Agentic reasoning and the age of electricity
Currents: AI & Energy Insights - October 2024
Welcome back to Currents, a monthly column from Reimagine Energy dedicated to the latest news at the intersection of AI and Energy. Every last week of the month, I send out an expert-curated summary of the most relevant updates from the sector.
They say good things come to those who wait! This month’s Currents is a week late but filled with exciting developments in AI and Energy.
1. Industry news
The agentic reasoning era has begun, accompanied by a surge in energy demand
The release of OpenAI’s o1 marked a shift in the world of large language models, now moving from “thinking fast” to “thinking slow.” This advancement offers more powerful AI capabilities, but also comes with a considerable increase in energy consumption for each query. What does this mean for our energy landscape?
Recent data from the International Energy Agency’s World Energy Outlook 2024 highlight a surge in global electricity demand. Over the next decade, we’re set to add the equivalent of Japan’s annual electricity consumption to the grid each year. “In energy history, we’ve witnessed the Age of Coal and the Age of Oil – and we’re now moving at speed into the Age of Electricity” said Fatih Birol, the IEA’s executive director.
While AI holds immense potential for humanity, some experts express concern that “generative AI is making the hardest problem we’ve ever had to solve that much harder to solve,” as noted in this MIT Technology Review article.
The need of large AI players to quickly access power generation is also creating social and health concerns. In Southwest Memphis, for example, xAI is building what aims to be the world’s largest supercomputer. But without approval to draw from the main grid yet, they deployed natural gas-powered mobile generators, further contributing to local pollution.
What I’m thinking: I’m a techno-optimist with respect to these issues. AI is propelling humanity at an exponential pace and I believe we can find technological solutions to the energy demand problem. In fact, most of the solutions are already available, we just haven’t found an effective way to deploy them at scale. We’re just at the dawn of the LLM age, recent research suggests we can reduce the energy cost of LLMs by 95%, and more and more research will come out on this topic. While it’s true that AI’s energy demand is creating immediate challenges, we can’t turn back now. Pandora’s box is open. All we can do is roll up our sleeves and get to work.
The supply debate: renewables vs. nuclear
The surge in energy demand has increased the focus on energy supply as well. I’ve gathered insights to understand the current state and future of energy production:
Last month, the last British coal power plant closed. The UK’s National Grid System Operator claims that blackout risks are very low, thanks to newly installed wind capacity and interconnections with Denmark. However, challenges remain, as wind lulls can last for days, posing stability issues for the electricity system.
An insightful analysis by fellow Substack author Thomas Pueyo explores solar energy’s potential. Projections suggest we’re at the beginning of the S-curve for solar electricity. Its growth is accelerating, and the 2020s through 2030s will see a massive increase in installed capacity. The cost of solar energy is dropping as planned; it may shrink by eight times in the coming decade.
However, a key challenge for renewables is their decreasing market value. As more solar capacity is installed, oversupply during sunny periods drives electricity prices down. This makes new solar installations less attractive and existing projects less profitable.
Because of their decreasing market value, some energy experts are arguing that wind and solar are the energy past, while the future is nuclear power. While I don’t fully agree with their analysis, I believe it’s important to consider alternative perspectives for informed energy policy decisions in the coming decades.
What I’m thinking: Even after the latest surge in energy demand, I still believe the solution lies in energy efficiency, renewables, storage, and demand-side flexibility. Nuclear will play a role but can’t be the complete solution. We need to cut carbon emissions today, and most nuclear plants take over ten years from commissioning to completion. We’re already too late to rely solely on nuclear to power our future while keeping global temperatures below catastrophic levels. While technological advances in nuclear could happen in the next 5-10 years, I wouldn’t bet humanity’s future on that actually happening.
Unlocking AI’s potential in the energy sector
Every month, I keep an eye out for AI solutions being deployed in the energy sector:
In the U.S., utilities are beginning to recognize the potential of AI to support operations and enhance resilience. They’re now looking at predicting outages and extreme weather events. However, accessing data from these utilities remains a challenge. Data that could enable demand-side flexibility programs or reduce consumption through digital products is hard to come by. Utilities often make data access difficult because they aren’t interested in lowering demand; in fact, the opposite may be true. I found a great analysis of this phenomenon in this Substack post by the founder of the Mission:data coalition, Michael Murray.
Despite these hurdles, progress is still being made, month after month. Edo and EPRI have earned an award from the California Energy Commission to advance Virtual Power Plants (VPPs) in schools.
Last month I found out about WattTime, an exciting project that provides an API to access marginal operating emissions rates—the emissions rate of electricity generators responding to changes in load on the local grid at a certain time. With last month’s data expansion, they now cover 210 countries.
What I’m thinking: If we want to make our energy system smarter, we need to facilitate data access. This is the only way to achieve cost-effective decarbonization. One of the largest barriers for software and AI companies wanting to work with energy is data accessibility. Regulation is lagging behind, but many bottom-up projects are advancing this from various angles. It’s fantastic that the data provided by WattTime is now available at a global scale. While the market might not be ready for this now (most of our customers at Ento are rather interested in an annual average CO2 emissions factor), it soon will be. Tools like this will form the infrastructure upon which we build the future energy system’s applications.
2. Scientific publications
Reliably estimating the impact of an active Control strategy in a building. This article introduces a new method to measure energy savings from building control strategies. Traditional measurement and verification methods can take up to two years and are often affected by unrelated changes in building performance. This innovative approach uses a randomized switchback design, frequently alternating between baseline and intervention strategies at fixed intervals—like daily switches. By employing sequential testing and predefined stopping criteria, the method provides reliable energy savings estimates much faster than conventional techniques. In a real-world case study in Chicago, the researchers detected an 11% annual energy savings within 45 weeks, with results remaining consistent even after adding more data.
What I’m thinking: At Ento, we also struggle to accurately quantify savings from applying our control strategies. Implementing this randomized switchback design might allow us to demonstrate energy savings to our clients much more quickly.
Diffusion Model Predictive Control. Google Deepmind researchers introduced Diffusion Model Predictive Control (D-MPC), an approach that combines diffusion models with Model Predictive Control. By learning multi-step action proposals and dynamics models using diffusion techniques, D-MPC outperforms existing model-based planning methods and rivals state-of-the-art reinforcement learning algorithms on standard benchmarks like D4RL.
What I’m thinking: The ability of D-MPC to quickly adjust to new objectives and changing conditions without retraining aligns perfectly with the challenges we face in optimizing energy usage in real-time. I’m looking forward to seeing how the algorithm performs in real-world buildings.
3. Reimagine Energy publications
Check out my latest article comparing Google’s Solar API with traditional physics-based modeling for solar PV generation:
4. AI in Energy job board
This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.
Energy Data Science Intern at EnergyHub
Data Scientist: Energy System Modeler - PhD at Microsoft
Founding Product Engineer at ElectronX
Senior Data Engineer at Grundfos
PhD position in Control of heating, ventilation, and air conditioning systems of a building at IMT Atlantique
Conclusion
With so much going on in the sector it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.) please reply to this email or reach out to me on LinkedIn and I’ll be happy to consider them for inclusion!