Deep Learning Based Bidding-Oriented Probability Density Forecasting Approach for Renewable Energy Generation in Virtual Power Plant

With the implementation of the peak carbon and carbon neutral strategy, more and more renewable energy sources (RES) will be gathered in virtual power plant (VPP), which is of great significance to the in-depth research on the forecasting of RES. Most of the existing research on day-ahead RES forecasting aims to construct robust models to enhance forecasting accuracy. The forecasted RES power is subsequently utilized as an input for renewable energy generation, with the overarching objective of maximizing the benefits obtained from the electricity market. Nevertheless, prevailing day-ahead RES forecasting models often do not integrate this overarching goal during the training or forecasting phase. This study proposes a bidding-oriented probability density forecasting approach for renewable energy generations. The bidding optimization problem is formulated as an additional layer of the forecasting model. The training loss of the deep neural networks is designed to maximize the bidding benefits of renewable energy generators. The overall gradient of the bidding benefits with respect to neural network parameters is derived with rigid proof. Numerical simulations verify the effectiveness and superiority of the proposed approach in enhancing the bidding benefits of renewable energy generations.