In order to cope with the complexity and uncertainty brought by the coordination of load and storage in the new power system, this paper proposes an intelligent scheduling optimization algorithm based on multi-source heterogeneous data fusion. This algorithm achieves semantic alignment between structured and unstructured data through knowledge graph (KG), and utilizes multi head attention mechanism to dynamically allocate feature weights, effectively improving the quality of data fusion. At the level of decision optimization, the algorithm constructs a graph reinforcement learning (GRL) model, embeds the topology information of power grid into the state space, and designs a multi-objective reward function that takes into account economy, safety and new energy consumption. In order to protect cross-regional data privacy, the algorithm introduces federated learning mechanism and realizes collaborative optimization through parameter sharing. The experimental results show that the algorithm performs well in IEEE 118-bus system and actual regional power grid model, and the system state prediction accuracy is significantly improved compared with traditional methods. The average daily dispatching cost is reduced to 1.159 million, the rate of abandoning wind and light is reduced to 3.1%, the number of crossing the line is only 0.3 times/day, and the decision-making time is about 38ms, which meets the real-time dispatching requirements. In addition, the federated learning framework makes the model achieve similar performance to centralized training without sharing the original data, showing good generalization ability and privacy protection characteristics.
Chen et al. (Sun,) studied this question.