Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Non-Intrusive Load Monitoring with Features Extracted by Independent Component Analysis

Sahar Moghimian Hoosh, Ilia Kamyshev, Henni Ouerdane

TL;DR

The paper tackles energy disaggregation under data scarcity and high appliance multiplicity by using Independent Component Analysis (ICA) as a physics-aligned feature extractor and a dedicated ICA+ResNetFFN architecture. The unmixing step, $X' = XU^{T}$, aligns with Kirchhoff's laws for linear signal mixing, followed by a 64-dimensional residual network of 15 blocks to perform multi-label classification. Across synthetic and real datasets, the ICA-enabled model outperforms baseline NILM methods in F1-score and shows smoother training dynamics, though real data reveals challenges in maintaining uniform performance across varying appliance counts. The study demonstrates the value of physics-guided feature extraction for accurate, scalable NILM and points to future work on expanding to more appliances and real-world deployment.

Abstract

In this paper, a novel neural network architecture is proposed to address the challenges in energy disaggregation algorithms. These challenges include the limited availability of data and the complexity of disaggregating a large number of appliances operating simultaneously. The proposed model utilizes independent component analysis as the backbone of the neural network and is evaluated using the F1-score for varying numbers of appliances working concurrently. Our results demonstrate that the model is less prone to overfitting, exhibits low complexity, and effectively decomposes signals with many individual components. Furthermore, we show that the proposed model outperforms existing algorithms when applied to real-world data.

Enhancing Non-Intrusive Load Monitoring with Features Extracted by Independent Component Analysis

TL;DR

The paper tackles energy disaggregation under data scarcity and high appliance multiplicity by using Independent Component Analysis (ICA) as a physics-aligned feature extractor and a dedicated ICA+ResNetFFN architecture. The unmixing step, , aligns with Kirchhoff's laws for linear signal mixing, followed by a 64-dimensional residual network of 15 blocks to perform multi-label classification. Across synthetic and real datasets, the ICA-enabled model outperforms baseline NILM methods in F1-score and shows smoother training dynamics, though real data reveals challenges in maintaining uniform performance across varying appliance counts. The study demonstrates the value of physics-guided feature extraction for accurate, scalable NILM and points to future work on expanding to more appliances and real-world deployment.

Abstract

In this paper, a novel neural network architecture is proposed to address the challenges in energy disaggregation algorithms. These challenges include the limited availability of data and the complexity of disaggregating a large number of appliances operating simultaneously. The proposed model utilizes independent component analysis as the backbone of the neural network and is evaluated using the F1-score for varying numbers of appliances working concurrently. Our results demonstrate that the model is less prone to overfitting, exhibits low complexity, and effectively decomposes signals with many individual components. Furthermore, we show that the proposed model outperforms existing algorithms when applied to real-world data.

Paper Structure

This paper contains 14 sections, 4 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Datasets
  4. Algorithms and Models
  5. Proposed model: ICA+ResNetFFN
  6. Baseline models
  7. Temporal Pooling NILM
  8. FIT-PS+LSTM
  9. Fryze+CNN
  10. Experiments
  11. Results
  12. Experiment 1: Synthetic Appliances
  13. Experiment 2: Real Appliances
  14. Conclusion

Figures (8)

  • Figure 1: Architecture of the proposed model, ICA+ResNetFFN. The number of parameters for this model is 65,000.
  • Figure 2: Binary cross entropy loss for four deep learning models.
  • Figure 3: $F_1$-score (sample average) for four deep learning models.
  • Figure 4: Distribution of $F_1$-scores (samples averaging) across different number of simultaneously working appliances for four deep learning models and 2 classical machine learning models.
  • Figure 5: Binary cross entropy loss for four deep learning models.
  • ...and 3 more figures