Abstract: The power load forecasting method based on sequence decomposition can improve its forecasting accuracy, but it also brings error accumulation. At the same time, the existing methods ignore the correlation between the historical forecast errors and the current forecast results. A short-term load forecasting method based on differential decomposition and error compensation (DD-EC-GRU) is proposed. First, the first-order difference of the original load sequence is used as an input feature to transform the load forecasting problem into a load's variation forecasting problem. Based on this, multiple sets of fake load sequences are introduced on the basis of a set of actual load sequences. The gated recurrent unit is used to construct a multi-objective iterative prediction network. Finally, comprehensively considering the changing trend and stationarity of the iterative prediction errors of each sequence, an error compensation network based on the sequence similarity and artificial neural network integrated model is constructed to improve the prediction accuracy. The effectiveness of each components of DD-EC-GRU is verified on three actual load power datasets. Compared with various popular algorithms, the DD-EC-GRU has higher prediction accuracy and stronger adaptability.