Multi-scale spatio-temporal transformer: A novel model reduction approach for day-ahead security-constrained unit commitment

Security-constrained unit commitment (SCUC) in large-scale power systems faces significant computational challenges, particularly with increasing renewable energy integration. This paper introduces a multi-scale spatio-temporal transformer (MSTT) model for efficient SCUC problem reduction through three key innovations: a multi-scale ST attention mechanism integrating both hierarchical temporal attention and electrical distance-based spatial attention to capture complex system dependencies, a physics-informed position encoding method incorporating power system domain knowledge including electrical distance, power flow sensitivity, and generator stability characteristics, and an adaptive reduction strategy with dynamic threshold adjustment mechanism that automatically balances computational efficiency and solution reliability based on system states and prediction confidence. Experimental results on IEEE test systems demonstrate that the MSTT model achieves up to 69.5 % computational time reduction while maintaining solution optimality (base-normalized cost (BNC) ≤ 0.05 %), significantly outperforming existing approaches. •A novel MSTT model captures complex spatio-temporal dependencies in SCUC through multi-scale attention.•Physics-informed position encoding integrates power system domain knowledge with deep learning.•Dynamic threshold adjustment mechanism enables adaptive model reduction for efficient SCUC solutions.•Comprehensive validation shows superior accuracy and 69.5 % time reduction on IEEE test systems.•MSTT bridges deep learning and power optimization for intelligent large-scale dispatch operations.