Locational marginal price forecasting using Transformer-based deep learning network

Location marginal pricing (LMP) in a deregulated electricity market is the result of bidding by each node. A deep learning neural network is proposed for multistep-ahead LMP prediction with gap time periods at a specific node in a certain zone of the PJM power market. The proposed method is based on Transformer model which is a well-known sequence to sequence(seq2seq) architecture for deep learning that perform well on a variety of natural language processing tasks. The proposed method can use the covariates of the prediction time period to generate prediction results in parallel, rather than autoregressively generate results like canonical Transformer, which can avoid the error accumulation of the results. The experimental results show that the proposed method all achieves higher accuracy than conventional deep learning methods in different prediction steps.