| 引用本文: | 余建明,刘 赫,单连飞,等.基于ALBERT和RE2融合模型的电网调度意图识别方法[J].电力系统保护与控制,2022,50(12):144-151.[点击复制] |
| YU Jianming,LIU He,SHAN Lianfei,et al.Method of power grid dispatch intention recognition based on ALBERT and RE2 fusion model[J].Power System Protection and Control,2022,50(12):144-151[点击复制] |
|
| 摘要: |
| 针对电网调度业务意图缺乏有效识别方法的问题，提出一种基于ALBERT(A Lite BERT)和残差向量-字词嵌入向量-编码向量(RE2)融合模型的电网调度意图识别方法。首先，基于ALBERT预训练的动态词向量计算调度专业语言文本特征，建立调度意图分类模型，通过训练调度专业语言构建基于RE2的文本相似度计算模型。然后，采用RE2相似度模型计算召回文本与分类文本的匹配结果对ALBERT意图分类权重进行计算重组，建立融合ALBERT和RE2的意图识别模型。最后，通过某调控中心调度专业语言验证，并与其他方法对比，所提电网调度意图识别方法具有更强的分类能力和泛化能力，对于20种调度意图识别的平均精准率、召回率和F1值分别达到了98.11%、97.96%、98.03%。 |
| 关键词: 电网调度 ALBERT RE2 意图识别 融合模型 |
| DOI：DOI: 10.19783/j.cnki.pspc.211099 |
| 投稿时间：2021-08-15修订日期：2021-12-23 |
| 基金项目:国家电网公司总部科技项目“智能调控机器人助手关键技术研究与示范应用”(5700-202028362A-0-0-00) |
|
| Method of power grid dispatch intention recognition based on ALBERT and RE2 fusion model |
| YU Jianming,LIU He,SHAN Lianfei,ZHANG Yue,QIAO Yongtian,JIANG Tao |
| (1. NARI Group Corporation Co., Ltd., (State Grid Electric Power Research Institute Co., Ltd.,), Nanjing 211106, China;
2. Beijing KeDong Electric Power Control System Co., Ltd., Beijing 100192, China; 3. National Electric Power
Dispatching and Control Center of State Grid Corporation of China, Beijing 100031, China) |
| Abstract: |
| There is a problem of the lack of effective identification methods for power grid dispatching business intentions. Thus a method for power grid dispatching intention recognition based on ALBERT (A Lite BERT) and a fully integrated residual vector-embedding vector-encoded vector (RE2) fusion model is proposed. First, the dispatching professional language text features are calculated based on the ALBERT pre-trained dynamic word vector, a dispatching intention classification model is built. A text similarity calculation model is built through training dispatching professional languages based on RE2. Then the similarity value is calculated using the dispatching language similarity model between the recalled and the classified text to reorganize the ALBERT intention classification weight. Finally, through the professional language verification of a dispatching and control center, the proposed power grid dispatching intention recognition method has stronger classification ability and generalization ability than other methods. The average accuracy rate, recall rate and F1 value of 20 kinds of dispatching intent recognition reached 98.11%, 97.96%, and 98.03% respectively.
This work is supported by the Science and Technology Project of the Headquarters of State Grid Corporation of China (No. 5700-202028362A-0-0-00). |
| Key words: power grid dispatch ALBERT residual vector-embedding vector-encoded vector intention recognition fusion model |
