Deceiving Customers is Never a Good Idea
In the last few years, more and more energy management software companies claim to provide Non-Intrusive Load Monitoring (NILM) or Artificial intelligence based appliance recognition algorithms. A closer look allowed us to discover there are lots of false claims.
NILM is not to be confused with a simple apportionment of energy based on a statistical analysis of energy breakdown found in types of buildings with similar conditions. Let’s shed some light on what a NILM algorithm is and how to identify technology humbugs.
Definition
As defined by Wikipedia and many research articles, NILMs or appliance recognition software are algorithms that detect changes in the electrical values (power, current, voltage) going into a building to infer what appliances are used in the building as well as their individual energy consumption.
In an early study from 2010 by Wattics and University College Dublin published under the name of RECognition of electrical Appliances and Profiling in real-time (RECAP) Wattics used a single energy meter clipped at the main electrical board to disaggregate appliances. The picture below shows that signatures based on variations of power could be associated with the turning ON/OFF of certain appliances.
Advantages and Limits
The key advantage of NILM algorithms is the ability to infer and breakdown energy consumption with one meter rather than clipping current transformers to each circuit to monitor. This aims at reducing the installation complexity and cost.
Unfortunately, although NILM looks great on paper, in practice this type of algorithms have a number of key limitations:
- Data Transmission. NILM cannot be achieved with daily or monthly electrical readings from the meter. The electrical meter must report data at a high frequency that is at least several data packets per minutes or second-based transmission, in order to capture and build a database of electrical load signatures.Unfortunately, the vast majority of electrical meters can report every 5-min, half-hourly, hourly or daily. In addition, if electrical meter data is queried and transmitted at high frequency to a cloud server for creating signatures and analysis, the transmission and data storage cost would likely exceed the cost of installing an extra set of current transformers.
- Profiling Issues. Commercial and industrial facilities have a large number of dynamic loads with a different mode of operations that makes almost impossible to capture signatures. Although NILM algorithms have been successful in a more confined domestic environment, the application of NILM is an industrial environment requires specialised meters with high-frequency power quality data capture that may include harmonics.
- Other Commercial Challenges. The expectation from non-technical users is that the software is going to recognise all or most of the loads in a building. Even if with high-sampling data frequency hardware, this is not often the case because certain small loads may not have a distinguishable signature or may not be part of the signature database hence making difficult to provide a clear picture to the client of what can/cannot be achieved.
- IPMVP saving validation. In case the energy professional requires validation of savings such as IPMVP-based calculation, NILM cannot be used as an alternative method for Measurement & Verification at the appliance level, hence not helping to achieve ISO 50001 for energy-efficiency projects or other certifications.
- Tenant Billing. NILM cannot be applied for apportionment of utility billing per tenant or per appliance because there is no certification such as UL or MID that is required to produce bills.
Wattics and NILM
NILM was part of the Wattics platform in early 2011-12 and was used in multiple projects. This helped to test at scale and identify the advantages but also that the technology could be misleading for industrial and commercial sectors although performing at an acceptable level for domestic use.
Considering that the Wattics business model was focusing on business analytics a decision was made to remove NILM from the technology. We made this decision in order to provide more precise analytics and set up a highly accurate data standard for projects in commercial and industrial sectors.
Instead, the NILM technology set the basis for the Sentinel machine-learning analytics that currently assists the energy professionals to identify abnormal energy patterns in real-time.
Conclusion
Although good on papers, NILM algorithms only work with specialised systems that can collect and process data at high frequency hence often fail to deliver the promised cost savings and application at scale.
NILM is not to be confused with a much simpler apportionment of energy based on a statistical analysis that split energy based on similar type of sites or buildings.
Currently, NILM cannot be applied using minute-based off-the-shelf meters for industrial and commercial sectors as a precise submeter data source disaggregation.
