智能电网窃电行为在线高效检测关键技术研究

Electricity theft not only brings about annual economic losses of up to 20 billion yuan for China, but also causes large-scale blackouts, fires and even casualties. However, existing electricity theft detection methods cannot well satisfy utility companies’ goal of locating electricity thieves quickly and accurately at low costs. This project focuses on developing key technologies of efficiently detecting electricity theft online in smart grid. Specifically, this project first studies how to build the interaction model and select strategies for electricity theft stakeholders. A multi-agent game whose players are electricity theft stakeholders is established to accurately and vividly model the interaction process of stakeholders, based upon which a reinforcement learning based strategy selection approach is further proposed to provide theoretical guidance for stakeholders selecting strategies. Then, this project studies how to deploy electricity theft monitors under limited monitoring resources. A particle swarm optimization-based monitoring point deployment approach is proposed to achieve the goal of complete observability of electricity theft behavior across the whole network, which provides technical support for detecting electricity theft efficiently. Afterwards, this project studies how to locate electricity thieves fast and accurately. A suspicion assessment-based electricity thieves inspection approach is proposed, which shortens the inspection time to a large extent. Finally, this project develops both simulation and physical experiment platforms to verify the effectiveness, practicality and efficiency of the proposed approaches. This project provides basic theories and advanced technologies for building a highly reliable and secure smart grid.

窃电行为不仅给我国带来每年高达200亿元的经济损失，也会引发大面积停电和火灾，甚至造成人员伤亡。然而，现有窃电行为检测方法尚不能很好地满足电网公司以较低成本快速准确定位窃电用户的要求。本课题针对智能电网窃电行为在线高效检测展开研究，具体为：研究窃电行为利益相关者互动建模及策略选择问题，提出基于多主体博弈与强化学习的窃电行为利益相关者最优策略选择方法，对利益相关者的策略选择提供理论指导；研究有限监测资源下的窃电行为监测点合理优化部署问题，提出一种基于粒子群优化算法的窃电行为监测点优化部署方法，为快速准确定位窃电用户提供技术支撑；研究最优监测器配置方案下的窃电用户快速准确定位问题，提出一种基于窃电嫌疑评估的窃电用户快速定位方法，极大地缩短窃电用户查找时间；最后，设计和开发仿真及物理实验平台，验证所提方法的有效性、实用性及高效率，为构建高可靠高安全智能电网提供基础理论和技术手段。

窃电检测是智能电网安全领域的重要问题之一。作为智能电网关键基础设施，智能电表计算、存储资源有限，且存在一些安全漏洞。这使得恶意用户能够利用物理攻击或者网络攻击来篡改智能电表读数从而实现其窃电目的。(1) 项目组全面综述和分析了窃电原因、窃电危害、窃电方法，并重点阐述了窃电检测技术最新研究进展、后续研究方向及相关技术挑战。(2) 针对现有基于机器学习的窃电检测方法不能很好地捕捉用户用电模式的周期性特征而导致检测准确率相对较低、误报率相对较高等问题，提出一种基于卷积长短期记忆网络（ConvLSTM）的窃电行为检测方法ETD-ConvLSTM。通过将用户用电量时间序列转换为具有时空特性的二维数据矩阵，该方法能够兼顾全局信息和局部信息来捕捉用户用电模式的周期性特征，从而提高检测精度。(3) 针对现有窃电检测方法一般都假设用户窃电量较大，而在少量窃电（Small-amount Electricity Theft, SET）攻击攻击模型下检测准确率大大降低甚至完全失效等问题，提出了一种基于休哈特—累积和联合控制图的窃电用户检测算法。该算法由窃电检测阶段和恶意用户识别阶段组成；其中窃电检测阶段旨在检测邻域网中是否有窃电情形存在，恶意用户识别阶段旨在准确识别出所有的恶意用户。在两个阶段，通过分析中央观测电表的供电总量测量值和由所有用户家中的智能电表上传的用电量值，估计休哈特控制图及累积和控制图的相关参数，从而实现快速检测/准确定位窃电用户。