[Paper Review] Subtask Gated Networks for Non-Intrusive Load Monitoring

Changho Shin, Sunghwan Joo+4 more|arXiv (Cornell University)|Nov 16, 2018

Smart Grid Energy Management23 references20 citations

TL;DR

This paper proposes Subtask Gated Networks (SGN), a deep learning architecture that improves non-intrusive load monitoring (NILM) by jointly training a regression network for power estimation and a classification subnetwork for on/off state detection. By gating the regression output with the classification probability, SGN achieves 15–30% better performance than state-of-the-art methods on REDD and UK-DALE datasets, especially when standby power is modeled.

ABSTRACT

Non-intrusive load monitoring (NILM), also known as energy disaggregation, is a blind source separation problem where a household's aggregate electricity consumption is broken down into electricity usages of individual appliances. In this way, the cost and trouble of installing many measurement devices over numerous household appliances can be avoided, and only one device needs to be installed. The problem has been well-known since Hart's seminal paper in 1992, and recently significant performance improvements have been achieved by adopting deep networks. In this work, we focus on the idea that appliances have on/off states, and develop a deep network for further performance improvements. Specifically, we propose a subtask gated network that combines the main regression network with an on/off classification subtask network. Unlike typical multitask learning algorithms where multiple tasks simply share the network parameters to take advantage of the relevance among tasks, the subtask gated network multiply the main network's regression output with the subtask's classification probability. When standby-power is additionally learned, the proposed solution surpasses the state-of-the-art performance for most of the benchmark cases. The subtask gated network can be very effective for any problem that inherently has on/off states.

Motivation & Objective

• Address the limitations of existing deep learning models in NILM that do not explicitly model appliance on/off states.
• Improve energy disaggregation accuracy by leveraging the inherent on/off behavior of household appliances.
• Develop a novel architecture that integrates regression and classification subtasks through a gating mechanism for better generalization.
• Demonstrate that explicit modeling of on/off states via a subtask network enhances performance beyond standard multitask learning approaches.
• Evaluate the method on standard NILM benchmarks (REDD and UK-DALE) under multiple metrics including MAE and SAEδ.

Proposed method

• Propose a subtask gated network (SGN) that combines a main regression network with a separate on/off classification subnetwork.
• Gate the regression output by element-wise multiplication with the classification subnetwork’s predicted probability of being 'on'.
• Train the model end-to-end using a composite loss function that includes both regression loss (L_output) and on-state classification loss (L_on).
• Introduce a variant that explicitly models standby power to improve performance on appliances with low-power states.
• Apply hard gating in some variants to enforce binary decisions, enhancing robustness in noisy conditions.
• Use the classification subnetwork as a dynamic gate, ensuring that regression outputs are suppressed when the appliance is predicted to be off.

Experimental results

Research questions

• RQ1 Can explicit modeling of appliance on/off states improve the accuracy of non-intrusive load monitoring using deep learning?
• RQ2 How does gating the regression output with on/off classification probability affect disaggregation performance compared to standard multitask learning?
• RQ3 Does incorporating a dedicated loss for on-state classification (L_on) lead to better generalization and reduced noise in regression outputs?
• RQ4 To what extent does modeling standby power improve performance on appliances with low-power states?
• RQ5 How robust is the subtask gated network across different error metrics (e.g., SAEδ) and varying thresholds (δ) in real-world applications?

Key findings

• Subtask Gated Networks achieve 15–30% better performance than state-of-the-art deep learning models on the REDD and UK-DALE datasets using MAE and SAEδ as metrics.
• The inclusion of the on-state classification loss (L_on) significantly improves the reliability of the gating mechanism, preventing the classification subnetwork from becoming a mere scaling factor.
• The classification subnetwork’s output is highly concentrated near 0 and 1, indicating effective learning of on/off states, especially for cycling appliances like fridges.
• For appliances like dishwashers and kettles, the gating mechanism effectively suppresses noisy regression outputs during off-states, improving accuracy.
• The model with explicit standby power modeling further improves performance on appliances with low-power states, demonstrating the benefit of fine-grained state representation.
• The SGN model maintains superior performance across all tested values of δ in SAEδ, indicating robustness to varying error tolerance levels in practical applications.