Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
97 tokens/sec
GPT-4o
53 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
5 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Transfer Learning for Non-Intrusive Load Monitoring (1902.08835v3)

Published 23 Feb 2019 in cs.LG and stat.ML

Abstract: Non-intrusive load monitoring (NILM) is a technique to recover source appliances from only the recorded mains in a household. NILM is unidentifiable and thus a challenge problem because the inferred power value of an appliance given only the mains could not be unique. To mitigate the unidentifiable problem, various methods incorporating domain knowledge into NILM have been proposed and shown effective experimentally. Recently, among these methods, deep neural networks are shown performing best. Arguably, the recently proposed sequence-to-point (seq2point) learning is promising for NILM. However, the results were only carried out on the same data domain. It is not clear if the method could be generalised or transferred to different domains, e.g., the test data were drawn from a different country comparing to the training data. We address this issue in the paper, and two transfer learning schemes are proposed, i.e., appliance transfer learning (ATL) and cross-domain transfer learning (CTL). For ATL, our results show that the latent features learnt by a complex' appliance, e.g., washing machine, can be transferred to asimple' appliance, e.g., kettle. For CTL, our conclusion is that the seq2point learning is transferable. Precisely, when the training and test data are in a similar domain, seq2point learning can be directly applied to the test data without fine tuning; when the training and test data are in different domains, seq2point learning needs fine tuning before applying to the test data. Interestingly, we show that only the fully connected layers need fine tuning for transfer learning. Source code can be found at https://github.com/MingjunZhong/transferNILM.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (3)
  1. Michele DIncecco (1 paper)
  2. Stefano Squartini (17 papers)
  3. Mingjun Zhong (20 papers)
Citations (205)

Summary

  • The paper demonstrates that Appliance Transfer Learning (ATL) and Cross-Domain Transfer Learning (CTL) improve NILM model generalization across varying datasets.
  • It employs a seq2point CNN model that maps aggregate mains data to single-point predictions, leveraging fine-tuned dense layers for domain adaptation.
  • Experimental results on REFIT, UK-DALE, and REDD confirm that transfer learning minimizes training data needs while boosting model accuracy for scalable energy monitoring.

Transfer Learning for Non-Intrusive Load Monitoring

The paper "Transfer Learning for Non-Intrusive Load Monitoring" explores advanced methodologies in the domain of monitoring household appliance energy consumption using Non-Intrusive Load Monitoring (NILM). This approach addresses the challenge of accurately disaggregating energy consumption without the need for intrusive and costly per-appliance sensors, focusing instead on using the aggregate mains reading for inferences. A significant challenge highlighted is the inherent unidentifiability of NILM, due to the non-uniqueness of recovery of appliance-level consumption from aggregate data.

This research contributes to NILM literature by proposing two transfer learning schemes: Appliance Transfer Learning (ATL) and Cross-Domain Transfer Learning (CTL), which are intended to enhance the generalizability of the NILM models and improve their performance across different domains, such as geographical regions or different appliances within the same household.

Methodology and Experimental Design

The paper employs seq2point learning, a distinctive deep learning model that maps sequences of mains data to a single point prediction—a midpoint in the input window, using Convolutional Neural Networks (CNNs). This method, established in prior studies as effective within a single domain (e.g., trained and tested on datasets from the same country), lacked validation across diverse data domains.

To address this, the authors conduct extensive experiments using multiple datasets including REFIT, UK-DALE, and REDD. REFIT serves as a robust training dataset due to its extensive coverage of different houses in the UK. For ATL, the ATL strategy considers applying the CNN layers pre-trained on one appliance, such as a washing machine, to other appliances, founded on the observation that appliances share common operational power characteristics such as ON-OFF changes and power levels.

CTL, on the other hand, investigates cross-domain applicability by evaluating seq2point model performance across different geographic locations. Initially, models trained on REFIT were directly tested on UK-DALE and REDD. Fine-tuning was then applied by adjusting only the fully connected layers of the neural network, leveraging already learned CNN features to align the model with specific domain characteristics.

Findings and Implications

The experimental results indicate that ATL can efficiently transfer knowledge from one appliance (e.g., washing machine) to others (e.g., kettle or microwave), reducing the need for training large and diverse data for each specific appliance. This implies reduced hardware dependency, as models developed for one appliance can be rapidly adapted for others with minimal additional training.

In terms of CTL, when testing seq2point performance on REFIT and UK-DALE—both hailing from the same country—minimal-to-no fine-tuning was required. However, in scenarios involving significantly different domains such as REFIT and REDD, fine-tuning only the dense layers led to improved model accuracy, evidencing the method’s potential for cross-context adaptation, pivotal for widespread real-world application.

The practical implications are substantial, hinting at the feasibility of creating generalized NILM models deployable worldwide, tailored with minor training adaptations based on specific local usage patterns. This aligns with efforts towards global energy efficiency, offering a scalable and cost-effective solution for energy monitoring and management.

Conclusion

In conclusion, this paper advances the NILM field by demonstrating tangible benefits and methods of applying transfer learning strategies to enhance model generalization and efficiency. Further research directions may include exploring different network architectures for deeper understanding and optimization of transfer learning strategies in NILM, particularly across more diverse domains and appliance sets. This paper lays fundamental groundwork for such explorations and highlights the critical importance of flexible, generalized learning methodologies in sustainable energy solutions.

Github Logo Streamline Icon: https://streamlinehq.com