A short-term load forecasting method based on load characteristics clustering and elastic net analysis is proposed in this paper. By analyzing and clustering the historical load characteristics, the annual days are classified and its clusters are specified, and the lack of pertinence of the types of the day cluster selection is avoided. At the same time, Elastic net analysis is adopted to identify and select the dominant factors for short-term load forecasting. Furthermore, the neural network forecasting model is established on the basis of input variable optimization. Taking the actual load of a city in Guangdong province as an example, the effectiveness of the proposed method in improving the daily load curve forecasting accuracy is verified by comparing with other methods. Results show that the model established is long-term effective, dispensing with repeated training, which is applicable for short-term load forecasting.