Day-ahead reactive power optimization of new power systems based on physics-informed graph attention network

• A GAT algorithm incorporating physical information of the power grid is proposed. • A unique optimal solution selection method based on the AHP-improved TOPSIS algorithm is proposed. • A day-ahead optimization strategy screening method that considers both objectives and risks is proposed. • The proposed algorithm demonstrates strong adaptability to different power grid topologies. • The proposed algorithm exhibits excellent scalability and significant potential for engineering applications. The inherent stochasticity and intermittency of renewable energy generation exacerbate voltage fluctuations in power grids, posing significant threats to their secure and economic operation. To address the day-ahead reactive power optimization problem under uncertainty, this paper proposes a novel stochastic reactive power optimization method based on a physics-informed graph attention network (GAT). First, the training labels for the GAT are generated by incorporating physical information derived from the power grid. Subsequently, the GAT model is employed to capture the complex nonlinear relationships between the states of reactive power regulation devices and the data of renewable energy sources and loads, thereby generating the optimal solution set. Building upon this, an optimal solution selection method based on the analytic hierarchy process (AHP)-improved technique for order preference by similarity to ideal solution (TOPSIS) algorithm is proposed. Finally, to extract the ultimate optimization strategy from the set of strategies generated under multiple stochastic scenarios, a screening method that balances both objectives and risks is proposed. Validations on a modified IEEE 39-bus system and an actual power grid demonstrate the proposed method's superior optimization performance and computational efficiency, confirming its suitability for day-ahead reactive power optimization applications.