Graph data-driven power flow model for offshore wind farm considering internal and external characteristics

Purpose Existing data-driven power flow modeling methods for offshore wind farms (OWFs) are only applicable to fixed external grid scenarios and ignore the imbalance of power flow samples, which severely restricts the fitting accuracy and application range of the model. To solve these defects, this paper aims to propose an optimized data-driven OWF power flow modeling method. It is committed to adapting to dynamic changes of internal and external operating conditions and solving the accuracy limitation caused by unbalanced samples, so as to improve the generalization and prediction accuracy of power flow prediction in complex working conditions. Design/methodology/approach This paper extracts hidden grid impedance information from historical power flow data to construct a new data set integrating internal and external characteristics of OWFs. Combining graph deep learning advantages and wind farm structural features, a multi-head graph attention network with Huber loss function is established for power flow fitting. Comparative verification proves that the proposed method can achieve accurate power flow prediction under variable working conditions and unbalanced samples. Findings Impedance of power grid has a significant impact on power flow distribution in the OWFs. By extracting the grid impedance characteristics from the historical power flow data (PFD), a new PFD set considering internal–external characteristics is constructed. The data-driven model trained with this data set can improve the internal–external scenario adaptability of the model. To avoid the deviation of power flow model fitting caused by the imbalance of PFD samples and misjudge key information such as voltage over-limit in serious cases, a graph attention network model based on Huber loss function is proposed to raise the attention of some small samples and further improve the global prediction accuracy of the model. Originality/value This study innovatively breaks the limitation of fixed grid scenarios in traditional data-driven power flow models and effectively solves the low accuracy problem caused by unbalanced power flow samples. The constructed data set covering internal and external operating characteristics enhances the model’s environmental adaptability. The introduced Huber loss-based graph attention network optimizes the feature extraction capability of wind farm topological structures. It provides a high-precision, strong-generalization data-driven modeling scheme for OWF power flow prediction and offers reliable technical support for the safe operation and grid-connected dispatching of wind power systems.