Standards Development

Scope

The proposed document will establish a framework for quantification of environmental impact of AI and its long-term sustainability, and encourage AI developers and users to improve efficiency of AI use. It will also provide a summary of the state of the art of AI technology for direct control and optimisation of energy use in energy systems. The document will provide life-cycle assessment of AI development, deployment and use. 

Emissions that are produced directly by combustion of fossil fuels are Scope 1 emissions. These are observed in transport system and in fossil-fuel energy generators, and the like. AI may help reduce Scope 1 emissions via smart interventions (demand-side response, optimisation of combustion, etc.) Scope 2 are indirect emissions from electricity use, and AI will play a major role in reducing these emissions. Scope 3 are emissions produced during a life cycle of a technology – these emissions are important in assessment of AI solution and will be in scope of this project. Emissions of Scope 4 are the avoided emissions – AI has great potential in quantifying avoided emissions (carbon savings), and the report will address this as well.

In 2015, Semiconductor Industry Association and Semiconductor Research Corporation published report "Rebooting the IT Revolution:

A Call to Action" (https://www.semiconductors.org/wp-content/uploads/2018/06/RITR-WEB-version-FINAL.pdf), where they provided a crude estimate of the future energy use and the required amount of energy for computing due to Internet of Things, directly related of the number of raw bit transitions. They did not consider the AI applications though, and their estimate of the energy production was not accurate, as they did not expect the fast growth of construction of power stations (even those that run on coal, as China is building dozens of those currently because of high energy demand). Given these assumptions, the report still stated that by 2040, digital solutions would exceed the global electric capacity, even without large-scale deployment of extensive AI.

Since this report, the number of AI solutions has grown nonlinearly, and sustainability of energy use for AI becomes crucial for defining the ways of AI development, as well as for the sustainability for the EU power supply.

There are EU-specific aspect of energy use for AI where there are gaps for standardisation - specific energy supply (hydro, nuclear), integrated energy systems (interconnectors), and the sustainability goals (international, such as the Green Deal, and national, such as the Net-Zero-Carbon target). 

The following schematic illustrates the Trans-European energy network with major electric interconnectors and gas pipes. Major international electric connections are marked by blue lines, and international gas pipes by red lines, national and minor links are denoted by grey lines. Major interconnectors are critical for energy linkage in Europe, and their fuel supply are key in assessment of energy carbon footprint, in particular, for AI use.

Figure: Trans-European energy network. 

Nowhere else there is such a dense interconnected network of high complexity that has to be modelled appropriately to estimate the carbon footprint of energy use when energy is imported from a country with another fuel mix. For example, Eastern European countries heavily rely on brown coal, whereas in France up to 75% of electricity is generated by nuclear energy. Their carbon factors will be affected by fuel use as well as the life-cycle energy use.

AI will be instrumental in improving energy efficiency in the Trans-European Energy Network, energy use in demand-side response programmes, and support digital energy services.

To assess the carbon impact of AI, it is necessary to consider life-cycle assessment (LCA) that would take into account the energy use through all stages of AI development: initial deployment, training, distribution and use of the AI solution. This will require a framework of standardisation similar to those that are being developed for other energy-demanding technologies, such as batteries (IEC 63369), with necessary EU specificity – such as the upcoming AI Act.  

Purpose

Energy is key in use of computer technologies, and this is essential for AI in particular, as training and deployment of neural networks may be very energy intensive. In the context of the current energy crisis, as well as long-term challenges of reduction of carbon emissions, AI applications must be assessed for sustainability. 

On the one hand, AI consumes large amount of energy, which should be quantified and reduced where possible. This may be achieved by using pre-trained modules, parallelisation, and optimal algorithms. On the other hand, AI can be used for direct control and optimisation of energy use in many systems, such as built environment with sensor networks, and the like. To make AI green and sustainable, it is necessary to estimate its direct carbon footprint via energy use and its potential as a tool for reducing energy use in Europe. 

It is necessary to establish a framework for quantification of sustainability of AI and encourage AI developers and users to improve efficiency of AI use. According to the EU Energy System Integration Strategy published in 2020, the ongoing reform of the European energy system will link energy sources and infrastructure to support decarbonisation and build a climate neutral EU by 2050. It will help to build modern infrastructure, make European industry more sustainable and competitive, create jobs, and provide clean energy for citizens. AI will play a key role in this integration, and in its turn, it should be sustainable and net-zero-carbon.

The proposed technical report will review and quantify environmental impact of AI use and deployment in the context of the EU integrated energy systems and outline the balancing measures by which AI can be used to optimise energy use at national and international level. Life-cycle assessment of AI will be chartered, and suggestions will be made as to how improve the carbon footprint of the AI technologies in terms of indirect carbon emissions via electricity use.

The work will be coordinated with equivalent activities of ISO SC42 (where a general item is going to be submitted at the ned of 2022), taking into account specific aspect of the European energy system that are not available elsewhere. In particular, sustainable energy supply (hydro, nuclear, wind, solar) provided via the European interconnectors will be taken into account for assessing the AI carbon footprint. On the other hand, AI solutions for optimisation of energy use will be reviewed and quantified to balance the energy use of extensive applications. The proposed documents will also address many of the UN Sustainable Development Goals, as outlined in Section 11.