load data into multiple stationary series to reduce the noise and complexity of the data. Secondly, the one-dimensional sequence data was transformed into a multidimensional structure, and the CNN was used to extract the internal relationship between multiple time points. Then, data was input into the LSTM model for training, and weighted fusion was performed using the Self-Attention mechanism to improve the prediction accuracy. Finally, the predicted values of each series are added together to obtain the final load prediction value. In order to verify the effectiveness of the method, a prediction experiment was carried out on the real load data of the power grid interval of fifteen minutes in a certain region of China, and the prediction results were compared with other common model prediction results. Experimental results show that the accuracy of this method is higher than that of other methods in univariate short-term power load forecasting problems.