Abstract: In response to the challenges of decreasing accuracy in real-time electricity load forecasting caused by objective factors and the issues of lengthy model training time and weak generalization capabilities due to high-dimensional nonlinearity of electricity load, this paper proposes a load forecasting digital twin model based on transformer and active adversarial transfer learning. Firstly, we construct a digital twin framework for load forecasting that allows real-time updates of the sample set and model. Next, we train the transformer-based load forecasting model using original historical load samples, and update the sample set and model in real-time with changing objective factors. We employ variational adversarial active transfer learning to select and train real-time load samples not yet learned, which saves model training time and enhances model generalization capabilities. Finally, we validate the effectiveness of the proposed method using electricity load data from a city.