AI-Driven Cooling: Keep Comfort High, Peaks Low

PJM’s grid hit its highest summer peak since 2011 this year, and it wasn’t an isolated story. From Delhi to Osaka to Washington, D.C., 2025’s heat and humidity pushed air-conditioning demand into the danger zone—exactly when power systems are already juggling aging infrastructure, tight capacity margins, and the complexity of renewable energy integration.

Here’s the part most energy teams miss: the “cooling problem” isn’t just about selling more efficient air conditioners. It’s about operating the grid differently during extreme heat—with forecasting that understands humidity, controls that reduce waste, and programs that reward the right behavior at the right time. That’s where AI in energy and utilities stops being a buzzword and becomes a practical tool: better demand forecasting, faster grid optimization, and more resilient operations during heat-driven peaks.

RMI’s 2025 update on the Global Cooling Efficiency Accelerator (GCEA) shows meaningful progress on the equipment side—especially around real-world testing and standards that reflect how people actually use AC. The bigger opportunity for utilities and system operators is to pair that next generation of cooling with AI-driven load intelligence, so “more cooling” doesn’t automatically mean “more blackouts.”

Cooling demand is now a reliability problem (not a comfort upgrade)

Extreme heat turns air conditioning into a public health necessity and a grid reliability risk at the same time. When temperatures stay high overnight and humidity rises, AC doesn’t cycle off the way planners expect. Loads stack up across neighborhoods, distribution assets run hotter, and peak windows stretch.

2025 offered a blunt reminder:

• Record demand strained major grids, including PJM (serving roughly 65 million people).
• Prolonged blackouts hit cities such as Kolkata.
• Regional grid faltering showed up across parts of Indonesia and Southeast Asia.

What’s changed is the shape of the problem. Peak is no longer a neat, predictable “late afternoon spike.” In humid climates, a big chunk of energy use comes from latent load—the energy required to remove moisture. When AC systems aren’t designed or controlled to handle humidity efficiently, people compensate by overcooling, driving consumption up even when the equipment is rated “high efficiency.”

“Overcooling to manage humidity drives excessive energy use — even in the most efficient units. Managing humidity and temperature together is key.”

— Bill McQuade, President of ASHRAE

That quote lands because it explains why utilities can’t solve cooling peaks with simple efficiency labels. They need measurement that reflects real usage and operations that respond in real time.

Why testing standards matter to utilities (and how AI makes them usable)

Better standards create better devices, but utilities need better standards because they improve planning accuracy. If ratings don’t reflect humidity performance or part-load behavior, grid planners and program designers end up incentivizing the wrong outcomes.

GCEA’s 2025 work focused on standards that match reality:

• A draft proposal to ISO on load-based testing for updated AC testing methods and metrics.
• Engagement with the ISO 21280 draft standard incorporating a sensible heat ratio requirement (so units are evaluated for both temperature and humidity).
• Consensus-building with international research groups to validate practical load-based testing and approaches for humidity loads.

The grid value of “load-based” cooling metrics

For utilities, load-based metrics are more than a technical detail. They translate into:

• More accurate peak contribution estimates from residential and commercial cooling
• Better program design for demand response and time-of-use pricing
• Improved distribution planning, because planners can model hot-humid stress instead of assuming dry-bulb temperature tells the whole story

Where AI fits: turning better metrics into better forecasts

Standards updates help equipment engineers. AI helps operators. The moment you have performance curves that reflect part-load and latent load behavior, you can train forecasting and optimization systems to answer questions operators actually face:

• How much of tomorrow’s peak is driven by humidity versus temperature?
• Which feeders will see the highest sustained nighttime load?
• If we shift setpoints by 1–2°F across enrolled customers, how much peak do we reduce without humidity discomfort backlash?

Modern load forecasting models can fuse weather forecasts (including dew point), AMI data, building characteristics, and program enrollment to predict cooling demand at substation or feeder levels. That’s not theoretical—it’s the difference between “we might hit a peak” and “we will overload these three feeders between 6–10 p.m., unless we dispatch these resources.”

Real-world AC performance: the lesson is bigger than efficiency

Lab ratings don’t tell you what happens in apartments during monsoon season or heat waves with sticky nights. GCEA’s field-testing work (including a nine-month field test in Palava City, India) showed a clear spread between typical units, high-efficiency units, and super-efficient concepts.

The punchline for utilities: real-world performance can mean major grid relief and customer savings, even when nameplate efficiency looks similar.

Why “super-efficient” often means “better controls,” not just better hardware

Most cooling peaks are operational problems:

• Users set colder temperatures because humidity feels uncomfortable.
• Units short-cycle or run inefficiently at part load.
• Smart features (if present) aren’t tuned for local climate or grid conditions.

That’s why the GCEA’s industry engagement around integrating smart control logic is so important. Controls are where AI can produce immediate value because they’re software-defined and adaptable.

A utility doesn’t have to wait for the entire market to replace hardware. It can help accelerate impact by supporting:

• Connected thermostat programs that target humidity-aware comfort
• Direct load control that pre-cools intelligently before peak events
• OEM partnerships that enable grid-responsive modes in room ACs

The reality? The fastest path to peak reduction is often control strategy + customer experience, not a perfect piece of equipment.

How AI reduces heat-wave risk across the grid stack

AI delivers the most value when it connects planning, operations, and customer programs into one feedback loop. Cooling is the ideal use case because the driver (weather) is predictable, but the response (human behavior + equipment behavior) is messy.

1) Demand forecasting that understands humidity and behavior

A serious cooling forecast model should ingest more than temperature:

• Dew point / relative humidity (for latent load)
• Overnight minimum temperature (for sustained load)
• Solar irradiance and cloud cover (for building heat gain)
• Customer segmentation (income, building type, occupancy schedules)

This is where utilities can move from one-size-fits-all forecasts to feeder-level, customer-class-aware predictions.

Practical output: a day-ahead “cooling risk map” that flags where voltage issues, transformer overheating, or capacity constraints are most likely.

2) Grid optimization: dispatch the cheapest “cooling capacity” first

When cooling spikes, operators typically lean on peakers, imports, or emergency procedures. AI-assisted optimization can prioritize lower-cost, lower-emissions options:

• Dispatch battery storage into the peak window
• Pull demand response from enrolled AC loads
• Shift flexible loads earlier (pre-cool) or later (post-peak)
• Coordinate with distributed energy resources for local relief

Done well, this becomes a repeatable playbook: predict → position → dispatch → verify.

3) Predictive maintenance during heat stress

Heat waves accelerate failures. Transformers, capacitors, underground cable segments, and substation components all operate closer to limits. AI models that combine thermal loading, SCADA signals, outage history, and asset health data can:

• Identify which assets face the highest probability of failure during the next event
• Recommend targeted inspections or switching plans
• Reduce restoration time by pre-positioning crews and materials

This is the unglamorous part of AI in utilities that pays back quickly: fewer failures when customers need cooling most.

The playbook: pairing super-efficient cooling with AI programs

Utilities don’t need to choose between better equipment and better operations. They need both—coordinated. Here’s a practical sequence I’ve found works when organizations want impact within 12–18 months, without waiting for full market transformation.

Step 1: Build a “cooling peak truth” dataset

Start by identifying what you already have, then fill gaps:

• AMI interval data (15-min or hourly)
• Weather reanalysis + forecast feeds (humidity included)
• Feeder and transformer loading data
• Customer program participation (thermostats, rebates, DR)

Output: a single dataset that can explain last summer’s peaks in plain terms.

Step 2: Segment customers by cooling behavior, not just rate class

Two households with the same usage can behave differently under humidity stress. AI clustering can group customers by:

• Peak coincidence
• Sensitivity to dew point
• Nighttime load persistence
• Response to price signals or events

Output: targeted program offers that improve enrollment and performance.

Step 3: Design DR events around comfort (humidity-aware)

If customers hate your events, your portfolio becomes unreliable. Use control logic that:

• Limits humidity drift n- Uses pre-cooling and short setbacks instead of long uncomfortable ones
• Provides “opt-down” options for vulnerable customers

Output: higher event participation and fewer overrides.

Step 4: Align rebates with real-world performance

As standards evolve, utilities can modernize incentives:

• Reward units that maintain comfort at lower wattage under part-load
• Encourage smart controls that manage humidity efficiently
• Support pilots in diverse climates to validate savings and peak impact

Output: rebates that reduce peak, not just annual kWh.

What to watch in 2026: standards, pilots, and market activation

GCEA’s stated 2026 focus—manufacturing capability, scaled pilots, and market activation—maps neatly to what utilities need to de-risk investment.

• Pilots across climates will create proof points utilities can translate into IRP assumptions, program filings, and capacity planning.
• Better test methods will reduce uncertainty in savings estimates, especially for hot-humid regions where latent load dominates.
• Market activation (finance, aggregation, awareness) creates the scale needed for meaningful peak reduction.

For the AI in Energy & Utilities series, this is a clean storyline: as heat extremes intensify, cooling becomes a primary driver of load growth, and AI becomes a practical operating layer that keeps reliability high while emissions fall.

Utilities that treat cooling as “just a summer issue” will keep getting surprised by peaks. Utilities that treat cooling as a data problem—forecastable, controllable, optimizable—will be the ones that ride out the next heat wave without emergency measures.

If you’re planning for summer 2026 now, ask your team one forward-looking question: Do we understand our cooling load well enough to control it—without making customers miserable?