Abstract: Edge computing is extensively concerned by the energy industry because of the advantages of fast data processing, low cost and high real-time, and the prediction on the edge server is helpful for the refined management and control of energy. For this reason, in allusion to the limitations of edge service resources, based on difference autoregressive integrated moving average (ARIMA) model and gated recurrent unit (GRU）neural network a combined prediction method of wind, light and load was proposed. Firstly, the ARIMA was used to extract the linear characteristics of source and load, and through fitting linear characteristics with the true value the residual with nonlinear features was obtained. Secondly, taking the residual as the training dataset of GRU a prediction model was established, and then leading in the pruning and quantification method the GRU model was optimized and compressed to reduce the size of the prediction model to suit the deploy of edge servers. Results of lots of simulation examples show that the constructed GRU compression model possesses the features of small scale and high prediction accuracy, so it is suitable to the deployment and application of edge servers.