Short-term residential household reactive power forecasting considering active power demand via deep Transformer sequence-to-sequence networks

Abstract

Recently, residential households such as consumers and new prosumers need increasing reactive power when charging electrical devices, exploiting flexible resources, and integrating renewable energy resources. Operators should supply enough reactive power by identifying influential factors and generating accurate forecasts for reactive power demand. Therefore, the scientific community concentrates more on reactive power consumption and brings increasing technical works, but related publications are still far from expectations compared with active power forecasting publications. In this context, this paper proposes a short-term reactive power forecasting scheme for residents. First, we reveal the uniqueness of the short-range residential household reactive power forecasting problem in distribution systems. Then, we customize an attention-based Transformer sequence-to-sequence network (Seq2Seq Net) to capture the volatility and uncertainty of reactive power sequential readings and improve the pressing accurate residential reactive power demand forecasts. The customized Transformer Seq2Seq Net can take advantage of the observed, known and static covariate factors, where the encoder-decoder architecture captures short-term memory, the attention mechanism captures long-term temporal dependencies, and gating residual connections augment the non-linear processing capability. In addition, we model the interactions between active and reactive power demand by incorporating a multi-task learning structure. Finally, we verify the effectiveness of the proposed solution by comparing hour- and day-ahead forecasts generated by the solution, benchmarks (i.e., one-hour, one-day, and one-week hold), comparative data-driven methods including machine learning (e.g., gradient boosting decision tree) and deep learning (e.g., temporal convolutional networks), and ablative models (that considers varying conditions in the scheme). The results showed that the hour-ahead forecasts generated by the proposed solution can reduce up to 188.2% error among multiple evaluation criteria, such as mean absolute error (MAE), root-mean-square error (RMSE), mean absolute percentage error (MAPE), normalized mean absolute percentage error (NMAPE), and Theil inequality coefficient (TIC), and that the day-ahead forecasts decrease up to 273.2% error, with the acceptable time costs on optimizing the model parameters and producing forecasts, illustrating the feasibility of the scheme.