An effective dimensionality reduction approach for short-term load forecasting

Abstract

Accurate power load forecasting has a significant effect on a smart grid by ensuring effective supply and dispatching of power. However, electric load data generally possesses the characteristics of nonlinearity, periodicity, and seasonality. For complex electric load systems, the presence of redundant information potentially reduces the real pattern extraction for load forecasting. Bearing in mind these issues, we propose an effective forecasting model in which a feature extraction module is introduced that is combined with the variational mode decomposition (VMD) with the variational autoencoder (VAE). In this combination, VMD is utilized for decomposing complex load series and VAE is used to filter the redundant information (noises) from each decomposed series. With two real data sets from China, we demonstrate that the proposed model can achieve highly accurate predictions, as we find the mean absolute percentage error (MAPE) values for one-step-ahead prediction to be 1% (Nanjing) and 0.8% (Taixing), respectively.