Scheduling Against the Grid: Energy-Aware Automation in 2026

What actually changed on the bill

The headline numbers are not subtle. In the United States, commercial customers paid an average of 13.51 cents per kilowatt-hour in April 2026, roughly 4.8 percent more than a year earlier, and analysts tracking the trend through 2026 described electricity inflation running at about double the general rate. The driver is not mysterious. Data centre construction pushed commercial and industrial demand past what existing grid infrastructure comfortably serves, and the Energy Information Administration expects commercial sector electricity sales to grow 3.3 percent in 2026 and a further 2.7 percent in 2027.

The sharpest increases are hiding in the part of the bill nobody reads. PJM, which operates the grid across thirteen states and the District of Columbia, cleared its capacity auction at the 329.17 dollars per megawatt-day price cap for the delivery year that began in June 2026, up about 22 percent on the previous year and the third consecutive auction to finish at a cap. Without the cap negotiated with Pennsylvania, the clearing price would have landed near 389 dollars. A subsequent auction for the 2028 to 2029 delivery year also hit its cap, at 325 dollars. PJM itself estimates the pass-through at 1.5 to 5 percent on some customer bills, depending on the state and the supplier, and that is on top of energy costs that are rising independently.

Europe is living the same volatility from the opposite direction. Record solar output has produced an extraordinary run of negative wholesale prices: Germany logged 573 hours below zero across 2025, and in the first half of 2026 Spain recorded roughly 600 negative hours, passing its entire 2025 total before the end of June, with Portugal at 462 and France at 370. European photovoltaic generation reached 129 terawatt-hours in the second quarter of 2026, close to 20 percent above any previous second quarter. Prices pressed so hard against the technical floor that exchanges moved it from minus 500 to minus 600 euros per megawatt-hour.

Put the two together and the conclusion is the same on both continents. The average price of electricity is becoming less useful as a planning number, because the spread between the cheapest and most expensive hours of the same day is where the money now sits.

Two different problems wearing one bill

Before automating anything, it is worth separating the two charges that commercial customers pay, because they reward completely different behaviour and a workflow built for one can make the other worse.

That last row is the single most common mistake in energy-aware automation. A workflow that dutifully waits for the cheapest window and then fires the compressors, the chargers and the wash cycle together can lower the kilowatt-hour cost while setting a demand peak that costs more than it saved, and a demand peak set once ratchets through the whole month. Staggering is not an optimisation to add later. It is part of the first version.

Which loads can actually move

The useful filter is not how much power something draws. It is whether the load produces something that can be stored. Cold is storable. Heat is storable. Charge is storable. A completed batch is storable. A haircut is not.

• Refrigeration and cold storage. Pre-cooling a chilled or frozen space by a degree or two during a cheap window and coasting through the expensive one is the classic move, and it needs no new hardware beyond the ability to change a setpoint on a schedule.
• Thermal batch processes. Bakery ovens, commercial laundries, paint curing, kilns, sterilisation cycles and industrial dishwashing all have some flexibility about which hour they run, even when the daily total is fixed.
• Fleet and vehicle charging. A depot with several fast chargers can see all of them draw at once at shift change. Staggering sessions by departure time, rather than plugging in and hoping, is where the large demand-charge reductions come from.
• Water pumping and irrigation. Tanks, reservoirs and greenhouse systems are literal batteries with a schedule attached.
• Digital workloads. Nightly exports, video rendering, backups, model fine-tuning, invoice OCR runs and long report jobs have no opinion about which hour they execute in, and in many businesses they are the easiest thing on this list to move because it is one line in a scheduler.
• Compressed air and machining runs. Small manufacturers often have a fixed weekly volume spread across a flexible calendar, which is exactly the situation scheduling is good at.

The companion question is which loads must never move: anything governed by food safety, pharmacy storage rules, clinical requirements, contractual delivery windows, or a person waiting. Those belong in a hard-constraint list that the automation is not allowed to touch, written down before the first workflow is built. If you are not sure how to pick the first candidate, the same triage logic applies as in any other process work, and our guide to what business processes to automate first is a reasonable starting frame.

The stack, in three layers

Energy-aware automation looks intimidating because it borders on building controls, but the architecture is the same shape as any other integration: a signal comes in, a decision gets made, and something acts. None of the three layers requires a specialist platform to begin with.

A first version that only reaches the third row through a human is still a real system. On a dynamic tariff, the following day's rates are typically published in the afternoon, which means a workflow that runs at five o'clock, ranks tomorrow's windows, checks the constraint list, and posts a plan to the operations channel gives a shift supervisor everything they need. Closing the loop to direct equipment control is a second phase, and it should be a second phase, because you want a month of plans that a human sanity-checked before software starts moving compressors on its own.

What the arithmetic looks like on a small site

Take an illustrative example rather than a vendor case study: a food producer with a cold store, two ovens, a wash line and a small delivery fleet, drawing a peak of about 180 kW and billed at 18 dollars per kW per month for demand plus a time-of-use energy rate.

1. Staggering alone. The oven pre-heat, the wash cycle and the first charging session currently overlap for about twenty minutes each morning. Separating them by fifteen minutes each removes roughly 35 kW from the monthly peak, worth about 630 dollars a month before a single kilowatt-hour moves.
2. Pre-cooling. Dropping the cold store setpoint during the cheap overnight window and letting it drift up within its safe band during the afternoon peak moves a meaningful share of compressor runtime out of the most expensive hours.
3. Charging by departure time. Instead of all vans charging on plug-in, the workflow reads tomorrow's route sheet and charges each vehicle in the cheapest window that still finishes before its departure.
4. Batch jobs. The nightly ERP export, the label print run and the backup are moved off the 6pm slot they inherited from an old installer default.

Published ranges for this kind of work sit at 20 to 40 percent off demand charges and 15 to 25 percent off the energy portion of flexible loads. Treat those as the outer envelope rather than a forecast, and note that the first item on the list, staggering, costs nothing but sequencing. That is the part worth stealing even if you do nothing else. For sizing the project honestly against the saving, the framework in what it costs to automate a business process applies here with one addition: the payback is measurable on a meter, which is rarer and more persuasive than most automation business cases.

Turn-up is the new turn-down

The interesting regulatory movement in 2026 is that grid programmes stopped being purely about using less. The fourth iteration of the United Kingdom's Demand Flexibility Service launched on 14 April 2026 with a significant change: it rewards customers for increasing consumption as well as reducing it, and for shifting usage into periods of high renewable generation. That is a direct consequence of the negative-price problem. A grid with too much midday solar does not need everyone to switch off; it needs someone to switch on.

Germany has been building the same capability into law. Section 14a of the Energy Industry Act requires controllable loads such as heat pumps and wallboxes to be able to receive external control signals, and since 1 April 2025 customers with controllable devices have been able to opt into time-variable grid fees, with different charge levels through the day. The metering rollout underpinning it is expected to reach around 80 percent of meters by October 2026 and complete in 2027. Across the bloc, dynamic tariffs are mandated in principle, with national implementation still uneven.

The practical implication for anyone building this is that the decision layer should not hardcode a tariff shape. Store the tariff as data, the constraints as data, and keep the logic generic enough that moving from a static calendar to a dynamic feed, or adding a demand-response event source, is a configuration change rather than a rebuild. Businesses that treated this as a one-off script in 2024 are the ones rewriting it now.

Where these projects go wrong

• No baseline. If you cannot show what the site did before, you cannot prove a saving and you cannot get paid by a flexibility programme. Capture at least a month of interval data first.
• Optimising kilowatt-hours and ignoring kilowatts. Covered above, and still the most expensive error.
• Hardcoded tariffs. Rate structures changed materially in several markets during 2026. A workflow with prices baked into a condition is a workflow that will quietly be wrong.
• Equipment wear. Price-chasing that cycles compressors and motors more often than the manufacturer intended trades an energy saving for a maintenance bill. Minimum run-times are not optional.
• Staff who were never asked. If the plan moves the wash line an hour later and nobody told the team on that line, the plan gets overridden by lunchtime and the project is declared a failure. Deliver the plan to people before you deliver it to machines.
• Treating it as a project rather than a process. Prices, tariffs and production mix all change. Somebody has to own the workflow, which is the same argument covered in whether your automations need ongoing maintenance, and the answer here is unambiguously yes.

A reasonable first ninety days

1. Weeks one to two. Pull twelve months of bills and separate energy from demand and network charges. Identify which of the two is actually growing at your site. Get access to interval meter data, from the supplier portal if not from your own submeters.
2. Weeks three to four. List every load above a few kilowatts, mark each one movable or fixed, and write the hard constraints for the movable ones. This list is the real asset; the code is comparatively trivial.
3. Weeks five to eight. Build the plan generator. It reads the tariff or price feed, sorts tomorrow's windows, assigns each movable load a start time that respects its constraints and staggers it against the others, and posts the result to the operations channel each afternoon. Humans execute it.
4. Weeks nine to twelve. Reconcile plan against meter. Where the plan was followed and the saving appeared, automate the action directly for that load. Where it was not followed, find out why before automating anything.

Nothing in that sequence requires a platform decision on day one, and the ordering is deliberate: the measurement comes before the control, and the humans come before the machines. It is an unglamorous project that produces an unusually honest number at the end of it, which is more than most automation programmes can say.

FAQ

Why are business electricity bills rising even when wholesale energy looks cheap?

Because the growth is in the parts of the bill that are not the commodity. Capacity, network and grid charges are rising faster than the energy component in several markets. PJM, the largest grid operator in the United States, cleared its capacity auction at the 329.17 dollar per megawatt-day cap for the delivery year that began in June 2026, up about 22 percent year on year and the third consecutive auction to hit a cap. Wholesale energy can fall in the same period while the fixed and capacity portions of the bill climb, which is why average commercial rates kept rising through 2026.

What is the difference between a time-of-use rate and a dynamic tariff?

A time-of-use rate is a published calendar. The utility defines peak, shoulder and off-peak windows in advance, often with weekday peaks in the late afternoon and off-peak from roughly 10pm to 8am, and those windows stay the same for months. A dynamic tariff passes through the wholesale market instead. Octopus Agile in the United Kingdom, for example, publishes 48 half-hourly rates for the following day at around 4pm, derived from the EPEX Spot day-ahead auction. Time-of-use automation can be a fixed schedule; dynamic tariff automation has to re-plan every day.

Which loads should a small business try to shift first?

Start with loads that store their output rather than deliver it live. Refrigeration and cold rooms store cold, water heaters store heat, EV depots store charge, and batch processes such as milling, curing, washing, drying and data jobs store finished work. Those can move by hours without a customer noticing. Loads tied to a person standing in front of them, such as a till, a treatment room or a service bay, should be the last thing you touch.

Do I need a battery to benefit from this?

No, and starting with one is usually the expensive route. Scheduling is free capital: moving a load costs nothing but coordination, while a battery is a purchase with a payback period that commercial peak-shaving projects typically measure in years. Sequence it the other way round. Shift what the calendar can shift, measure the demand peak that remains, and only then size storage against the residual peak you could not schedule away.

What does the automation actually need to connect to?

Three things: a price or signal source, a decision layer, and something that can act. The signal comes from a supplier API, a grid operator feed such as the ENTSO-E Transparency Platform in Europe or day-ahead market data from an independent system operator in North America, or a published tariff calendar. The decision layer is an ordinary workflow tool. The action is whichever endpoint your equipment exposes, which in practice means a charger management API, a building or energy management system, an ERP or MES job queue, a cloud scheduler for digital workloads, or a message to the person who starts the machine.

Can this be done without a full energy management system?

For many small and mid-size sites, yes. A dedicated energy management system is worth it when you have fast control loops, a large connected load and a complex tariff. Below that threshold, a scheduling workflow that reads tomorrow's prices, picks windows, writes the plan into your production calendar and tells staff or equipment what to run when captures a large share of the benefit. The cheapest first version is often a daily plan delivered to a human, not a closed control loop.

How do negative prices change the calculation?

They invert it. In markets with heavy midday solar, the profitable move is sometimes to consume more, not less. Spain recorded roughly 600 hours of negative wholesale prices in the first half of 2026 alone, passing its full-year 2025 total before the end of June, and European exchanges lowered the floor from minus 500 to minus 600 euros per megawatt-hour after prices pressed against the old limit. Automation built only to turn things off will miss half the opportunity, which is why grid programmes are starting to reward turn-up as well as turn-down.