Integrating Submetering and NILM for Building Smart Readiness

This article examines how submetering (Intrusive Load Monitoring- ILM) and Non-Intrusive Load Monitoring (NILM) can enhance smart readiness in residential buildings. Submetering provides precise data on energy use across different zones or systems, enabling detailed audits, cost allocation, and performance optimisation, though it often requires costly infrastructural changes. NILM offers a more scalable alternative by using a single meter and advanced algorithms to estimate energy use by individual appliances, though its accuracy can be challenged by complex loads. The proposed framework introduces both methods, allowing for a step-by-step increase in monitoring complexity as needed. By enabling finer control of individual loads in a building, this framework contributes to achieving smart readiness [1] in residential buildings for more sustainable and intelligent living environments by optimising energy efficiency, operational performance, and occupant engagement while reducing costs.

Keywords: Internet of Things (IoT), Intrusive Load Monitoring (ILM), Non-intrusive Load monitoring (NILM), Machine Learning.

Introduction

This article examines how submetering (i.e., Intrusive Load Monitoring- ILM) and Non-Intrusive Load Monitoring (NILM) enhance smart readiness in residential buildings. A framework is developed that can gradually increase monitoring complexity. There are various types of energy monitoring technologies. For example, ILM submetering provides precise data on energy usage across different zones or systems, enabling detailed audits, cost allocation, and performance optimization. However, it often requires costly infrastructural changes. NILM offers a more scalable alternative by using a single meter and advanced algorithms to estimate energy use by individual appliances, though its accuracy can be challenged by complex loads. The proposed framework combines both methods, allowing for a step-by-step increase in monitoring complexity as needed. Initial machine learning (ML) results from a residential case study show significant improvements in energy disaggregation accuracy. By optimising energy efficiency, operational performance, and occupant engagement while reducing costs, this framework offers a practical path to achieving smart readiness [1] in residential buildings, contributing to more sustainable and intelligent living environments.

To address climate change, governments worldwide are increasingly prioritising energy efficiency and reduction to lower global greenhouse gas emissions, with the goal of limiting global temperature rise to below 2°C [2]. Building operations are a significant contributor to rising CO₂ emissions, which have increased by 5.4% since 2015. This has prompted global policy shifts, including stricter emission standards for new buildings, such as the net-zero mandates detailed in the revised Energy Performance and Buildings Directive (EPBD) [3]. While smart readiness has become a focal point in optimising building performance, enhancing occupant comfort, and supporting sustainability goals [1], discussions on the importance of load monitoring remains limited.

Appliances in residential and commercial buildings account for an increasing amount of energy consumption with plug and process loads accounting for majority of total energy consumption in commercial buildings [4]. Significant energy savings could be actualised through effective load management through personalised recommendation and appliance usage feedback may result in more than 12% energy savings [4]. For example, highlighting the largest contributors to their total energy usage. Furthermore, energy suppliers can utilise this information to maintain grid stability through peak load management and demand response management through a better understanding of consumer behavioural patterns. An Intelligent Load Analytics framework can be utilised to provide fine-grained control over appliances in buildings, provide detailed insights on energy consumption and methods to reduce it. An overview of Intelligent Load Analytics framework (as shown in Figure 1) proposed in this paper consists of 3 modules: Load Monitoring, Intelligent Analysis and Feedback modules. Their application is shown in Figure 2.

Figure 1. Intelligent Load Analytics framework.

Load Monitoring Module

The purpose of load monitoring is to identify the appliances drawing power in a building and collect information about their power consumption. There are two approaches to monitoring power consumption of appliances: Non-Intrusive Load Monitoring (NILM) or Intrusive Load Monitoring (ILM). Introduced by Hart in 1992 [5], NILM techniques aim to separate aggregate data into individual loads where the NILM problem formulation is summarized by Equation (1):

Where Y(t) is the aggregate signal which consists of a sum of M individual appliance loads and the ε noise and approximation error.

Figure 2. Diagram displaying ILM and NILM implementations and subsequent edge or cloud processing.

Aggregate data is collected through a custom sensor installed directly to the main supply or data extracted through existing smart meters. NILM research has been popular due to lower installation and maintenance costs associated with single-point sensing. Additionally, this reduces the intrusion experienced by occupants in residential settings, improving consumer buy-in. Despite this, NILM suffers from significant downsides. Due to the complexity of the disaggregation problem, NILM research struggles to deal with a wide range of appliances, especially when multiple appliances of the same type are utilized simultaneously [6]. Low frequency NILM research often struggles to identify low powered appliances [7]. This is problematic in commercial buildings where there are many similar appliances with overlapping power profiles such as computers, laptops, and printers.

On the other hand, ILM systems circumvent disaggregation using direct sensing, using smart plugs or additional circuitry, to submetering to a specific zone, plug or appliance. Increasing abundance in higher resolution smart plugs in the market has made it more viable to retrofit ILM to existing infrastructure resulting in a surge in their use in home automation. Direct access to a plug or appliance simplifies appliance classification as there is less noise and offers a whole suite of advantages such as increased localisation and low latency. This offers a method that is scalable to a wide range of appliances and effective for appliances of the same type. However, the cost of installing sensors or smart plugs into households rapidly scales in both residential and commercial settings where the former often struggles with consumer buy-in due to its intrusive nature.

Modern smart plug solutions offer up to 1Hz sampling frequency capable of enhanced control of individual appliances and real-time monitoring. A scalable solution would involve applying machine learning and deep learning models for automatic appliance classification that can be updated OTA (Over-The-Air) to reduce setup and maintenance costs. Figure 3 shows a confusion matrix describing the performance of a lightweight random forest model that segments real-time power data into 60 second windows utilising statistical and frequency-based features from raw power data. The prediction of microwave and toaster has a lower performance due to the similarity of their power signals, but this can be resolved by utilising additional features such as power factor or apparent power data and utilising data of a higher sampling frequency.

To train these models, the datasets must be applicable to the target market. Using a dataset from a different country introduces challenges due to differences in electrical infrastructure and household appliance usage patterns driven by cultural behaviours. Electrical infrastructure often differs in voltage standards and grid stability. Furthermore, appliance types might have different power consumption characteristics and daily routines based on mealtimes and leisure activities, climate effect temporal aspects of data. All these factors may reduce generalization to different areas but may benefit from combining datasets by supplementing local data or using domain adaptation. For example, US homes use 120V compared to 230V in the UK with larger homes, prevalent use of air conditioning and different cooking habits on average compared to the UK. Indian homes often experience power outages and voltage fluctuations compared to UK homes. EU countries might have similar voltage standards but may differ in appliance types and energy efficiency regulations. Unfortunately, there are only few publicly available datasets that have appliance level data. The most prevalent UK datasets with appliance level data include UKDALE [7] and IDEAL [8] both of which have sampling rates larger than 5 seconds.

Figure 3. Confusion matrix displaying the performance of a random-forest model that automatically classifies raw power data to appliance labels. The diagonal highlights the accuracy of the model for each appliance.

The choice between ILM, NILM and their respective techniques relies on the downstream application and the data available to train the model. The decision hinges on the latency required from the down-stream application and temporal resolution of the training data which are often inter-linked. The latency can be broken down into real-time or near real time and batch processing. The former is important for time-sensitive tasks such as an outlier detection for catastrophic failure or unauthorized access to an appliance or zone which benefits from ILM infrastructure. Batch processing is effective in NILM due to the added disaggregation task in NILM which introduces additional complexity.

Intelligent Analysis Module

The purpose of this module is to collect data for insights, detect anomalies and optimize the framework. Appliance data collected can be used to train outlier detection models that can be used to identify anomalies in appliances that may indicate inefficiencies or faults that can trigger notifications to inform corrective actions. For example, load monitoring of appliances has also been shown to be a possible avenue for the early prevention of dementia by classifying activity levels and detecting outliers [9] and detecting and isolating electrical faults in shipboard appliances enabling early intervention to prevent costly maintenance and failure [10].

Moreover, statistical analysis can be performed on the data to create energy audits that can inform control algorithms that adjust appliance activation or usage based on either manually specified policies or policy configurations. Additionally, optimisation algorithms such as demand response, load forecasting and energy storage management can be utilised to predict, analyse and optimise existing processes in a building and the grid.

Feedback Module

This module helps tune and retrain existing models and provides actionable insights, recommendations and automated responses to improve efficiency. For example, dashboards for real-time monitoring can help reduce energy usage in buildings and provide custom reports to different stakeholders for optimisation. Automated feedback loops can be integrated with building management systems to automatically adjust settings such as lighting, HVAC, electric heaters or fans to build a continuous learning system that adapts to feedback from users.

Conclusion

Adopting ILM and NILM has and will advance energy management by changing behavioural patterns in consumers through actionable feedback, providing fine-grained control to building management systems and improving grid optimisation through detailed insights on consumption patterns.  Nevertheless, smart load monitoring, through ILM and NILM, offers detailed insights into energy usage patterns, enabling precise optimisation of building systems for efficiency and sustainability. By providing granular data on individual appliance consumption, these technologies lay a strong foundation for future smart readiness by facilitating targeted energy management and integration with intelligent building systems. Looking forward, NILM’s current capabilities are limited by dataset availability and model complexity however ongoing advancements in model architectures suggest that it could eventually close the gap between ILM and NILM where NILM could be most dominant technology offering both comprehensive coverage and precision.

References

[1]     European Commission (2020). Smart Readiness Indicator for Building https://energy.ec.europa.eu/topics/energy-efficiency/energy-efficient-buildings/smart-readiness-indicator_en.

[2]     Paris Agreement to the United Nations Framework Convention on Climate Change, Dec. 12, 2015, T.I.A.S. No. 16-1104.

[3]     United Nations Environment Programme (2022). 2022 Global Status Report for Buildings and Construction. Global Alliance for Buildings and Construction. https://www.unep.org/resources/publication/2022-global-status-report-buildings-and-construction.

[4]     Armel, K. Carrie et al. (2013). Is disaggregation the holy grail of energy efficiency? The case of electricity, Energy Policy, Volume 52, 2013, Pages 213-234, ISSN 0301-4215. https://doi.org/10.1016/j.enpol.2012.08.062.

[5]     Hart, George W.. “Nonintrusive appliance load monitoring.” Proc. IEEE 80 (1992): 1870-1891.

[6]     Hasan Rafiq, Prajowal Manandhar, Edwin Rodriguez-Ubinas, Omer Ahmed Qureshi, Themis Palpanas, A review of current methods and challenges of advanced deep learning-based non-intrusive load monitoring (NILM) in residential context, Energy and Buildings, Volume 305, 2024, 113890, ISSN 0378-7788, https://doi.org/10.1016/j.enbuild.2024.113890.

[7]     A. Marchiori, D. Hakkarinen, Q. Han, and L. Earle, “Circuit-Level Load Monitoring for Household Energy Management,” IEEE Pervasive Computing, vol. 10, no. 1, pp. 40–48, 2011.

[8]     Kelly, J., Knottenbelt, W. The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes. Sci Data 2, 150007 (2015). https://doi.org/10.1038/sdata.2015.7

[9]     Pullinger, M., Kilgour, J., Goddard, N. et al.The IDEAL household energy dataset, electricity, gas, contextual sensor data and survey data for 255 UK homes. Sci Data 8, 146 (2021). https://doi.org/10.1038/s41597-021-00921-y

[10]   Wang, J.; Spicher, N.; Warnecke, J.M.; Haghi, M.; Schwartze, J.; Deserno, T.M. Unobtrusive Health Monitoring in Private Spaces: The Smart Home. Sensors 2021, 21, 864.

[11]   P. A. Lindahl, D. H. Green, G. Bredariol, A. Aboulian, J. S. Donnal and S. B. Leeb, "Shipboard Fault Detection Through Nonintrusive Load Monitoring: A Case Study," in IEEE Sensors Journal, vol. 18, no. 21, pp. 8986-8995, 1 Nov.1, 2018. https://doi.org/10.1109/JSEN.2018.2869115.

[12]   Zeynep Duygu Tekler, Raymond Low, Lucienne Blessing, User perceptions on the adoption of smart energy management systems in the workplace: Design and policy implications, Energy Research & Social Science, Volume 88, 2022, 102505, ISSN 2214-6296, https://doi.org/10.1016/j.erss.2022.102505.