A data-physical fusion method for economic dispatch considering high renewable penetration and security constraints

• A physics-informed reinforcement learning method is proposed for economic dispatch. • An invertible safety layer is designed to ensure action feasibility in training. • A two-stage framework integrates the penalty method and Lagrangian relaxation. • Physical gradients improve model transfer and reliability under limited data. Economic dispatch plays a pivotal role in ensuring the safety and reliability of new power systems. Conventional model-based methods of economic dispatch encounter significant challenges due to the increasing uncertainties brought about by high renewable penetration. Reinforcement learning (RL) has shown great potential because it does not depend on probability distribution modelling of uncertainties. However, the disparity between training scenarios and real-world applications hinders the broader adoption of RL. To address these challenges, this paper presents a data-physical fusion method for economic dispatch considering high renewable penetration and security constraints. An invertible mapping is introduced in the action space of the Markov decision process model, comprising a balance layer and a reverse mapping. The invertible mapping can increase the exploration efficiency and guide the training quality of agents. Then, a physics-informed neural network based on the deep deterministic policy gradient algorithm is employed for fine-tuning and transfer learning. Physical models of unsatisfied constraints are integrated into the critic using Lagrange relaxation. The case studies demonstrate the superior performance of the proposed method in terms of both training speed and decision-making reliability compared to conventional data-driven methods. The fusion of physical models and neural networks enhances the interpretability of data-driven methods, thereby facilitating their broader application in new power systems.