Join 200+ CIOs, CDOs, Directors, and decision-makers as we unpack how AI is truly transforming
our industries—not through buzzwords or headlines, but through hard science.
As an engineering-led consultancy, we go deep into the research and bring it back
to real-world applications we see today in Oil & Gas,
Utilities, Supply Chain and critical industries.
Power Demand Modelling with Physical AI is a special edition, with Dr. Michael Flaxman (Advisor & Research @ SERIOUS AI). Michael is a rare combination of deep academic expertise and real-world engineering leadership. A Fulbright Scholar with dual PhDs from Harvard, and over two decades at the frontier of spatial planning, environmental modeling, and AI-driven analytics. Michael lays out why the future of energy systems demands Physical AI — and why traditional models, from old forecasting tools to language models, can't keep up with the real-world physics of the grid. From climate-adjusted forecasting to real-time DER control, utilities leading the next decade aren't waiting — they're building.
Heads up, this article touches some technical concepts - if you get lost, no worries - we’ve got you covered in the ‘simply put section’
This is Michael:
Electrical demand isn't creeping up — it's detonating.
Grid Strategies’ National Load Growth Report (December 2024) projects a fivefold increase in US demand in just five years.
Drivers:
- AI-fueled data center expansion
- Skyrocketing deployment of distributed renewables
Meanwhile, most utilities are still trying to fight a new war with pre-Internet weapons.
Today, the modeling problem spans every timescale:
- Seconds–Minutes: Real-time system balancing
- Hours–Days: Day-ahead markets and risk hedging
- Years–Decades: Billion-dollar infrastructure bets
And at every scale, reliability is life or death.
The old models can’t keep up.
This table shows exactly that. Energy forecasting is no longer a single problem — it’s many problems across different timescales, each demanding different methods. As the grid becomes faster, more volatile, and more weather-dependent, relying on slow, rigid models (like traditional NWP) isn’t good enough.
Why Traditional AI Fails for Energy Systems
Large Language Models (LLMs) running critical energy infrastructure scares us.
LLMs are probabilistic: they complete patterns based on training, not grounded physical laws. When confronted with unfamiliar situations, they can infer incorrectly or generate plausible-sounding but wrong answers.
Physical AI takes a different approach. It doesn’t guess. It fits real-world measurements — and shows clearly where and how it misses.
Utilities focused on leading the next decade are already embedding Physical AI into real operations — not just exploring it.
Simply Put:
Energy systems need models that fit real-world physics, not just language patterns — and leading utilities are already making the shift.
What Physical AI Actually Means
Forget the marketing decks.
Physical AI boils down to two things that actually matter:
Training directly on observational data:
Real physics, real sensors, real measurements.
Optimizing to real-world physical targets:
Not "plausibility." Hard metrics like:
- MAE (Mean Absolute Error): How wrong are you, precisely?
- IoU (Intersection over Union): How accurately do you detect structures like rooftop solar?
If a model doesn’t do both of these, it’s not Physical AI.
Simply Put:
Physical AI trains directly on real-world measurements and aims for measurable targets — not just "plausible" answers.
Conventional ML vs Physical AI
Legacy predictive models — random forests, gradient boosted trees (GBTs) — have long captured physical relationships effectively.
GBTs stack trees to sequentially correct errors.
They naturally handle non-linear behaviors, sharp thresholds, and complex interactions — all critical for modeling real-world systems.
They’re also transparent: an auditor can walk through every prediction.
By contrast, LLMs are optimized for language plausibility, not exactitude. Even basic facts or calculations can require careful handling ("prompt engineering") to ensure reliability — making them less suitable for direct control of critical systems without significant adaptation.
Simply Put:
Older models like GBTs already captured physical behaviors better than current LLMs — because they focused on real-world logic.
Pushing Further: Physical Foundation Models
Physical AI is leveling up:
Massive geo-foundation models like IBM’s Prithvi and Clay are trained on petabytes of earth observation data.
They recognize:
- Building footprints
- Vegetative cover
- Solar deployments
- Infrastructure risk zones
But if you think you can plug-and-play these models everywhere without tuning, you’re setting yourself up for bad surprises.
Fine-tuning — precise, context-specific fine-tuning — is essential.
Simply Put:
Big "foundation" models for the earth (like Prithvi and Clay) are powerful, but you have to fine-tune them carefully — no plug-and-play shortcuts.
Who’s Already Executing (While Everyone Else Talks)
Some utilities aren't "exploring options."
They're building the next decade right now:
- Climate-Adjusted Forecasting:
Xcel Energy and Duke Energy folded NOAA’s CMIP6 climate models into operations — already prepping for +15% summer load spikes. - Scenario Planning with RL:
Southern Company simulated 2,000+ grid futures using reinforcement learning — and slashed gas dependency by 14%, with no reliability tradeoff. - Real-Time DER Optimization:
Con Edison wired 650,000+ DERs into a real-time RL system — tightening solar curtailment prediction by 40% and running laps around static controls.
Illustrated: Southern Company — RL for Grid Strategy
Southern Company treats grid planning like a living optimization engine — not a static plan.
Initial State: All assets, loads, and links modeled into a dynamic grid map.
Action: Propose system tweaks — reconfigure assets, reroute flows, adjust generation mixes.
Simulate: Forecast system resilience, failure risks, and operational tradeoffs.
Measure:
- Does system stability improve?
- Does fossil dependency drop without reliability loss?
Reward: Reinforce grid strategies that survive the widest range of future scenarios.
Bottom line:
Old IRP models are static snapshots. Southern Company is using reinforcement learning to simulate, adapt, and build grids that can actually withstand the future.
Southern Company is actively deploying Physical AI to plan resilient, low-carbon energy systems at scale.
Illustrated: Con Edison — RL at the Grid Edge
Con Edison’s real-time grid looks more like a neural network than a legacy dispatch center.
- Live Inverter Streams: Hundreds of thousands of DERs feeding telemetry.
- System State: Real-time mapping of volatility and stability.
- Action: Fine-tune inverter behaviors, storage, and curtailment settings.
- Measure:
- Is voltage better?
- Did curtailment drop?
- Reward: Bake the good behaviors deeper into the system.
Bottom line:
The future grid won't be centrally commanded. It will self-organize, self-correct, and self-optimize.
Con Edison are operating real-time Physical AI systems at scale.
Simply Put:
Utilities like Xcel, Duke, Southern Company, and Con Edison aren’t waiting — they’re using real-world AI to rework forecasting, planning, and control today.
Weather Forecasting: The Hidden Monster
Forget "set and forget" assumptions.
Weather variability — wind, solar, load — is now the grid's wild card.
Traditional Numerical Weather Prediction (NWP):
- Slow
- Computationally brutal
- Poor scalability at finer resolutions
This diagram highlights how Energy forecasting depend on a full chain — from weather models to energy models to generation forecasts — with each step introducing potential errors. Improving only the weather forecast isn't enough; real gains come from optimizing the entire system together, using faster, more accurate Physical AI models.
Simply Put:
Fast, accurate weather predictions are now critical for grid operations — and Physical AI models are making old forecasting methods look slow and expensive.
Physical AI Weather Models Are Crushing It
The ECMWF’s AIFS2 system rewrote the playbook:
- 20%+ accuracy gains at 15-day forecasts.
- 1,000x lower compute demand.
- Predictions delivered in seconds.
If your trading desk and dispatch operators aren't already thinking about how to integrate faster, cheaper, more accurate forecasts (via Physical AI for example), you're leaving millions on the table.
Money on the Table
- 1,250 MW wind plant
- Small forecasting error reduction
- = $12M–$40M a year, depending on system dynamics
And the bigger your renewables share, the bigger the payday for every fractional improvement.
Why Physical AI Fits Utilities Like a Glove
- Speak physics, not poetry.
Physical AI models real-world variables — load curves, voltages, temperatures — not next-word guesses. - Explain everything.
Every decision chain is traceable — no "black box" mysticism. - Work with the engineers you already have.
Minimizing the need for extensive retraining and allowing teams to build on their domain expertise.
Simply Put:
Physical AI models the real variables engineers care about — like voltage and load — with full transparency and no black-box mystery.
Conclusion: No More Excuses
Physical AI isn’t a theory.
It’s not a buzzword.
It’s the way serious utilities will survive — and dominate — the next phase of the energy system.
As Duke Energy’s Lynn Good said in March 2025:
"The era of passive load forecasting is over — we’re now architects of demand shaping."
The only real question:
Are you building?
Get in touch
If you’re serious about deploying Physical AI, SERIOUS AI is built to deliver. We bring real-world deployment experience, engineering-grade talent, and a deep understanding of critical infrastructure. Our systems are transparent, resilient, and built to perform — if you’re ready to forecast better, optimize harder, and outbuild the competition, let’s talk.
