| 引用本文: | 赵健,曹子恺,高源,李梁.基于不变风险最小化的高比例分布式光伏配电网拓扑异常检测[J].电力自动化设备,2025,45(1):84-91 |
| ZHAO Jian,CAO Zikai,GAO Yuan,LI Liang.Invariant risk minimization based topological anomaly detection of distribution network with high proportion distributed photovoltaic[J].Electric Power Automation Equipment,2025,45(1):84-91 |
|
| 摘要: |
| 高比例分布式光伏接入配电网后，其出力波动导致终端用户功率数据偏移，不满足机器学习中数据独立同分布的要求，导致基于学习的拓扑异常检测方法收敛慢、精度低。为此，提出一种基于不变风险最小化(IRM)方法的高比例分布式光伏配电网拓扑异常检测方法。基于自适应在线深度学习构建配电网数据潮流(DF)模型，利用用户和配电变压器侧的历史量测数据反映量测点有限的配电网中各电气量间的潮流映射关系。基于IRM方法构建配电网DF模型，形成改进的IRM-DF模型，降低高比例分布式光伏接入产生的数据分布偏移对原有DF模型的影响，提升模型的准确性。利用IRM-DF模型代替配电网的节点网络方程，输出各节点关联性矩阵，并利用孤立森林算法筛选出矩阵中的异常点，确定异常拓扑位置。以仿真的低压配电网为例验证了所提方法的准确性和有效性。 |
| 关键词: 分布式光伏 配电网 不变风险最小化 自适应深度学习 数据潮流模型 |
| DOI：10.16081/j.epae.202410016 |
| 分类号:TM615 |
| 基金项目:国家自然科学基金资助项目(51907114) |
|
| Invariant risk minimization based topological anomaly detection of distribution network with high proportion distributed photovoltaic |
|
ZHAO Jian, CAO Zikai, GAO Yuan, LI Liang
|
|
College of Electrical Engineering, Shanghai University of Electric Power, Shanghai 200090, China
|
| Abstract: |
| After high proportion distributed photovoltaic(PV) integrated into the distribution network, the fluctuation of its output leads to the deviation of end-user power data, failing to meet the requirements of independent and identically distributed data in machine learning, which results in slow convergence and low accuracy in topology anomaly detection methods based on learning. To address this, a topology anomaly detection method of distribution network with high-proportion distributed PV is proposed based on IRM(invariant risk minimization) method. A distribution network data flow(DF) model based on adaptive online deep learning is constructed, utilizing historical measurement data from both users and distribution transformers to reflect the power flow mapping relationships between electrical quantities in the limited measurement points of the distribution network. Based on the IRM method, a DF model of distribution network is constructed, and an improved IRM-DF model is formed to reduce the impact of data distribution shifts caused by the integration of high proportion distributed PV on the original DF model, thereby improving the accuracy of the model. The IRM-DF model is used to replace the node network equation of distribution network, and the correlation matrix of each node is output. The isolation forest algorithm is used to filter out the anomalies in the matrix and determine the abnormal topology location. Taking the simulated low-voltage distribution network as an example, the accuracy and the effectiveness of the proposed method are validated. |
| Key words: distributed photovoltaic distribution network invariant risk minimization adaptive deep learning data flow model |
