This article is written by Lina Stankovic, senior lecturer and member of the Ethical AI Network, and Vladimir Stankovic, reader at the University of Strathclyde

You’ve probably never thought of your gas or electricity meter as being at the cutting edge of science. But in a way, these unassuming, practical gadgets offer a glimpse into the future, and can teach us valuable lessons about the pitfalls and benefits of machine learning.

Smart meters are the next generation electricity and gas meters that measure household energy consumption. What makes the traditional, well-known meters "smart", is that they use a secure data network to automatically and wirelessly send your meter readings to your energy supplier at least once a month in addition to being easier to read for the consumer. This means customers receive accurate bills, and also means fewer visits by utility engineers to manually read meters.

With the widescale rollout of smart meters, which now numbers more than 16 million across the UK, and much more worldwide, significant research and innovation is happening to make the most out of the data that is being collected.

Some of the main goals for researchers and utilities is to create timely and accurate billing, increase the understanding of energy use in the home, ease the transition to renewable energy and electric vehicles, and better management of generation and distribution of electricity.

This, in turn, enables households and utilities to cut costs and meet climate change and energy efficiency goals through reducing unnecessary energy use.

But how do we make all the above happen? The answer lies in artificial intelligence (AI).

What the kettle says about you

Today we have the ability to use the power of AI to analyse data from smart meters using machine learning to improve the way we consume energy. For example, it is easy for an algorithm to "learn" how much energy is consumed from historical data and predict patterns of consumption for the future using time-series machine learning algorithms, helping the AI engine make a decision about demand-response and thus ensuring energy security of supply. The next step up is inferring when and how often particular appliances are used in the home through load disaggregation or non-intrusive load monitoring (NILM) algorithms.

These algorithms break down individual "loads" — from your oven, your kettle and your television, etc — that contribute to the aggregate meter reading at any one time through signal processing, and supervised and unsupervised machine learning algorithms. This in turn enables the AI engine to make recommendations on appliance upgrades, generate a life-cycle assessment (LCA) per appliance or food product and statistically quantify the type of activity we spend our time on. However, this also raises clear ethical challenges.

It can also be misused for targeted advertising for coffee or tea, or inferring when a household has breakfast or dinner, which can again be used to infer what their working hours are

For example, it is now possible to infer how often an individual puts the kettle on, and how much water was used to fill the kettle. This is useful in order to provide energy saving recommendations and track forgetfulness arising from old age or as a precursor to dementia. However, it can also be misused for targeted advertising for coffee or tea, or inferring when a household has breakfast or dinner, which can again be used to infer what their working hours are. Similarly, it is also possible to infer when an individual is cooking, when and how often they do their laundry, have a shower, etc.

Besides inferring when an individual is at home, when they are sleeping, when they are hosting guests, and what are their daily energy usage patterns, it is also possible to infer, through load disaggregation algorithms, which appliances are being used and, indirectly, which domestic activities individuals are engaged in at home, from their smart meter data.

For this reason we need to ensure that the use of machine learning in smart meters doesn’t infringe on the privacy and security of users, and that the AI technology itself is transparent to users that include consumers, government and businesses producing the technology. These stakeholders must have a clear understanding of who has access to the smart meter measurements and the information generated from analysis.

This has a direct implication on trustworthiness of the technology and assigning liability.

Addressing privacy concerns

Over the past couple of year, frameworks and guidelines have been released on general data protection and ethical handling of experiments on individuals and personal data from governing and charitable bodies such as the UK research councils, Institute of Electrical and Electrical Engineering (IEEE), Institute of Engineering and Technology (IET), British Computer Society (BCS), AI4People, and EU High-level expert group on responsible AI.

It is very difficult to generalise the type and level of privacy and responsibility of an AI system, as context is absolutely critical — guidelines for handling experiments on animal testing are not quite the same as guidelines for AI algorithms online that recommend products based on a person’s shopping history.

Pertinent to the smart metering context, the Data Protection and Impact Assessment (DPIA) is a major step towards data privacy. Developed by the European Commission Smart Grid Task Force and the UK Information Commissioners’ Office, two bodies that are dedicated to identifying appropriate regulatory scenarios and recommendations for data handling, data security and data protection, the DPIA offers a template for Smart Grid and Smart Metering Systems within a data privacy and security framework that both protects and enables privacy.

Key to all this is the identification of risks — expected and unexpected — for the particular data at hand

The DPIA is a step in the right direction as it sets out processes to build and demonstrate compliance with GDPR. Firstly, the DPIA sets out types of data processing operations and places strong emphasis on identifying risks to privacy and personal data at multiple stages of a project, i.e. during the design of the system, before the usage of the new system, and throughout the usage of the system. This ensures that all risks are captured, including expected and unexpected types of usage of the system.

Key to all this is the identification of risks — expected and unexpected — for the particular data at hand. Experts are making significant breakthroughs in what kind of information can be inferred about a household and individuals from their smart meter data.

New applications are constantly being developed that push the boundaries of what smart meters can do. Some of these include:

With the growing popularity of smart home technologies, such as Google Assistant, Amazon’s Alexa and Apple’s Siri, it is easier to make additional inferences about an individual in their home and provide advanced control functionality over appliances and space heating that could potentially be harmful to the individual.

In conclusion, further effort is needed to clearly identify the benefits and risks that are opened up by AI engines leveraging machine learning algorithms before adopting them for decision making. The machine learning algorithms must be made transparent to users to ensure trust. Investigation into appropriate data transformation tools to alter the data is needed, so that it is useful but not harmful, and recommendations can be inferred, depending on the level of individual consent to sharing their data. It is also important to explain clearly to individuals what the benefits and risks are when their personal data is being shared for analysis.

As the technology evolves, more can be inferred from smart meter data, so the review of benefits and risks is a continuous process

As the technology evolves, more can be inferred from smart meter data, so the review of benefits and risks is a continuous process. This should be performed by a range of stakeholders and competencies, including those understanding the sensing elements of the meter and measurement signals to those making inferences using machine learning, and those managing and offering the services.

While machine learning technology opens up risks of misuse and misinterpretation of data, this technology is also the solution to provide accountability as personal meter data flows from the home, where it is captured to the cloud, to the technology provider(s) and finally end-user. — Lina Stankovic and Vladimir Stankovic

(Picture Credit: Unsplash)