Abstract: Efficient and accurate short-term load forecasting is essential to the safe and stable operation of the

power system. A combined prediction model based on empirical mode decomposition（EMD）and gated recurrent

unit（GRU）was proposed. First，the date factors ，meteorological factors and historical load factor were selected to

construct the input feature set；then the EMD algorithm was used to decompose the strong random history load data into a limited number of inherent modal function components and trend components with different features，together with date factors and meteorological factors as the input of the GRU. Double-layer GRU recurrent network design was adopted to increase the depth of the network to enhance the model learning ability，and each component data was separately predicted and superimposed to reconstruct the output prediction value. Taking the load data of a certain place in China as an actual calculation example，the experimental results illustrate that the prediction error of this method is only 6.11%，which is greatly improved compared with the GRU network model and the BP neural network model. Compared with the EMD-LSTM network model，when the prediction accuracy differs by 0.04%，the prediction time is shortened by 25.99%，and the training efficiency is significantly improved.

      Key words: short-term load forecasting；time series；empirical mode decomposition（EMD）；gated recurrent

unit（GRU）