A large model accelerated BERT distributed multi-objective economic dispatch method for novel power systems with large-scale renewable energy sources

With the worldwide structural transition of energy, traditional economic dispatch (ED) models and methods cannot adapt to modern power systems (PSs) demands. To response the limitations of conventional approaches in handling high-dimensional and nonlinear dispatch scenarios, this study proposes a large model accelerated bidirectional encoder representation from Transformers (BERT) distributed multi-objective economic dispatch method (BERT-DMOED) and a distributed novel PSs ED model with large-scale renewable energy sources. The BERT-DMOED leverages distributed computing and the deep feature representation and generalization capabilities of pre-trained BERT models to learn complex nonlinear mappings between load data and optimal generator outputs obtained from conventional multi-objective optimization results. The training of BERT-DMOED is performed with off-line learning. Once the training is complete, BERT-DMOED can quickly obtain dispatch schemes and assign to all generator sets, resulting in a significant reduction in computation time. Simulation results indicate that BERT-DMOED is compared with 7 comparison algorithms: (1) in the 118-bus PS, average generator cost and average carbon dioxide emission are reduced by 16.94 %–20.33 % and 73.06 %–75.49 %; (2) in the 1018-bus PS, average generator cost and average carbon dioxide emission are reduced by 21.28 %–23.47 % and 85.72 %–87.75 %; (3) the BERT-DMOED achieves optimality in Euclidean distance, spacing, and dispatching time.