A Transfer Learning-Based Method for Cross-Scenario Operation Step Generation and Generalization Optimization in Distribution Networks

Abstract

In response to the issue of limited generalization in automatic generation of operation steps across different stations and operating conditions in distribution networks, this paper proposes a temporal-topological joint modeling method driven by transfer learning. The method takes device state sequences and network topology as inputs, utilizing a Transformer to extract temporal features and Graph Convolutional Networks (GCN) to model topological constraints. During the transfer phase, a “freeze the lower layers – progressively unfreeze” fine-tuning strategy is applied, combined with distribution alignment and domain adversarial learning to achieve cross-scenario adaptation. Evaluations on the Caltech 28-Bus Digital Twin dataset show that, compared to No Transfer and Fine-tuning, the proposed method reduces the Mean Squared Error (MSE) from 0.15 to 0.06, the Mean Absolute Error (MAE) from 0.39 to 0.245, and the Generalization Performance Index (GPI) to approximately 0.47 in the target domain. When only 10% of the target domain samples are used, the Relative Normalized Squared Error (RNSE) is approximately 0.55. For the sake of generation quality, BLEU is 0.83, and the Semantic Consistency Score (SCS) is 0.92. These findings show that the suggested framework can reliably produce operation steps that are semantically coherent and logically consistent across various topologies and operating conditions, offering a workable technological route for the automation of operation step generation in distribution network operation scheduling.