Text-To-Text Transfer Transformer Based Method for Generating Startup Scenarios for New Equipment in Power Grids

Abstract

Power grids often develop specialized startup plans for new equipment to standardize the commissioning process, which presents distinct challenges compared to other types of startup plans. Rule-based generation methods have limited transferability, making it difficult to adapt to the rapid evolution of power grid infrastructure. Current deep learning-based generation methods are primarily improvements on rule-based approaches, but they are still constrained by those rules. This paper proposes a startup plan generation method for new power grid equipment based on the Text-to-Text Transfer Transformer (T5). The method leverages a T5 model pretrained on Chinese text and fine-tunes it using historical startup plans for new power grid equipment to generate applicable startup texts. Additionally, a dynamic loss adjustment strategy, based on professional terminology judgment, is introduced to address inconsistencies in professional terminology during the generation process. The Adam optimizer is used for backpropagation and parameter updates. The trained model generates startup plan texts for new equipment using a beam search decoding strategy. Tests and validations using real-world data from four types of new equipment startup plans demonstrate that this method can effectively generate startup plans, meeting the requirements of practical applications.