In September 2017 the French Prime Minister commissioned MP Cédric Villani to lead a mission on artificial intelligence, intended to inform a national strategy. Giving evidence on 12 October, we set out three lessons drawn from twenty years of observing the energy sector.
One preliminary matters: fundamental change in this sector plays out over long timescales. Investment horizons run to decades, infrastructure is not replaced with each technology cycle, IT systems are highly inertial, and so are organisations.
Lesson 1: regulation comes before technology
The main developments of the past twenty years were driven by regulatory decisions before they were driven by technology. That is obvious for market organisation, unbundling, wholesale markets, retail competition. It is equally true of what reshaped the energy landscape: renewables, active demand management, demand response and storage did not result from spontaneous adoption of technologies imposing themselves on merit, but from public policies that chose to support them.
The point goes further. Regulation is rarely the barrier to deployment: it tends to run ahead of demonstrated need. Rules for integrating demand response into the balancing mechanism were found quickly; the framework for individual and collective self-consumption advanced while technical potential remained small. Regulation can even be the driver of technologies that would not have deployed on their own, such as heat metering or communicating meters.
Lesson 2: the marginal value of complexity is often low
This is the most counter-intuitive point, and it holds for market rules as much as for technology.
On rules, France’s demand response trading framework required lengthy consultation and produced 150 pages of technical rules, for marginal use: under 5,000 MWh in 2016 against 550 TWh of block trading, a ratio of one to a hundred thousand. Likewise, nodal pricing in the United States and New Zealand took years to design for results that remain contested.
On technology, two examples. Flexibility serving distribution networks, which can draw on AI for aggregation or constraint forecasting, delivers real but modest net benefits, €10 to 20 million a year against €1.5 billion of annual distribution network investment.
Solar output forecasting is more telling still. Several companies differentiate themselves through model sophistication. But marginal gains in forecast quality carry little value in most use cases: the transmission operator needs neither site-by-site forecasts, since it balances nationally, nor minute-ahead anticipation, since reserves are mobilised automatically. A better-equipped player will therefore struggle to justify a significant price premium.
Lesson 3: AI eases adoption, it does not trigger it
The Nest thermostat, AI-enabled and self-learning, illustrates the limit precisely. Its success fell short of Google’s expectations: around $350 million of annual revenue, below the $400 to 670 million analysts expected to justify the $3.2 billion paid in 2014.
Several explanations compound each other. The service is not perceived as material enough. A promise of 12 to 15 % savings feels abstract to a customer with only a rough idea of their bill, which can vary more than that year to year for weather reasons. A customer on regulated tariffs wanting a 12 to 15 % reduction can more simply switch to a market offer. And the sales channels may be wrong, since the trade, architects, contractors, installers, is probably better placed to persuade.
What this implies for public policy
A well-judged incentive policy will need to lever energy policy priorities in setting its AI agenda, ensure that the complexity created stays controlled and justified, and focus on technologies that carry meaning and value for customers and society.
Blockchain warrants caution in that light. Using it to automate contracts between producers and consumers attracted considerable enthusiasm, but several questions remained open: what is the demonstrated level of demand, is it the only solution, how does it compare with alternatives? Above all, in its financial versions blockchain is extremely energy-intensive, at 39 kWh per transaction and 19 TWh a year for Bitcoin, the equivalent of two nuclear units. Deploying it in service of a virtuous power system deserves at minimum to be questioned.
To go further, download our full analysis below or contact our experts.
