基于彩色V-I轨迹特征和边缘机器学习非侵入式负荷识别方法
陆玲霞 , 孟繁举 , 于淼 , 任沁源 , 包哲静
工程科学与技术 ›› 2025, Vol. 57 ›› Issue (05) : 134 -141.
基于彩色V-I轨迹特征和边缘机器学习非侵入式负荷识别方法
Non-intrusive Load Monitoring Based on Colored V-I Trajectory Features and Edge Machine Learning
非侵入式负荷识别方法作为分析用户用电行为的主要途径,对开展能耗监测、实现用电安全评估具有重要意义。针对传统基于V-I轨迹特征的非侵入式负荷识别方法存在特征重叠和无法识别未知负荷的问题,提出一种基于彩色V-I轨迹特征和轻量级孪生网络的非侵入式负荷识别方法。首先,通过负荷电压电流数据构建具有方向信息的彩色V-I轨迹图像。然后,利用孪生网络计算待识别负荷的V-I轨迹图像和负荷特征库中V-I轨迹图像之间的相似度,以完成初步识别。随后,计算电流谐波特征之间的余弦距离,与阈值对比完成最终负荷识别。在以STM32MP1微处理器为核心的嵌入式Linux系统上,使用实验室电器负荷进行了实物验证。结果表明:彩色V-I轨迹能更详细地反映负荷特征,提高负荷识别准确率,并且由于改进的人工智能模型比较轻量化,对计算量需求大大减小,可以在嵌入式设备端对负荷特征库进行动态实时在线更新,轻松完成模型再训练。与依赖服务器的传统算法相比,无需返回PC或服务器重新训练模型并重新移植模型到嵌入式设备端。该方法仅依赖嵌入式终端便可准确识别未知负荷,避免在出现较多未知负荷后识别准确率下降,保证了负荷识别效果。系统运算一次负荷识别时间为0.2 s左右,可以满足实时性要求,具有重要的研究价值和实用性。
Objective A large amount of data is generated during the process of power consumption, which serves as the basis for informed user decisions and helps realize home energy efficiency monitoring, safety protection, and demand-side management. Therefore, obtaining real-time electricity consumption information through load identification has significant research value. Non-intrusive load monitoring (NILM) refers to a method that does not require the installation of a monitoring device for each individual load. Instead, it analyzes only the voltage and current data on the bus to obtain information about various factors, such as load type, operating status, and power consumption. Studies have shown that non-intrusive load recognition helps users reduce their energy consumption by up to 15%. As a primary tool for analyzing users’consumption behavior, non-intrusive load identification is crucial for both energy-use monitoring and electrical safety assessment. Methods based on the V-I trajectory have indicated promise, but several limitations remain. First, V-I trajectories often overlap: many appliances produce very similar shapes, especially after normalization, so using the trajectory alone makes it challenging to separate look-alike loads. This calls for integrating auxiliary cues, power and power-factor statistics, harmonic features, or other spectral or temporal descriptors. Second, many approaches are closed-set. They frame recognition as multiclass classification and identify only classes seen during training, while unknown loads are mishandled. The associated neural classifiers are also relatively complex. Third, training pipelines typically depend on server-class compute. When ported to embedded edge devices, limited resources prevent timely model updating or retraining. As the number of unseen loads grows, accuracy degrades without on-device or real-time adaptation, making it challenging to guarantee effectiveness and latency in practice and revealing a gap between lightweight AI and deployable systems. Methods This study proposed a NILM method that combined colored voltage-current (V-I) trajectory features with a lightweight Siamese network. The aim is to address the overlapping features caused by different loads. The method has two main objectives: to enhance the information conveyed by V-I images and to maintain a compact inference pipeline suitable for embedded, real-time NILM applications. This study introduced a simple and effective method for constructing colored V-I trajectory images. The approach involved creating colored V-I trajectory images that incorporated directional information using load voltage and current data, while employing the RGB color channels of the images. Compared to traditional V-I trajectories, this new method captured more load characteristics and clearly reflected the distinct features of each load. The improved colored V-I trajectory enabled a more detailed depiction of load characteristics, enhancing the accuracy of load identification. Given that the load V-I trajectory images were not overly complex and that NILM must operate online on embedded devices, a highly complex neural network structure was unnecessary. This study employed a Siamese network to calculate the similarity between the V-I trajectory image of the load to be identified and the V-I trajectory images in the load feature database, which enabled preliminary identification. The Siamese network model consisted of two components: the convolutional neural network (CNN) model and the backpropagation (BP) model. The CNN model extracted feature vectors from the colored V-I trajectory images, while the BP model computed the similarity between these feature vectors. The CNN model referenced was the classical lightweight neural network structure known as LeNet-5, which produced a 32-dimensional feature vector. LeNet-5 has a simple architecture and low complexity, which makes it suitable for real-time operation. The BP model took as input a 64-dimensional vector derived from merging two 32-dimensional feature vectors obtained by the CNN model, and its output was a continuous value between 0 and 1. An evaluation was made on whether the input vectors belonged to the same class by comparing the output results of the Siamese network with a predetermined threshold. The second stage incorporated harmonic features for identification after the initial recognition using the Siamese network to prevent misidentification of different loads with similar V-I trajectories. Specifically, the cosine distance between the harmonic features of the current was calculated and compared against a threshold to complete the load identification process. Initially, loads in the harmonic feature database corresponding to the V-I trajectory image that has matching harmonic features (including the 1st, 3rd, 5th, and 7th harmonics) were identified, and the similarity between these harmonic features was assessed using cosine distance. A greater cosine distance indicated higher similarity between the two loads. If the similarity exceeded the threshold, the load to be identified was classified as a known load in the database. In contrast, if the similarity fell below the threshold, the load was considered a new load, and its feature vector was added to the feature database. Through this dynamic updating of the feature database, the load was recognized as known if it appeared again. The method effectively avoided misidentifications that could have arisen from the similarities in V-I trajectories of different loads by integrating the features of V-I trajectories with harmonic features. Results and Discussions The Siamese network model was trained using V-I trajectories from the WHITED dataset and was deployed on an embedded Linux system powered by an STM32MP1 microprocessor. It was then validated using laboratory electrical loads. The results indicated that the colored V-I trajectory provided a more detailed representation of load characteristics, which enhanced the accuracy of load identification. In addition, the improved artificial intelligence model was lightweight, which significantly reduced computational requirements. The on-device feature database was updated online in real time, and local retraining or incremental updates were supported. Unlike server-dependent pipelines, this deployment eliminated the requirement for a round-trip to a PC or server for retraining and subsequent redeployment to the edge device. Conclusions The method employs the embedded terminal to accurately identify unknown loads, preventing a reduction in recognition accuracy as the number of unknown loads increases and ensuring effective load recognition. The system's runtime for identifying a single load is approximately 0.2 seconds, which satisfies real-time requirements and demonstrates substantial research value and practicality. The colored V-I trajectory combined with a lightweight Siamese network model provides a robust and deployable approach for real-time NILM on resource-constrained embedded hardware.
| [1] |
|
| [2] |
|
| [3] |
董坤,赵剑锋,孙睿晨, |
| [4] |
|
| [5] |
张丽,刘青雷,张宏伟.基于改进二进制粒子群算法的家庭负荷优化调度策略[J].中国电力,2023,56(5):118‒128. |
| [6] |
|
| [7] |
|
| [8] |
|
| [9] |
汪繁荣,向堃,刘辉.基于改进AP聚类与优化GRNN的非侵入式负荷分解研究[J].工程科学与技术,2020,52(4):56‒65. |
| [10] |
|
| [11] |
|
| [12] |
郭红霞,陆进威,杨苹, |
| [13] |
|
| [14] |
李如意,张鹏,刘永光, |
| [15] |
|
| [16] |
|
| [17] |
林平川,路磊,谷超, |
| [18] |
|
| [19] |
荆永震.非侵入式负荷识别系统研发[D].广州:广东工业大学,2019. |
| [20] |
|
| [21] |
王爱囡,袁麓,丁长波, |
| [22] |
|
| [23] |
王守相,郭陆阳,陈海文, |
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
刘宇,刘丛笑,赵欣, |
| [28] |
|
| [29] |
周任飞,汤鹏飞,刘三丰, |
| [30] |
|
| [31] |
|
| [32] |
宰州鹏,赵升,朱翔鸥, |
| [33] |
|
| [34] |
汪颖,杨维,肖先勇, |
| [35] |
|
| [36] |
|
| [37] |
裘星,尹仕红,张之涵, |
| [38] |
|
| [39] |
崔昊杨,蔡杰,陈磊, |
| [40] |
|
| [41] |
|
| [42] |
李双峰.TensorFlow Lite:端侧机器学习框架[J].计算机研究与发展,2020,57(9):1839‒1853. |
| [43] |
|
| [44] |
|
| [45] |
|
| [46] |
陆玲霞,强柱成,于淼, |
| [47] |
|
国家自然科学基金面上项目(52077194)
浙江省“十四五”第二批本科省级教学改革备案项目(JGBA2024014)
教育部产学合作协同育人项目(2501270945)
浙江大学本科“AI赋能”示范课程建设项目(2024-24)
浙江大学AI For Education系列实证教学研究项目(202402)
/
| 〈 |
|
〉 |