融合多尺度卷积的xLSTM方法在非侵入式负荷识别中的应用

doi:10.19595/j.cnki.1000-6753.tces.251692

Abstract
Figure/Table
References (0)
Related Citation (4)

Download: PDF (1351 KB) HTML (1 KB)
Export: BibTeX | EndNote (RIS)

Abstract

Non-intrusive load identification (NILI) is a core function of non-intrusive load monitoring (NILM) systems, aiming to recognize individual appliance operating states from aggregated electrical measurements collected at a single monitoring point. In practical residential and industrial environments, NILI performance is often constrained by complex operating conditions, including simultaneous appliance operation, similarity in steady-state power levels, strong temporal variability, measurement noise, and limited labeled training data. Conventional time-series models, such as long short-term memory (LSTM) and gated recurrent unit (GRU) networks, rely on single-form memory update mechanisms. Consequently, their ability to jointly represent short-duration transient behaviors and long-term operational patterns is limited, particularly under non-stationary and noisy load conditions. These challenges motivate the development of more expressive and flexible sequence modeling frameworks for robust load identification.
This paper presents an end-to-end load identification framework integrating multi-scale convolutional feature extraction, attention-based feature modulation, and extended long short-term memory (xLSTM) temporal modeling. The framework is designed to enhance feature discriminability and temporal dependency modeling while maintaining moderate computational complexity suitable for practical NILM applications. By jointly modeling local transient responses, global contextual information, and long-range temporal dependencies within a unified architecture, the proposed method improves robustness and generalization in diverse load scenarios.
In the feature extraction stage, a dual-branch multi-scale convolutional structure analyzes power signals at different temporal resolutions. One branch employs convolution kernels with larger receptive fields combined with down-sampling operations to capture coarse-grained global trends and long-term power variations. The other branch stacks multiple small-kernel convolutional layers to extract fine-grained local details associated with transient events and short-duration fluctuations. To enhance nonlinear representation capability, a self-gated activation function with learnable parameters adaptively modulates channel-wise feature responses. In addition, a coordinate attention mechanism encodes channel dependencies together with temporal positional information, enabling selective emphasis on discriminative load segments while suppressing irrelevant fluctuations. The outputs of the two branches are spatially aligned and fused through element-wise operations, forming a compact representation that preserves complementary multi-scale characteristics.
For temporal modeling, xLSTM units composed of scalar long short-term memory (sLSTM) and matrix long short-term memory (mLSTM) modules are employed. Unlike conventional recurrent networks that update memory states using a single scalar representation, xLSTM provides more flexible and structured state representations through coordinated scalar and matrix memory updates. This design enables simultaneous modeling of short-term transient patterns and long-term operational cycles, improving robustness under noisy, overlapping, and non-stationary load conditions. By stacking xLSTM layers after the convolution-attention encoder, hierarchical spatio-temporal feature representations and long-range dependencies are effectively integrated, facilitating accurate load identification across varying operating durations.
The proposed method is evaluated on four representative public datasets: UK-DALE (UK Domestic Appliance-Level Electricity), LIFTED (Labelled High Frequency Dataset for Electricity Disaggregation), PLAID (Plug-Load Appliance Identification Dataset), and WHITED (Worldwide Household and Industry Transient Energy Data Set). These datasets cover diverse acquisition conditions, including long-term residential monitoring, controlled laboratory environments, transient-rich signals with strong noise, and high-fidelity waveform measurements. A unified preprocessing, training, and evaluation strategy is applied across all datasets to ensure fair comparison. Performance is evaluated based on five metrics: weighted precision, weighted recall, weighted F1-score, overall accuracy, and Matthews Correlation Coefficient (MCC). This comprehensive set of metrics is chosen to holistically reflect classification effectiveness under challenging conditions, including class imbalance and diverse misclassification patterns.
Experimental results demonstrate that the proposed framework achieves competitive and stable performance across different datasets and evaluation metrics. Compared with conventional recurrent models, convolutional neural networks, and representative hybrid deep learning approaches, the proposed method exhibits improved robustness and more balanced metric performance, particularly in scenarios involving long-duration appliance operation, similar power levels, and strong noise interference. Ablation studies further confirm that joint modeling of transient and steady-state features, the selection of discriminative segments, and the representation of long-term dependencies are enhanced, improved, and strengthened by multi-scale convolution, attention mechanisms, and xLSTM, respectively. The synergistic integration of these components results in the observed improvements in generalization capability and performance stability.
In summary, the proposed multi-scale convolution-attention-xLSTM framework offers a robust and generalizable solution for non-intrusive load identification under complex operating conditions. By jointly addressing feature extraction, temporal modeling, and noise robustness within a unified architecture, the method significantly improves load identification accuracy across diverse datasets and application scenarios, thereby advancing the development of practical NILM systems and energy monitoring applications.

Key words： Non-invasive load identification multi-scale convolution xLSTM time series modeling sLSTM mLSTM

Received: 30 September 2025

PACS:

TM714

	Service

	E-mail this article
	Add to my bookshelf
	Add to citation manager
	E-mail Alert
	RSS
	Articles by authors
	Li Jun
	Dai Hui

Cite this article:

Li Jun,Dai Hui. Application of xLSTM Method with Multi-scale Convolution in Non-invasive Load Identification[J]. Transactions of China Electrotechnical Society, 0, (): 692-.

URL:

https://dgjsxb.ces-transaction.com/EN/10.19595/j.cnki.1000-6753.tces.251692 OR https://dgjsxb.ces-transaction.com/EN/Y0/V/I/692

[1] 孔宇,张恒旭,施啸寒,等.基于多时间尺度分层协同的电力系统开放式推演框架[J].电工技术学报,2025,40(07):2063-2077.
Kong Yu, Zhang Hengxu, Shi Xiaohan, et al.Open Evolution Simulation Framework for Power System Based on Multi-Time Scale Hierarchical Coordination[J].Transactions of China Electrotechnical Society, 2025, 40(07): 2063-2077.
[2] 申刘飞,翟雨佳,吴星徵,等.海上超导风电制氢一体化研究进展与发展趋势[J].电工技术学报,2025,40(11):3362-3380.
Shen Liufei, Zhai Yujia, Wu Xingzheng, et al.Progress and Development Trend of Integrated Research on Hydrogen Production from Offshore Superconducting Wind Power[J].Transactions of China Electrotechnical Society, 2025, 40(11): 3362-3380.
[3] 李延珍,王海鑫,杨子豪,等.基于非侵入式负荷分解的家庭负荷两阶段超短期负荷预测模型[J].电工技术学报,2024,39(11):3379-3391.
Li Yanzhen, Wang Haixin, Yang Zihao, et al.Two-Stage Ultra-Short-Term Load Forecasting Model of Household Appliances Based on Non-Intrusive Load Disaggregation[J].Transactions of China Electrotechnical Society, 2024, 39(11): 3379-3391.
[4] 高洪超,康重庆,李嘉宇,等.新型电力系统下虚拟电厂的技术演进研判与运营挑战分析[J].电工技术学报,2025,40(19):6057-6071.
Gao Hongchao, Kang Chongqing, Li Jiayu, et al.Analysis of Technological Evolution and Operational Challenges for Virtual Power Plants in New Power Systems[J].Transactions of China Electrotechnical Society, 2025, 40(19): 6057-6071.
[5] 赵旭东,王艺博,王博闻,等.考虑差异化工业高载能负荷灵活性挖掘的市场实施及调度模型研究综述[J].电工技术学报,2025,40(07):2043-2062+2161.
Zhao Xudong, Wang Yibo, Wang Bowen, et al.Review of Market Implementation and Scheduling Models Considering the Flexibility Extraction of Differentiated Industrial Energy-Intensive Loads[J].Transactions of China Electrotechnical Society, 2025, 40(07): 2043-2062+2161.
[6] 朱仁庆,沈小军,董子航.虚拟电厂的隐私保护技术研究现状与展望[J].电力系统自动化,2025,49(20):16-33.
Zhu Renqing, Shen Xiaojun, Dong Zihang.Research Status and Prospects of Privacy Preservation Technologies for Virtual Power Plants[J].Automation of Electric Power Systems, 2025, 49(20): 16-33.
[7] 鲍海波,杨舒惠,陈子民,等.事件检测类非侵入式负荷监测算法综述[J].电力系统自动化,2023,47(13):94-109.
Bao Haibo, Yang Shuhui, Chen Zimin, et al.Review on Event-inspection Based Non-intrusive Load Monitoring Algorithms[J].Automation of Electric Power Systems, 2023, 47(13):94-109.
[8] Kotsilitis S, Kalligeros E, Marcoulaki E C, et al.An efficient light-weight event detection algorithm for on-site non-intrusive load monitoring[J].IEEE Transactions on Instrumentation and Measurement, 2022, 72: 1-13.
[9] 李延珍,王海鑫,杨子豪,等.基于多阶段数据递推分析的用户用电行为特性挖掘方法[J].电机与控制学报,2025,29(02):35-46.
Li Yanzhen, Wang Haixin, Yang Zihao, et al.User behavior characteristic mining method based on multi-stage data recurrence analysis[J].Electric Machines and Control, 2025, 29(02): 35-46.
[10] Etezadifar M, Karimi H, Aghdam A G, et al.Resilient event detection algorithm for non-intrusive load monitoring under non-ideal conditions using reinforcement learning[J].IEEE Transactions on Industry Applications, 2023, 60(2): 2085-2094.
[11] Sadeghianpourhamami N, Ruyssinck J, Deschrijver D, et al.Com-prehensive feature selection for appliance classification in NILM[J].Energy and Buildings, 2017, 151: 98-106.
[12] 解洋,梅飞,郑建勇,等.基于V-I轨迹颜色编码的非侵入式负荷识别方法[J].电力系统自动化,2022,46(04):93-102.
Xie Yang, Mei Fei, Zheng Jianyong, et al.Non-intrusive Load Monitoring Method Based on V-I Trajectory Color Coding[J].Automation of Electric Power Systems, 2022, 46(04):93-102.
[13] 雷怡琴,孙兆龙,叶志浩,等.电力系统负荷非侵入式监测方法研究[J].电工技术学报,2021,36(11):2288-2297.
Lei Yiqin, Sun Zhaolong, Ye Zhihao, et al.Research on Non-Invasive Load Monitoring Method in Power System[J].Transactions of China Electrotechnical Society, 2021, 36(11): 2288-2297.
[14] 黄莉,刘昱源,周赣,等.基于负荷暂态事件的低压台区户相变识别技术[J].电力系统自动化,2025,49(14):152-162.
Huang Li, Liu Yuyuan, Zhou Gan, et al.Identification Technology for Consumer-transformer Relationship and Phase of Low-voltage Distribution Station Area Based on Load Transient Events[J].Automation of Electric Power Systems, 2025, 49(14): 152-162.
[15] 胡正伟,王志红,畅瑞鑫,等.基于在线自组织增量学习的非侵入式负荷识别方法[J].工程科学与技术,2024,56(04):316-324.
Hu Zhengwei, Wang Zhihong, Chang Ruixin, et al.Non-intrusive Load Identification Method Based on the Online Self-organizing Incremental Neural Network[J].Advanced Engineering Sciences, 2024, 56(04): 316-324.
[16] 刘宇,刘丛笑,赵欣,等.基于维特比算法改进的稳暂态混合非侵入式负荷识别方法[J].电工技术学报,2023,38(19):5241-5255.
Liu Yu, Liu Congxiao, Zhao Xin, et al.An Improved Steady-and Transient-State Mixed Non-Intrusive Load Monitoring Using Viterbi Algorithm[J].Transactions of China Electrotechnical Society, 2023, 38(19):5241-5255.
[17] 王磊,马佳琪,韩肖清,等.考虑多状态特征的非侵入式负荷识别方法[J].电网技术,2024,48(11):4720-4728.
Wang Lei, Ma Jiaqi, Han Xiaoqing, et al.Non-intrusive Load Monitoring Method With Multi-state Characterization of Loads[J].Power System Technology, 2024, 48(11): 4720-4728.
[18] Yan Z, Hao P, Nardello M, et al.A Generalizable Load Recognition Method in NILM Based on Transferable Random Forest[J].IEEE Transactions on Instrumentation and Measurement, 2025.
[19] 宰州鹏,赵升,朱翔鸥,等.基于颜色编码与谐波特征融合的非侵入式负荷识别方法[J].电气技术,2022,23(12):9-16.
Zai Zhoupeng, Zhao Sheng, Zhu Xiang’ou, et al.Non-intrusive load monitoring based on color coding and harmonic feature fusion[J].Electrical Engineering, 2022, 23(12): 9-16.
[20] 周步祥,赵雯雯,臧天磊,等.基于低频功率差量特征与双长短期记忆网络的非侵入式负荷监测方法[J].电力自动化设备,2023,43(08):167-173+209.
Zhou Buxiang, Zhao Wenwen, Zang Tianlei, et al. Non-intrusive load monitoring method based on low-frequency power difference characteristic and dual long short-term memory network[J].Electric Power Automation Equipment, 2023, 43(08):167-173+209.
[21] 林顺富,詹银枫,李毅,等.基于CNN-BiLSTM与DTW的非侵入式住宅负荷监测方法[J].电网技术,2022,46(05):1973-1981.
Lin Shunfu, Zhan Yinfeng, Li Yi, et al.Non-intrusive Residential Load Monitoring Method Based on CNN-BiLSTM and DTW[J].Power System Technology, 2022, 46(05): 1973-1981.
[22] 杨桂兴,王维庆,姚红雨,等.基于1DCNN-BP的非侵入式负荷识别算法[J].高电压技术,2023,49(07):3031-3039.
Yang Guixing, Wang Weiqing, Yao Hongyu, et al.Research on Non-intrusive Load Identification Method Based on 1DCNN-BP[J].High Voltage Engineering, 2023, 49(07): 3031-3039.
[23] 范睿,孙润稼,刘玉田.考虑空调负荷需求响应的负荷恢复量削减方法[J].电工技术学报,2022,37(11):2869-2877.
Fan Rui, Sun Runjia, Liu Yutian.A Load Restoration Amount Reduction Method Considering Demand Response of Air Conditioning Loads[J].Transactions of China Electrotechnical Society, 2022, 37(11): 2869-2877.
[24] Yaniv A, Beck Y.Advances in non-intrusive load monitoring for the industrial domain: Challenges, insights, and path forward[J].Renewable and Sustainable Energy Reviews, 2025, 210: 115136.
[25] Beck M, Pöppel K, Spanring M, et al.xlstm: Extended long short-term memory[J].Advances in Neural Information Processing Systems, 2024, 37: 107547-107603.
[26] Ma N, Zhang X, Liu M, et al.Activate or not: Learning customized activation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021: 8032-8042.
[27] Nie Z, Yang Y, Xu Q.An ensemble-policy non-intrusive load monitoring technique based entirely on deep feature-guided attention mechanism[J].Energy and Buildings, 2022, 273: 112356.
[28] Wang Z, Zheng G.Residential appliances identification and monitoring by a nonintrusive method[J].IEEE transactions on Smart Grid, 2011, 3(1): 80-92.
[29] Chicco D, Jurman G.The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation[J].BMC genomics, 2020, 21(1): 6.
[30] Kelly J, Knottenbelt W.The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes[J].Scientific data, 2015, 2(1): 1-14.
[31] Yan L, Han J, Xu R, et al.LIFTED: Household appliance-level load dataset and data compression with lossless coding considering precision[C]//2020 IEEE Power & Energy Society General Meeting (PESGM). IEEE, 2020: 1-5.
[32] Medico R, De Baets L, Gao J, et al.A voltage and current measurement dataset for plug load appliance identification in house-holds[J].Scientific data, 2020, 7(1): 49.
[33] Kahl M, Haq A U, Kriechbaumer T, et al.Whited-a worldwide household and industry transient energy data set[C]//3rd international workshop on non-intrusive load monitoring. 2016: 1-4.
[34] 张帅,程志友,田甜,等.基于马尔可夫转移场和轻量级网络的非侵入式负荷识别[J].电力系统保护与控制,2024,52(17):51-61.
Zhang Shuai, Cheng Zhiyou, Tian Tian, et al.Non-intrusive load identification based on the Markov transition field and a lightweight network[J].Power System Protection and Control, 2024, 52(17): 51-61.
[35] 魏广芬,李谊林, KUZENGURIRA T.Tapiwa,等.基于彩色图像特征提取及融合的非侵入式负荷识别[J].电网技术,2025,49(11):4854-4864.
Wei Guangfen, Li Yilin, KUZENGURIRA T.Tapiwa, et al. Non-Intrusive Load Recognition Based on Color Image Feature Extraction and Fusion[J].Power System Technology, 2025, 49(11): 4854-4864.
[36] 沈鑫,王钢,赵毅涛,等.融合SENet注意力机制和GA-CNN的非侵入式负荷识别方法[J].中国电力,2025,58(05):33-42.
Shen Xin, Wang Gang, Zhao Yitao, et al.A Non-Invasive Load Recognition Approach Incorporating SENet Attention Mechanism and GA-CNN[J].Electric Power, 2025, 58(05): 33-42.