Integrating Graph Signal Processing and Multitask Temporal Convolutional Networks for Household Nonintrusive Load Monitoring

• Published in: IEEE transactions on instrumentation and measurement Vol. 73; pp. 1 - 12
• Main Authors: Su, Yongxin, Peng, Haotian, Tan, Mao, Chen, Jie
• Format: Journal Article
• Language: English
• Published: New York IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Computational modeling; Convolution networks; Data mining; Data models; Energy management; Feature extraction; Graph representations; graph signal processing (GSP); Graphical representations; Hidden Markov models; Home appliances; load disaggregation (LD); Load modeling; Model accuracy; Monitoring; multitask learning; nonintrusive load monitoring (NILM); Pattern matching; Signal processing; Sparsity
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3379372

Highly accurate nonintrusive load monitoring (NILM) models are essential for energy management, optimization decisions, and system monitoring. However, the sparsity of load features and spatio-temporal relationships hidden in loads have not been fully tackled, hindering the accuracy of NILM models. To solve the above problems, a load disaggregation (LD) framework that cascades graph signal processing (GSP) with multitask temporal convolution network (TCN) is proposed. In this framework, the GSP is designed as a direct load feature extractor to tackle the sparsity of load features. The multitask TCN can leverage GSP outputs and total power to extract spatio-temporal relationships among loads, and then generate precise LD results for each load simultaneously. Afterward, the implementation of the GSP-based direct load feature extractor is designed, including the construction of graph representation of load features, pattern matching, and direct load features correction. Then the implementation scheme of multitask TCN is proposed, consisting of load features fusion strategy, the spatio-temporal relationships extractor design, and the loss function setting and training strategy. The experiment shows that our model can concurrently output LD results for eight appliances. Meanwhile, compared to existing advanced methods, our model has an over 13% reduction on mean absolute error (MAE).