Real-Time Locational Marginal Price Forecast: A Decision Transformer-Based Approach

The locational marginal prices (LMPs), which are jointly determined by locational demands and generation bids, are difficult to be forecasted from the market participants’ perspective. In this paper, a special neural network (NN) architecture named the decision transformer, is proposed to forecast real-time LMPs by learning temporal correlations among historical market data sequences. LMP forecasting problem is formulated as the problem of learning future actions in a model consisting of explicit locational demand data and implicit bid data. The LMPs are the consequence of generators’ strategic bidding behaviors, which are considered the result of a decision-making process based on the knowledge of historical market data and current fuel prices. The transformer NN contains a sequence-to-sequence architecture, which can learn temporal correlations by paying different attention to market data at different timestamps. Case studies on Southwest Power Pool (SPP) and ISO-New England (ISO-NE) datasets validate the performance of the method.