Powerformer: A Section-adaptive Transformer for Power Flow Adjustment

TL;DR

Powerformer targets robust state representations for sectional power flow adjustment by introducing a section-adaptive attention mechanism and a multi-factor attention framework that disentangles electrical factors and fuses graph topology via GNN propagation. The model integrates state features, section power flow, and transmission topology, enabling precise dispatch decisions with a Dueling DQN-based policy. Empirical results across IEEE 118-bus, 300-bus China, and 9241-bus European systems show high success rates (e.g., 98.19% on 118-bus) and substantially faster inference than traditional OPF, validating both effectiveness and practicality for real-time operation. The work also provides ablation analyses and visualization evidence, highlighting the benefits of state factorization and section-aware attention for scalable, topology-aware power-flow control with potential extensions to distribution networks and PV/storage scenarios.

Abstract

In this paper, we present a novel transformer architecture tailored for learning robust power system state representations, which strives to optimize power dispatch for the power flow adjustment across different transmission sections. Specifically, our proposed approach, named Powerformer, develops a dedicated section-adaptive attention mechanism, separating itself from the self-attention used in conventional transformers. This mechanism effectively integrates power system states with transmission section information, which facilitates the development of robust state representations. Furthermore, by considering the graph topology of power system and the electrical attributes of bus nodes, we introduce two customized strategies to further enhance the expressiveness: graph neural network propagation and multi-factor attention mechanism. Extensive evaluations are conducted on three power system scenarios, including the IEEE 118-bus system, a realistic 300-bus system in China, and a large-scale European system with 9241 buses, where Powerformer demonstrates its superior performance over several baseline methods.

Figures (10)

• Figure 1: Illustration of 10 transmission sections $\{\Phi_1,...,\Phi_{10}\}$ in the IEEE 118-bus system, where each section is represented by a set of transmission lines that share the same color and are closely located electrically. An illustrative section $\Phi_7$ is the set of purple lines, i.e., $\Phi_7 = \,${bus33-bus37, bus19-bus34, bus30-bus38, bus23-bus24}, that partitions the system into distinct left and right components.
• Figure 2: The illustration of the Transformer and Powerformer architectures. (a) Transformer architecture with the multi-head self-attention mechanism (MHA). (b) Powerformer architecture with the multi-factor section-adaptive attention (MFSA).
• Figure 3: The visualization highlights the distribution of Concat, Attention, and Powerformer operations for combining state and 10 types of transmission section representations on the IEEE 118-bus system.
• Figure 4: The learning curves of all methods for 10-section task on three power systems. The experimental results use the median performance and one standard deviation (shaded region) over 5 random seeds to ensure a fair comparison.
• Figure 5: The visualization effectively emphasizes the active power of all scenario samples, both before and after adjustment, across all ten transmission sections of the IEEE 118-bus system using our method. Among them, 170 scenarios have been successfully adjusted into the capacity of the transmission section, i.e., 98.19% test success rate reported in Table \ref{['tab:representation_comparison']} and \ref{['tab:total']}.