Apply Graph Signal Processing on NILM: An Unsupervised Approach Featuring Power Sequences

Bochao Zhao, Xuhao Li, Wenpeng Luan, Bo Liu2023 · DOI

TL;DR

This paper introduces STS-UGSP, a novel unsupervised approach for Non-Intrusive Load Monitoring (NILM) that leverages graph signal processing (GSP) on state transition sequences (STSs) extracted from power data. By focusing on the dynamic transitions of appliance usage rather than just steady-state power, STS-UGSP achieves improved disaggregation accuracy, particularly for appliances with complex or transient power signatures.

What Problem Does This Paper Address?

Non-intrusive load monitoring (NILM) aims to disaggregate individual appliance energy consumption from a single aggregate power measurement, offering a cost-effective alternative to installing individual sensors. While existing NILM methods, especially those for low-rate power data (e.g., 1 Hz or less), have seen progress using techniques like Hidden Markov Models (HMMs), deep learning, and graph signal processing (GSP), there's still a need to enhance feature selection for better performance. Current GSP-based methods often rely on steady-state power sequences or simple power change events, which can struggle with appliances exhibiting complex start-up/shut-down transients or those with similar power profiles. This paper addresses the challenge of improving NILM accuracy by proposing a novel feature extraction method and an unsupervised GSP framework that better captures the dynamic behavior of appliances.

Methods

The proposed STS-UGSP approach consists of three main stages:

  1. Adaptive State Transition Sequence (STS) Extraction: This is a core innovation. Instead of focusing on steady-state power or just power changes, the method extracts "rising" and "falling" power sequences that capture the complete transient events of an appliance starting or stopping. This is achieved through an adaptive algorithm that considers power variations and neighboring samples, with parameters to constrain sequence length.
  2. GSP-Based STS Clustering: The extracted STSs are then mapped to nodes in a graph. The similarity between these STSs is quantified using Dynamic Time Warping (DTW) distances, which are then used to weight the edges of the graph. Graph signal processing techniques are applied to cluster these STSs, grouping similar "rising" and "falling" sequences independently.
  3. Forward-Backward STS Matching: After clustering, a novel forward-backward matching algorithm pairs the "rising" STSs with their corresponding "falling" STSs. This algorithm considers both the power magnitude and temporal information of the sequences, using trade-off factors to balance their influence. This matching process is crucial for accurately identifying complete appliance operational cycles.

Datasets

The proposed STS-UGSP method was validated on three publicly available datasets:

  • AMPds: Collected from Canada, sampled at 1-minute resolution.
  • REFIT: From the UK, sampled at 8-second resolution.
  • REDD: From US houses, sampled at 1Hz resolution.

These datasets represent varying levels of complexity, including different sampling rates, unknown appliances, and measurement noise, allowing for a comprehensive evaluation.

Key Insights and Ideas

  • Dynamic Transients Matter: The paper strongly suggests that focusing on the transition periods (start-up and shut-down) of appliance usage, rather than just steady-state power, provides richer information for disaggregation.
  • DTW for Sequence Similarity: Using Dynamic Time Warping (DTW) to measure the similarity between State Transition Sequences (STSs) is effective, especially when sequences have unequal lengths due to varying sampling rates or appliance usage patterns.
  • Graph Representation of Transitions: Mapping STSs to graph nodes and using DTW for edge weights allows GSP to effectively capture relationships between these dynamic events.
  • Forward-Backward Matching for Completeness: The proposed matching algorithm ensures that complete appliance cycles (from start to finish) are identified, improving accuracy by linking the initial "rising" transient to its corresponding "falling" transient.
  • Unsupervised Advantage: The unsupervised nature of the approach is key, as it doesn't require labeled data or sub-metering, making it more practical for real-world deployment in households.

Key Results

  • Superior Performance: STS-UGSP generally outperformed four benchmark methods (SS-UGSP, UGSP, SHMM, and CO) across the AMPds, REFIT, and REDD datasets in terms of F-measure (Fm) and disaggregation accuracy (Acc.).
  • Improved Accuracy: For instance, on the AMPds dataset, STS-UGSP achieved an overall Fm of 0.84, outperforming benchmarks. It showed particular strength in disaggregating appliances with distinct transient states, like HP.
  • Robustness to Noise and Unknown Loads: The GSP-based approach, especially STS-UGSP, demonstrated better robustness to unknown loads and measurement noise compared to SHMM and CO, which showed sharper drops in performance on challenging datasets like REFIT.
  • Efficiency Considerations: While GSP-based clustering can be computationally intensive with large graph sizes, the STS extraction is efficient, suggesting a hybrid approach where extraction is done on edge devices and clustering/matching on the cloud.

Limitations and Future Work

  • Granularity Dependence: The paper notes that for very low sampling rates (10 minutes or longer), state transition detection or STS extraction becomes unavailable, limiting the applicability of the method.
  • STS Matching Refinement: Future work could involve refining STS matching by further mining temporal connections among state transitions of multi-state loads.
  • Hybrid Features: Exploring the combination of STS features with steady-state power sequences could further improve power signal reconstruction.
  • Low-Frequency Measurements: Applying the proposed method to measurements with various low frequencies is suggested for future research.
  • Computational Cost: While STS extraction is efficient, GSP-based clustering's computational cost increases exponentially with graph size, suggesting optimization strategies for larger datasets.

Summary updated

BibTeX
@misc{zhao2023apply,
  author = {Bochao Zhao and Xuhao Li and Wenpeng Luan and Bo Liu},
  title = {{Apply Graph Signal Processing on NILM: An Unsupervised Approach Featuring Power Sequences}},
  year = {2023},
  doi = {10.3390/s23083939},
}
Figure