摘要
针对分布式光伏点多面广、集中式光伏的异常检测技术难以在分布式光伏系统中应用等问题,该文提出一种采用模糊C均值聚类算法(FCM)和贝叶斯优化算法(BOA)优化时间卷积网络-门控循环单元神经网络(TCN-GRU)的分布式光伏发电异常检测方法。首先,对原始数据进行异常值处理和相关性分析筛选出最佳特征;其次,为了降低天气波动性对光伏出力预测结果的影响,提出一种基于权重的FCM-Frechet算法对数据进行两阶段相似日集群划分,将其划分为晴朗、多云和阴雨相似日;然后,为了提高不同相似日下光伏常态出力预测的准确性,提出一种基于BOA优化的TCN-GRU网络模型;最后,采用实际光伏电站进行案例分析,利用设定的规则进行异常判断。结果表明,该文所提方法相较于CNN-LSTM和Transformer-BiLSTM模型的准确率分别提高了11.24和3.92个百分点,验证了所提方法在分布式光伏异常检测上的有效性。
To address the challenges that the anomaly detection technology for centralized photovoltaic systems is difficult to apply in distributed photovoltaic systems due to the widespread and diverse nature of distributed photovoltaic points,this paper proposes a distributed photovoltaic power generation anomaly detection method based on the fuzzy C-means(FCM)clustering algorithm and Bayesian optimization algorithm(BOA)optimized TCN-GRU network.The method involves feature selection,the FCM-Frechet model for clustering similar days,and the BOA-TCN-GRU model for predicting normal photovoltaic power output,combined with a dual dynamic threshold method to determine anomalies in distributed photovoltaic power generation systems.This approach reduces the need for annotated datasets and improves the accuracy of photovoltaic system anomaly detection.Firstly,the original data is treated for outliers and Pearson correlation analysis is conducted to select current,horizontal total radiation,temperature,humidity,horizontal scattered radiation,and wind speed as inputs to construct a similar day clustering model.Secondly,to mitigate the impact of weather volatility on photovoltaic power prediction results,a weighted FCM-Frechet algorithm is proposed for two-stage similar day clustering,categorizing weather into clear,cloudy,and rainy similar days.Then,using mutual information to screen redundant features,a BOA-optimized TCN-GRU network model is proposed.TCN processes local and global spatial features of time series,while GRU handles temporal features,achieving spatiotemporal joint modeling of features and improving the prediction accuracy of normal photovoltaic power output.Finally,the dual dynamic threshold method with an adaptive factor is used to determine photovoltaic system anomalies.The effectiveness of the proposed method is verified using the Alice Springs dataset from the Australian desert area and a building photovoltaic power generation dataset from Changsha City.The results show that the proposed method achieves higher detection accuracy.This paper implements two simulations using Pytorch.The first simulation aims to predict normal photovoltaic power output and detect anomalies under different similar days using the proposed BOA-TCN-GRU model.The results indicate that under cloudy conditions,the proposed method reduces RMSE and MAE by 0.2239 kW and 0.1706 kW,respectively,compared to FCM-CNN-LSTM-Attention;under rainy conditions,RMSE and MAE are reduced by 0.3857 kW and 0.3989 kW,respectively.The proposed method demonstrates better weather generalizability in normal photovoltaic power prediction tasks.On clear days,the FPR is reduced by 40.38%and 33.71%compared to CNN-LSTM and Transformer-BiLSTM,respectively,with an accuracy increase of 11.24 and 3.92 percentage points.The second simulation aims to verify the generalization ability of the proposed anomaly detection model.The results show that the accuracy rates of CNN-LSTM and Transformer-BiLSTM are 81.32%and 85.43%,respectively,and the proposed method can still detect photovoltaic anomalies on other datasets,with an accuracy rate reaching 90%.From the simulation analysis,the following conclusions can be drawn:(1)The weighted FCM-Frechet clustering model effectively reduces the impact of weather volatility on photovoltaic prediction results by clustering similar days.(2)Considering environmental conditions,the BOA-optimized TCN-GRU network model,combined with the dual dynamic threshold method with an adaptive factor,improves the accuracy and versatility of the model in photovoltaic anomaly detection.(3)Experimental results show that the proposed model improves accuracy by 11.24 and 3.92 percentage points compared to CNN-LSTM and Transformer-BiLSTM.This paper identifies an issue with the inaccuracy of the model between moving clouds and photovoltaic output under cloudy weather conditions.The next step will involve researching high-precision spatiotemporal modeling methods for the relationship between cloud movement and photovoltaic output in cloudy weather and proposing detection methods for identifying photovoltaic anomalies under this model.
作者
彭昱
符琛
郭昕
黄守道
苏盛
Peng Yu;Fu Chen;Guo Xin;Huang Shoudao;Su Sheng(School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha,410114,China;School of Intelligent Manufacturing,Hunan First Normal University,Changsha,410205,China;College of Electrical and Information Engineering,Hunan University,Changsha,410082,China)
出处
《电工技术学报》
北大核心
2025年第17期5389-5401,共13页
Transactions of China Electrotechnical Society
基金
国家自然科学基金(62473065)
湖南省自然科学基金(2023JJ40053)
湖南省教育厅优秀青年基金(23B0321)
湖南省教育厅资助科研项目(23C0424)资助。
关键词
分布式光伏异常检测
时间卷积网络
门控循环单元
相似日聚类
贝叶斯优化
Distributed photovoltaics anomaly detection
temporal convolutional network
gate recurrent unit
similar day clustering
Bayesian optimization
