Uncertainty reduction in power forecasting of virtual power plant: From day-ahead to balancing markets

Abstract Adjusting prediction data before bidding is a straightforward and cost-effective method to reduce uncertainty and imbalance between bidding data and real-time power supply. To avoid profit loss for virtual power plant, this study proposes an uncertainty optimization model that minimizes the remaining uncertainty after power generation forecasts. The proposed model specifically addresses different weather conditions—rainy, overcast, cloudy, partly cloudy, and sunny—by minimizing the distance between actual and forecasted VPP generation. The model is applied to adjust prediction data of a VPP with an average generation capacity of 1.5 GW in Tokyo, Japan. Bidding data for winter 2024 are predicted using three deep neural network-based methods. The results indicate a significant reduction in both uncertainty and energy storage capacity after using the uncertainty optimization model. Moreover, the findings show that the proposed uncertainty optimization model increases the profit growth rate for prediction methods characterized by high uncertainty.