Two-Layer Transfer-Learning-Based Architecture for Short-Term Load Forecasting

In this paper, by utilizing corresponding load data of source zones, a two-layer transfer-learning-based short-term load forecasting (STLF) architecture is proposed to improve forecasting accuracy of load in target zone. In the inner layer, the latent parameter is introduced to represent the latent factors that results in the differences in electricity consumption behavior between different zones. With the latent parameter as extra input, a latent parameter-assisted model suitable for both the load data of target zone and all source zones is built. Variant weights are assigned to datasets according to their fitness to the latent parameter-assisted model. To solve the weights, an iterative algorithm is developed in the outer layer. Case studies demonstrates that the proposed STLF architecture always improves the forecasting accuracy of classic STLF algorithms, especially when the load data of target zone is insufficient.