With the development of smart grid technology, short-term power load forecasting becomes more and more important. To improve the accuracy of short-term electric load forecasting for individual users, this study proposes a load forecasting model, which is based on K-means and Convolutional Neural Network (CNN). Firstly, K-means is applied to group users into two categories. For users with strong daily correlation, the historical loads of adjacent time points and same time points in adjacent days are taken as input to the CNN model to extract abstract features for prediction. For users with weak daily correlation, the historical loads of the adjacent time points are utilized as features. To assess the performance of the proposed method, we conducted comparison experiments on real data with random forest and support vector regression. The experimental results show that the MAPE of the proposed approach is reduced by 20%.