Demand Charges Are Not
the Whole Story

Most facility teams quote their electricity cost as a single blended number: last month's bill divided by last month's kilowatt-hours. A California commercial and industrial bill is not one price but a stack of distinct products, each with its own unit, its own clock, and its own physics, and a savings projection that treats the stack as one number is usually wrong before its first assumption.

Section 01One bill, several products

The tariff is the contract and the bill is its monthly execution. On Pacific Gas and Electric's system, a commercial or industrial account whose maximum demand has exceeded 499 kilowatts for three consecutive months within the most recent twelve is assigned to Schedule B-19, the medium general demand-metered time-of-use rate.¹ Cross 999 kilowatts on the same test and the account moves to Schedule B-20.² Assignment follows measured load history, not preference: the meter decides which contract governs the site. Southern California Edison runs the same logic under different names, with its large general service schedule serving accounts that regularly register demands above 500 kilowatts.³

What these schedules share is the architecture this paper is about. Energy is priced per kilowatt-hour, at rates that differ by season and by hour of consumption. Demand is priced per kilowatt, on the highest measured interval of the month, under several parallel definitions. A customer charge accrues per meter per day regardless of consumption. Around that core sit the adjustments and overlays that most first-pass models never open: a power factor adjustment benchmarked to 85 percent, rate columns that differ by service voltage, standby provisions that attach when part of the load is self-supplied, and event-based programs that trade bill credits for exposure on called days.¹

The practical consequence is that two facilities with identical annual kilowatt-hours can pay materially different bills, and the same intervention can be worth materially different amounts at each. A blended average erases exactly the structure that determines value. That is not an argument that bills are unknowable. It is an argument that they must be read, and that the reading has an order.

Section 02The three families of charges

Strip any California C&I bill to its skeleton and three families of charges remain. The first is energy: dollars per kilowatt-hour, multiplied by metered consumption inside each time-of-use window. Energy dollars respond to two levers, consuming less and consuming at different hours, and the tariff prices the second lever explicitly.

The second family is demand: dollars per kilowatt, applied not to consumption but to the rate of consumption, sampled as short-interval averages and billed on the month's maxima. On PG&E's B-19 and B-20 this family has several members that are set independently: a summer peak-window demand charge, a summer part-peak demand charge, and a maximum demand charge that applies to the highest interval whenever it occurs, in both seasons.¹ SCE draws the same distinction in different vocabulary, separating facilities-related demand charges, which apply year-round and recover the delivery infrastructure sized to a site's peak, from time-related demand charges tied to on-peak windows.³

The third family is fixed: the customer charge, billed per meter per day, indifferent to anything the facility does.¹ It is usually the smallest of the three and the one that quietly caps what any operational measure can reach, because dollars that do not vary with load cannot be removed by varying load.

Which family dominates is a property of the load, not of the tariff. The ratio of average to peak demand, the load factor, decides it. A continuous process running near its peak around the clock accumulates enormous kilowatt-hours against demand lines it was going to pay anyway; energy dominates, and the interesting questions are about the clock. A short-shift operation with sharp morning starts pays for capacity it touches a few hundred hours a year; demand dominates, and the interesting questions are about intervals. Sales material tends to lead with demand charges because they are legible and dramatic. On high-load-factor sites they are frequently the smaller story, which is why the title of this paper is a warning in both directions.

Section 03Demand is measured, not experienced

Demand charges reward mechanical precision, so here are the mechanics as the tariff states them. PG&E averages demand over fifteen-minute intervals, and the maximum demand for a billing month is the highest of those averages.¹ For intermittent or severely fluctuating loads the utility may shorten the interval to five minutes, and resistance welder loads receive their own computation under the utility's rules.¹ The meter is patient: it does not record why an interval spiked, only that it did.

Because the parallel demand lines are set independently, they can be set by different events on different days. The summer peak-window line might be fixed by an ordinary hot Thursday evening; the facility maximum might be fixed the following Tuesday at six in the morning, when two chillers and an air compressor restart together after a maintenance window. The paper bill reports the resulting kilowatts, not the causing timestamps. Interval data reports both, which is why no serious bill analysis starts anywhere else.

Two further mechanics matter. When a billing month straddles a season boundary, the tariff prorates: maximum demands are evaluated separately for the summer and winter portions and weighted by the billing days each represents.¹ And schedule assignment travels with measured history rather than intuition: eligibility tests are written on recorded maxima, so a site whose load profile changes can find its account transferred to a different schedule with a different structure.^1,2

The engineering consequence is unglamorous: staging. Sequencing large motor starts, interlocking noncritical equipment during defined windows, and alarming on interval trajectory are inexpensive relative to any generation or storage asset, and they act on the same billed lines. They are also limited. Staging cannot remove the capacity a continuous process genuinely needs, and a control strategy that depends on perfect discipline eventually meets the one shift supervisor who overrides it in August.

A single fifteen-minute interval can set a month's demand line. Every other interval that month is billed as energy.

Section 04The clock matters as much as the meter

The time-of-use structure is where California bills have changed most in the past decade, and where models built on older intuition fail. On PG&E's B schedules, summer runs June 1 through September 30 and winter covers the remaining eight months. The peak window is 4 p.m. to 9 p.m. every day of the year, weekends and holidays included. Summer adds part-peak shoulders from 2 p.m. to 4 p.m. and 9 p.m. to 11 p.m. Winter adds a super off-peak window from 9 a.m. to 2 p.m. every day in March, April, and May, the hours and months when midday supply on the state's grid is most abundant.¹

SCE prices the same evening hours as on-peak in summer, but on weekdays excluding holidays rather than every day.³ The difference sounds clerical and is not. A seven-day operation faces summer peak pricing on every evening in one territory and only on weekday evenings in the other, so two identical plants in different service territories should never share a savings model. A spreadsheet ported across that boundary without opening the tariff misprices every summer weekend.

The windows encode grid physics. Evening peaks exist because midday is saturated with solar supply and system stress arrives after sunset; spring midday hours are priced at the bottom of the year because that is when surplus runs deepest. Any measure that changes when a site consumes is buying and selling against this clock, and its arithmetic has to be done window by window:

• Scheduling and load shifting act first on energy dollars, moving consumption out of the evening window into cheaper hours, and can reach the peak-window demand line if the discipline holds every single day of the season. They do nothing for the facility maximum if the site's absolute peak sits mid-morning, and they carry a real operating cost: production scheduled around a tariff is production constrained by it.
• Storage is naturally shaped for a defined evening window: charge in cheap hours, discharge across 4 to 9 p.m. Its honest limits are duration and discipline. The window is five hours long every day, one short or missed discharge can reset a month's peak-window line, the charging energy is itself billed somewhere on the clock, and round-trip losses are purchased kilowatt-hours. Whether it touches the facility maximum depends entirely on when that maximum occurs.
• Solar generation removes midday kilowatt-hours. Its strongest energy case is summer daytime; its weakest is precisely the spring super off-peak window, when output is high and the displaced energy is the cheapest of the year. Its contribution inside the evening peak tapers toward zero across the window, which is why it is so often paired with storage, at added capital and control complexity. It reliably compresses energy dollars; on its own it rarely controls demand lines.
• On-site generation, whether combustion equipment or electrochemical, can address energy and demand lines simultaneously if it runs whenever the site would otherwise set its maxima. In exchange it imports a fuel price, a maintenance calendar, an air-permitting question, and the standby provisions discussed below. Its bill arithmetic is the most powerful of the four and the most entangled.

No technology reads the tariff on its own. The tariff decides which lines each measure is even eligible to touch, and the site's interval data decides how much of that eligibility is real.

Section 05The lines almost nobody models

Power factor. Below the energy and demand blocks sits a line most projections skip. PG&E's B-19 and B-20 rates assume an average power factor of 85 percent, computed monthly from the ratio of lagging reactive consumption to real consumption. Each percentage point above the benchmark reduces the total monthly bill by a small per-kilowatt-hour rate; each point below increases it symmetrically.¹ Plants heavy with lightly loaded induction motors sag below the line, and correction equipment is a mature, modest capital item. The honest caveat runs the other way: the adjustment rate is small per point, so the line deserves a magnitude check before anyone buys hardware to chase it. Some readers will discover they already earn a small monthly credit.

Service voltage. Rate columns differ by delivery voltage: secondary, primary, and transmission-level service each carry their own numbers on the same schedule, reflecting who owns the transformation and where losses land.¹ A model built on the secondary column for a primary-service account is wrong in every line at once, and this is among the most common template errors in vendor spreadsheets.

Unbundling, and the bill after a supply switch. A bundled utility bill stacks generation, distribution, and transmission components inside its printed rates, and the tariff publishes the unbundled split.¹ The split governs customers of community choice aggregators and direct access providers, who buy generation from their provider while the utility continues to bill delivery, usually on one consolidated bill.⁴ Those customers also carry the Power Charge Indifference Adjustment, a charge applied so that the costs of resources procured before a customer's departure do not shift onto those who remain.⁴ Two misreadings follow: comparing a provider's generation-only rate against a utility's all-in rate, and assuming a supply switch made the delivery and indifference lines disappear. Neither survives contact with an actual bill.

Standby. Self-supplying part of the load changes the contract itself. PG&E applies standby provisions to B-19 and B-20 customers whose premises are regularly supplied in part by a nonutility source, with monthly reservation charges under its standby schedule.^1,2 SCE goes further and makes its standby schedule mandatory for most customers who would otherwise take large general service and who supply part or all of their requirements from a generating facility.³ These provisions are neither punitive nor decorative: they price the capacity the utility holds ready for the generator's outage days, and exemptions are tariff-specific. A generation project modeled at full retail displacement with no standby line is overstated by construction.

Event overlays. These schedules also carry event-based pricing structures that grant bill credits in exchange for sharply higher charges during called event windows.¹ Bill models love the credits and forget the events. The only defensible treatment is to model both sides against the site's actual ability to respond when called.

Section 06Seven misreadings that inflate projected savings

The individual errors above compound into a small canon of misreadings that appear over and over in savings projections. Each one inflates the baseline or the delta. None survives interval data and a current tariff sheet.

1. The blended-rate fallacy.Dividing the bill by kilowatt-hours and pricing every avoided unit at the average. The average contains demand and fixed dollars the measure may never touch, so the error flatters everything, most of all whatever is being sold.
2. One demand number.Collapsing peak-window, part-peak, and facility maximum charges into a single figure. They are set independently, often by different intervals on different days; a measure can clear one and miss the others entirely.
3. Averages instead of intervals.Modeling on monthly or hourly averages when each line is set by the single worst fifteen-minute average. Averages are flattering; maxima are what gets billed.
4. July, annualized.Extending summer arithmetic across the year. Summer is four months on PG&E's B schedules, winter has its own windows, and spring afternoons are priced at the year's floor. A measure whose value lives from June to September must say so.
5. The disappearing delivery bill.Comparing a generation-only price against an all-in rate, or assuming a supply switch removed delivery charges and the indifference adjustment. Delivery remains with the utility, and the PCIA remains on the bill.
6. The bill after the project.Modeling the tariff as static while the project changes the load. Standby provisions attach to partial self-supply, power factor moves with new equipment, and schedule eligibility follows measured history. The tariff reacts to what gets built.
7. Last year's sheet.Building on stale rate pages. Tariff sheets revise repeatedly; the B-19 rate pages in force as this paper publishes took effect March 1, 2026. A bill model that does not state its sheet date cannot be audited.

Section 07How to read the bill correctly

The discipline that avoids all seven fits in four steps. First, obtain the governing documents for the account as it actually is: the schedule, the elected options, the service voltage, and the current rate sheets, confirmed against the bill header rather than assumed from memory.

Second, reconstruct the current bill from interval data and the tariff, line by line, until the reconstruction matches the utility's arithmetic to within rounding. This step is where misassigned voltage columns, forgotten options, and stale sheets surface. An analyst who cannot reproduce the existing bill has no basis for projecting a different one.

Third, model any proposed measure against the specific lines it touches: which windows, which seasons, which interval maxima, in both directions, including the new charges the measure itself creates. Fourth, label every number as sourced or estimated, carry the tariff sheet date inside the model, and re-run the model when the sheets revise. Estimates are legitimate; unlabeled estimates are how projections inflate.

It is worth naming who benefits from the blended read. Everyone selling something does, because an unstructured baseline makes every product look better against it, and no seller is compensated to point out that the line their equipment addresses may be a modest fraction of the stack. That is not misconduct; it is compensation doing what compensation does. The correct read has no natural salesman, which is the strongest argument for buying it from someone independent of the answer.

The bill is the most honest dataset a site owns. It is measured, it is priced under a public contract, and it is auditable to the interval. Read correctly, it is the starting point of every defensible power decision, including the decision to change nothing yet.