AI and Energy

Recently the energy consumption of AI has been receiving ever more attention. This consumption is enormous and is only expected to rise as a whole, due to the fast-growing use of AI. On the other hand, AI can also play a crucial role in the energy transition. With the help of AI, energy consumption in society could be optimized, for example, and consequently diminished. AI can also ensure that the power grid will be used more efficiently, so that grid congestion can be fought and delays of sustainability projects can be reduced.[1] This blog will dive deeper into the somewhat paradoxical relationship between AI and energy. 

Advanced AI systems, and especially generative AI, need enormous quantities of computing power and therefore energy. Although it is hard to estimate, given the complexity of the systems and the reluctance of companies to share information on their AI systems, how much energy AI is really using, estimates are as plain as day.[2] Research of Stanford University shows that training the ChatGPT 3.5 model costs around 500 tons of CO2 emissions.[3] By comparison, the energy consumption of an average Dutch household is approximately 17 tons of CO2 per year.[4] The use of an AI system like ChatGPT also requires quite a lot of energy. One single prompt to ChatGPT is said to require ten times as much energy as a search request via Google.[5] The total energy consumption of ChatGPT depends on the kind of prompt that is given; generating text costs considerably more energy than generating an illustration.[6] When the increasing use of AI is observed, the conclusion soon follows that the energy consumption of AI can become a cause of concern in the near future. For example, it is estimated that in 2027, AI will use as much energy worldwide as the whole of the Netherlands.[7]   

A solution for limiting great energy consumption need not necessarily lie in a reduced consumption of AI, but in making these systems more sustainable. This can be done in various ways. By optimizing the training process of AI, for example, and only using relevant data, energy consumption can be reduced considerably. This can be done by data pruning, which excludes irrelevant data. In addition, research of Princeton University shows that by ‘growing and pruning’ an AI system energy can also be saved in the use of that system.[8] This technique, which can be compared to the human brain’s operation, prunes unused connections over time, leaving only the most efficient and effective parts of the model. The simpler the system, the less energy is consumed. 

Despite the high consumption of energy, AI also offers enormous opportunities for the energy transition. New AI systems are being developed, which contribute to accelerating the energy transition. Among other things, the following applications can be considered: 

• Predicting the demand for energy: by combining historical data with real-time data, an AI system can carefully predict the demand for energy. With these predictions the waste of energy can be minimized and overburdening during peak hours can be prevented, thus guaranteeing a consistent supply of energy.[9]
• Optimising energy production: the use of renewable energy sources ensures an inconsistent generation of energy. In order to allow a constant distribution of energy, predicting sustainable generation is crucial. AI may be used in this process to create a detailed weather forecast, for example. With this prediction the generation of wind energy can be aligned to the generation of other sources of energy, thus guaranteeing a constant distribution of energy.[10]
• Optimising maintenance: the maintenance of energy systems is very expensive, amongst others because some parts of the system are hard to reach. AI offers ways to make inspection and maintenance to energy systems more effective. Predictive maintenance uses inspection and monitoring data, possibly in combination with external data like weather conditions, to predict the remaining lifetime of parts of the system. Dependent on the outcomes, it is determined where maintenance is required (at once) and which parts of the system can still last. A stable grid is essential for integrating renewable energy sources.[11]
• More efficient use of available capacity: the historical rather conservative approach of grid use leads to it that only 30% of the potential is used on average. AI may play a role in a more efficient use of the grid. By combining several data flows, it can be made visible where and when capacity is available. This will lead to more flexibility on the grid, allowing more grid capacity to be used, without detracting from the grid’s reliability.[12]

In the Netherlands the deployment of AI for the energy transition has already begun. For instance, grid operators are already using AI to optimize maintenance and inspections.[13] Elsewhere too, the use of AI in the energy transition is pursued. The British government gives financial support to AI innovations that support the green transition of the United Kingdom.[14] One supported project engages in improving weather forecasts for solar energy producers, for example.

In the AI Act, the European Union also pays attention to more sustainable use of AI. The standardisation requests to be issued by the European Commission, fulfilling the obligations of high-risk systems, emphasize (among other things) the reduction of energy consumption by AI during their life cycle (see Article 40 of the AI Act). For non-high-risk systems too, the European Union considers sustainable AI to be important. Article 95 of the AI Act prescribes that in the preparation of codes of conduct, minimising the impact of AI systems on environmental sustainability must be paid attention to, including as regards energy-efficient training and use of AI systems. For general purpose AI models, the technical documentation has to contain information about the model’s known or estimated energy consumption (Articles 51 and 53 and Annexes XI and XIII AI Act). 

Despite the fact that the AI Act does not impose specific obligations regarding reduction or sustainable use of energy by AI systems, these provisions do contribute to the broader target of the European Union to develop technologies that are both innovative and environment-friendly, and it can certainly not be excluded that more specific rules in this field will follow in the future. Last year, for example, the European Commission issued a call for tenders for measuring and encouraging energy-efficient and emission-low AI in the EU, which stated the developing of an AI energy and emission label as a main objective.[15] To be continued in any case. 

The AI Act also targets the positive influence AI can have on energy in general. The recitals already mention that the use of AI can “support socially and environmentally beneficial outcomes, for example in (..) energy.” However, this use of  AI in the energy sector has risks too. The AI Regulation provides for those risks. Systems that are used as safety components in the supply of electricity are regarded as critical infrastructure in the AI Regulation (see Annex III AI Act). Security components are systems protecting the physical integrity of this infrastructure. If an AI system is regarded as a critical infrastructure, it is classified as a high-risk AI system and has to meet all the accompanying requirements. Also in the absence of critical infrastructure, the use of AI in energy systems entails regulatory risks. For example, when using data personal data have to be taken into account, and consequently compliance with the GDPR. The use of AI also entails risks in the field of cyber security, and this will have to be reckoned with legally. These aspects will have to be observed when using AI in energy systems.

AI and energy; these two topics may not have much in common at first sight, but the opposite is true. The deployment of AI in the energy transition offers promising options to improve the efficiency and sustainability of our energy supply. However, if such options are used it is crucial to consider the downside as well. In some cases the energy consumption of AI systems may be quite big and may outdo the advantages of using them. To what extent is the use of AI in the energy transition useful if it entails an enormous peak in energy consumption? Especially if we consider this, it seems worthwhile to strive for a sustainable application of AI. If approached correctly, AI may be a strong ally in the energy transition.

[1]Welke rol speelt AI bij het optimaliseren van ons energieverbruik?, solvari.nl.

[2]AI slurpt energie: 'Kan over vier jaar net zoveel stroom als Nederland gebruiken', nos.nl.

[3]AI-index, stanford.edu. 

[4]Wat is je CO2-voetafdruk?, milieucentraal.nl.

[5] AI’s Power Demand: Calculating ChatGPT’s electricity consumption for handling over 365 billion user queries every year, bestbrokers.com.

[6] AI is an enery hog. This is what it means for climate change, technologyreview.com.

[7]AI slurpt energie: 'Kan over vier jaar net zoveel stroom als Nederland gebruiken', nos.nl.

[8] ‘Grow-and-prune’ AI mimincs brain development, slashes energy use, princeton.edu. 

[9]Het net versterken: De rol van AI in energietransformatie, itresearches.com.

[10]Artificial Intelligence voor de energietransitie: 8 uitdagingen voor Nederland, innocationquarter.nl. 

[11]AI als versneller van de energietransitie. Kansen voor een CO2-vrij energiesysteem, nlaic.com.

[12]Chipreus Nvidia komt met AI-oplossing om stroomnet efficiënter te benutten, welingelichtekringen.nl.

[13]Stedin voorspelt en plant onderhoud met ai, computable.nl. 

[14] Government backing for AI businesses to deliver net zero with innovative technologies, gov.uk. 

[15] Artificial Intelligence Act: Call for tenders to measure and foster energy efficient and low emission artificial intelligence in the EU, digital-strategy.ec.europa.eu.