Distributionally robust optimization dispatch strategy for virtual power plants based on data generation augmentation

The newly constructed renewable energy-oriented virtual power plants (VPPs) suffer from a scarcity of operation data for renewable energy and the uncertainties from both generation and load, which severely impact on the reliable operation and economic dispatch of VPPs. To address this issue, this paper proposes a distributionally robust optimization (DRO) scheduling strategy for VPPs based on data generation augmentation. Firstly, to mitigate the problem of operation data scarcity, this paper integrates the physical model of photovoltaic (PV) with deep learning methodology integrating convolutional neural networks (CNN) and multi-head attention mechanism. Historical meteorological data are utilized to calibrate the parameters of the physical model, thereby enhancing the accuracy of data generation. Secondly, considering the dual uncertainties of sources and loads, a DRO dispatch model for VPP is established based on the Wasserstein distance. Furthermore, this paper proposes to reformulate the dispatch model into a tractable mixed-integer linear programming problem by leveraging linear decision rules and strong duality theory. Finally, simulations verify that the proposed strategy demonstrates advantages in data generation accuracy and robustness under different meteorological conditions. Moreover, the DRO dispatch method can adapt to the dynamic variability characteristics of PV power and load power, thereby reducing the operating costs and simultaneously enhancing operating reliability of VPPs.