A Transfer Learning Combination Model for Annual National Electricity Consumption Forecasting in China

Accurate forecasting of annual national electricity consumption data is crucial for energy and economic development planning. However, the small amount of data and the complex influencing factors pose great challenges. Here we propose a novel transfer learning combination model for the 5-year forecasting of annual national electricity consumption data in China. To improve the forecasting accuracy, we adopted a data transfer learning approach to extend the training set of China by migrating relevant data from 17 developed countries. To increase the data information of training sets, a new data preprocessing procedure containing trend calculation and data extension was designed. To fully utilize the advantages of different models, a multi-model integration framework with a neural network weighting unit is designed. The comparison results show that the proposed model has the lowest errors, with the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) of 0.0105, 0.0002, and 0.3047, respectively.