Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Federated Sequence-to-Sequence Learning for Load Disaggregation from Unbalanced Low-Resolution Smart Meter Data

Xiangrui Li

TL;DR

This work tackles NILM under hourly, low-rate smart-meter data by introducing a privacy-preserving federated sequence-to-sequence approach that leverages weather information to disaggregate total household load into 12 appliances. It advances the field by employing the L2GD federated framework to balance global and local models, substantially reducing communication rounds while handling data heterogeneity. Empirical results on the Pecan Street dataset show that weather features significantly improve disaggregation accuracy and that L2GD outperforms FedAvg and FedProx in heterogeneous settings, achieving efficient, privacy-preserving NILM for low-resolution meters. The approach has practical implications for scalable, real-world energy monitoring and program design without compromising user privacy.

Abstract

The importance of Non-Intrusive Load Monitoring (NILM) has been increasingly recognized, given that NILM can enhance energy awareness and provide valuable insights for energy program design. Many existing NILM methods often rely on specialized devices to retrieve high-sampling complex signal data and focus on the high consumption appliances, hindering their applicability in real-world applications, especially when smart meters only provide low-resolution active power readings for households. In this paper, we propose a new approach using easily accessible weather data to achieve load disaggregation for a total of 12 appliances, encompassing both high and low consumption, in scenarios with very low sampling rates (hourly). Moreover, We develop a federated learning (FL) model that builds upon a sequence-to-sequence model to fulfil load disaggregation without data sharing. Our experiments demonstrate that the FL framework - L2GD can effectively handle statistical heterogeneity and avoid overfitting problems. By incorporating weather data, our approach significantly improves the performance of NILM.

Federated Sequence-to-Sequence Learning for Load Disaggregation from Unbalanced Low-Resolution Smart Meter Data

TL;DR

This work tackles NILM under hourly, low-rate smart-meter data by introducing a privacy-preserving federated sequence-to-sequence approach that leverages weather information to disaggregate total household load into 12 appliances. It advances the field by employing the L2GD federated framework to balance global and local models, substantially reducing communication rounds while handling data heterogeneity. Empirical results on the Pecan Street dataset show that weather features significantly improve disaggregation accuracy and that L2GD outperforms FedAvg and FedProx in heterogeneous settings, achieving efficient, privacy-preserving NILM for low-resolution meters. The approach has practical implications for scalable, real-world energy monitoring and program design without compromising user privacy.

Abstract

The importance of Non-Intrusive Load Monitoring (NILM) has been increasingly recognized, given that NILM can enhance energy awareness and provide valuable insights for energy program design. Many existing NILM methods often rely on specialized devices to retrieve high-sampling complex signal data and focus on the high consumption appliances, hindering their applicability in real-world applications, especially when smart meters only provide low-resolution active power readings for households. In this paper, we propose a new approach using easily accessible weather data to achieve load disaggregation for a total of 12 appliances, encompassing both high and low consumption, in scenarios with very low sampling rates (hourly). Moreover, We develop a federated learning (FL) model that builds upon a sequence-to-sequence model to fulfil load disaggregation without data sharing. Our experiments demonstrate that the FL framework - L2GD can effectively handle statistical heterogeneity and avoid overfitting problems. By incorporating weather data, our approach significantly improves the performance of NILM.
Paper Structure (13 sections, 6 equations, 4 figures, 5 tables, 2 algorithms)

This paper contains 13 sections, 6 equations, 4 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related works
  3. methodology
  4. Sequence-to-Sequence Model
  5. Federated Learning
  6. Experiments
  7. Data Description
  8. Model Configurations
  9. Numerical Results and Discussion
  10. Performance Metrics
  11. Comparison of Federated learning frameworks in homogeneous and heterogeneous datasets
  12. Effectiveness of Weather Feature to Time-Series Model
  13. Conclusion and Future Work

Figures (4)

  • Figure 1: Proposed Sequence to Sequence model.
  • Figure 2: Attention units implemented in Sequence-to-Sequence model.
  • Figure 3: The overview of L2GD framework.
  • Figure 4: Average training loss with communication rounds increasing.