Graph reinforcement learning with auxiliary temporal-graph convolutional neural network for unit commitment

In power systems, Unit Commitment is an important means of ensuring the safe operation of power systems. For large power grids, the solutions obtained in the past based on traditional algorithms such as metaheuristic algorithms and mathematical optimization algorithms are difficult to guarantee and deficient in terms of time complexity and safety. The method of predicting start–stop or dispatching schedules based on deep learning models can hardly fully satisfy the constraints. To address these issues, this paper proposes a Graph Reinforcement Learning with Auxiliary Temporal-Graph convolutional neural network method to learn the temporal and spatial characteristics of the power grid. The method interacts with the environment through a reinforcement learning agent based on a graph convolutional neural network, and makes economic and safe start–stop decisions based on the topology of the power grid and historical load data. The start–stop decision scheme is adapted by designing a reward function and a start–stop adaptation module to ensure that the power system constraints are met. The proposed method has been tested on the IEEE 30-bus and IEEE 118-bus datasets. The method proposed in this paper can solve the unit commitment problem quickly, effectively and safely. Compared with traditional models, the indicators of the model are optimal and the solution can fully satisfy the constraints. Compared with the commercial Mixed Integer Programming solver Gurobi, there is a slight difference in cost, but the solution time is effectively shortened, which is more suitable for large-scale power grids. • Considered power grid topology to capture its physical spatial patterns. • Applied graph reinforcement learning for decision-making under incomplete info. • Developed a priority list-based startup/shutdown module meeting constraints. • Integrated TCN and GCN for unit commitment optimization.