In order to cope with the uncertainty of renewable energy utilization and customer loads, a multi-timescale prediction method is proposed, where the prediction process is carried out in three phases: day-ahead, intra-day rolling and real-time, with timescales of 1 h, 15 min and 5 min, respectively. First, a prediction method based on difference statistics is used to accomplish the three stages of forecasting meteorological parameters. Second, a regression prediction model combining signal decomposition and machine learning is established for the day-ahead and intraday stages of load prediction, and a machine learning time series prediction model is established for the real-time stage. Next, the best prediction methods for typical daily loads in the day-ahead and intra-day rolling stages are determined based on the prediction accuracy metrics of the test set. Finally, the prediction method is applied to the energy forecast of typical days to verify the feasibility of the method. The results show that the determination coefficient R² of the prediction results of the meteorological parameters for a typical day in all three phases is above 0.8; in the day-ahead and intraday rolling phases, the prediction tasks of multivariate loads should be performed with different signal decomposition methods, and the R² of the load prediction results in the real-time phase is above 0.9, and the mean absolute percentage error (MAPE) is close to 0.