基于神经机器翻译的电网故障诊断

Intelligent power grid fault diagnosis has important significance for improving the automation level of power grid dispatching operation decision-making. How to extract and use the depth timing sequence knowledge of alarm information, improve the solving ability and enhance the level of intelligence is an urgent problem to be resolved in the current power grid fault diagnosis. Based on the idea of end-to-end neural machine translation, this project studies deep network models and diagnostic methods that can extract and represent deep timing sequence knowledge of power grid fault diagnosis. The character decomposition clustering is used to separate the objects and attributes in the text description of the alarm information and obtain the description participle of the grid fault event, which makes it possible to extract deep knowledge. Based on the simulation analysis of power system faults including the description of action process in electric secondary system, the characterization mechanism of fault event description participle to the fault evolution process is revealed. On the basis, the grid fault diagnosis sequence learning model that can extract and represent the depth timing sequence knowledge is studied, and a diagnosis method based on supervised sequence learning is proposed. The sequence learning model is improved to construct a reinforcement learning depth strategy value evaluation network of fault diagnosis. The reinforcement learning methods of power grid fault diagnosis is proposed for the target of self-learning to enhance solving ability under fewer sample conditions. The research results will improve the intelligent level of power grid fault diagnosis.

智能化电网故障诊断对提高电网调度运行决策自动化水平具有重要意义。如何提取使用告警信息中的深度时序知识，提高求解能力、提升智能化水平是当前电网故障诊断急需解决的问题。而目前的诊断模型和方法缺乏深度知识提取和表示能力。本项目基于端到端神经机器翻译的求解思路，研究可提取和表示电网故障诊断深度时序知识的深度网络模型及诊断方法。拟采用字符分解聚类将告警信息文本描述中的对象和属性相分离，获得电网故障事件描述分词，使深度知识提取成为可能。基于含二次系统动作过程描述的电网故障仿真分析，揭示故障事件描述分词对故障演化过程的表征机理。在此基础上，研究可提取和表示深度时序知识的电网故障诊断序列学习模型，提出基于有监督序列学习的诊断方法。改进序列学习模型，构建电网故障诊断强化学习深度策略价值评估网络，以自学习为目标提出电网故障诊断强化学习方法，增强较少样本条件下的求解能力。研究成果将提升电网故障诊断智能化水平。

智能化电网故障诊断对提高电网调度运行决策自动化水平具有重要意义。本项目采用电网调度实时告警信息为数据源，以端到端电网故障诊断为目的，采用序列学习和强化学习，研究了基于神经机器翻译的电网故障诊断。项目研究成果总结如下：（1）利用告警信息的时序分布特性可以有效提高电网故障诊断深度学习算法的训练效果和速度；（2）告警信息的量化表示是智能化的电网故障诊断的基础，不同的量化表示方法对于电网故障诊断的求解效果有一定的影响，如何表示告警信息的时序关系是数据驱动电网故障诊断的重点研究内容。（3）基于神经机器翻译模型可以实现端到端的电网故障诊断，能够自动提取告警信息文本中的电气设备名称，命名规则和关联关系，完成基于预设模板的结果输出。（4）采用强化学习可以实现复杂电网故障样本较少时断路器拒动故障的准确诊断。（5）不同模型和算法的验证表明，采用混合不同样本集数据，可以提高基于数据驱动电网故障诊断的模型迁移能力。