Abstract: Accurate short-term load forecasting can guarantee efficient energy management and reliable operation for electric power companies. With the extensive application of information and communication technology in the power system, the available data of power system increases day after day, providing enough data for accurate power load forecasting. However, these data are usually lack of structure and obvious features. To solve the problem, a short-term power load forecasting method based on similar day and SAE-DBiLSTM model is proposed. Firstly, the electric power data is preprocessed and the power consumption features are described by similar day, base day load data and weather information, abstract features are further extracted using stack self-coding network(SAE). Secondly, these abstract features are input into the deep bidirectional long and short-term memory network(DBiLSTM) for training of the prediction model. Finally, the performance of the proposed model and several other models are compared, including DBiLSTM, SAE-ELM, SAE-DGRU, SAE-DLSTM and SAE-DBiLSTM, using the dataset of 2016 National College Student Electrical Mathematical Modeling Contest. Experimental results show that the proposed SAE-DBiLSTM model achieved the highest prediction performance in different regions. Moreover, the proposed model is simple and reliable for forecasting short-term regional power load.