Facility energy use rarely follows a tidy bell curve. Buildings breathe with people and processes. Loads rise at 7 a.m. as lights, AHUs, and hot water kick in, then drop after the morning rush, then surge again when production lines start second shift or when restaurants hit six sigma certification dinner service. The result often looks like two humps instead of one. A bimodal chart captures that shape and turns it into a diagnostic tool. When you can see both peaks clearly, you can ask sharper questions, isolate patterns that daily averages bury, and take action with less guesswork.
I learned to trust this visual the hard way on a university campus where the central plant kept missing its expected turndown overnight. Our daily kWh graphs looked fine, even slightly better than plan. Yet the chillers logged an extra 6 to 8 percent runtime every night. When we replotted the data as an hourly histogram with two distinct peaks, one centered near 10 a.m. and another around 7 p.m., the mystery unraveled. Custodial crews had shifted to a later schedule, and a handful of buildings held temperature setbacks too tight in the early evening. Those two changes nudged a second peak into existence. Without a bimodal view, we would have chased phantom equipment faults for weeks.
What a bimodal chart shows that an average does not
Averages are quick. They are also blunt. A facility that averages 1.2 MW across a day might spend four hours at 2.0 MW and six hours near 0.5 MW. That kind of swing dictates how you negotiate demand charges, plan maintenance windows, and set control strategies. A bimodal chart, typically drawn as a distribution of power or energy by time of day, reveals:
- Distinct operating regimes, such as occupied vs unoccupied or production vs sanitation The width and skew of each regime, which suggest stability or volatility of loads
Those regimes matter. A single normal distribution implies one dominant operating state. A bimodal shape, by contrast, hints at two competing states, two schedules, or two drivers like weather and people. If the two peaks sit far apart, your facility behaves like two buildings stitched together. If the peaks drift by season, your controls likely overreact to outdoor air temperature or solar gains. If a third shoulder appears during audit season or holiday hours, you have a temporary behavior worth catching.
In energy management, the humps often line up with simple realities. Offices and classrooms pull hard during late morning then ease into the afternoon. Kitchens spike near lunch and dinner. Data halls hold steady but tilt upward during maintenance windows. Manufacturing lines rise with shift change, drop for lunch, rise again for finishing and sanitation runs. When you plot consumption by hour across weeks or months, those patterns solidify into peaks.
The anatomy of a bimodal chart for facilities
There is no single correct way to draw one. The goal is clarity. The most useful forms come down to two categories: distributions by value and distributions by time.
A value distribution starts with interval data, fifteen minute or hourly reads at the meter. Collect a few weeks to a few months, then plot a histogram of kW values. If the building reliably idles around 200 kW overnight and sits near 800 kW when occupied, you will see two humps around those values. This version ignores time of day, which can be a feature. By removing the clock, you focus on states. You can then layer metadata such as weekday or weekend, OA temperature bins, or occupancy flags to see which observations belong to which peak.
A time distribution fixes the clock front and center. Take the same interval data, group by hour of day, and compute a distribution or a band for each hour across many days. You might display medians with interquartile ranges, or simple means with standard deviation bands. When the building has two distinct active periods, your chart will show two blocks of elevated consumption, often with troughs in the early morning and mid afternoon. For portfolio managers, this approach scales well across sites because it normalizes by behavior rather than equipment specifics.
In both versions, seasoning matters. Two to four weeks may be enough for a stable process plant. For an office building with a lively calendar and weather swings, collect at least eight weeks per season. You are chasing recurring patterns, not one-off events.
When a bimodal chart is the right tool
Not every facility needs one. A hospital with 24/7 clinical operations might hold a broad, single peak that reflects steady base load. A small retail space with tight hours might produce a narrow daytime peak and a trivial nighttime tail that a simple daily profile captures. The chart shines when you suspect:
- Double occupancy periods, like early classes and evening labs on campuses Split production shifts or frequent overtime Aggressive morning warmup or preheat cycles After-hours cleaning or kitchen prep that extends beyond setpoints Secondary systems, like data closets or process chillers, that anchor a high base load
I lean on the chart whenever I inherit a building I do not trust. Two weeks of data can expose a lopsided base load, a swing that deserves better sequencing, or a second peak that rides on human behavior rather than HVAC physics. One manufacturer used the view to confirm that a late sanitation crew forced the boiler plant to wake at 3 a.m. every day to reheat DHW loops, adding 4 to 6 percent to gas consumption. They retimed the loop setback and installed local electric boosters in two wash stations. Gas use dropped enough to pay back in a single winter.
Building the chart step by step
Start with clean interval data. Utility AMI feeds often arrive with gaps and estimated points. Export fifteen minute kW or kWh intervals for at least 30 days. If you only have monthly bills, you cannot construct a useful distribution. Handle data hygiene first: remove nulls, flag estimates, and interpolate only over short gaps. I avoid filling holes longer than an hour unless I have parallel submeter data to anchor the estimate.
Define your context. You can plot all days together, or you can separate weekdays from weekends, or even tag days by weather. A building where cooling is dominant will show two different personalities in July and in January. If you have outside air temperature at hourly resolution, bin the data into cool, mild, and hot buckets. Two to three bins keep the plot readable.
Choose sensible bin widths. For a histogram of kW values, bins that are too narrow carve noise; too wide and you blur the humps together. As a rule of thumb, set the bin width to 3 to 7 percent of peak load. In a 1 MW building, a 30 to 70 kW bin works. For time of day charts, use hourly averages and derive medians with interquartile shading. The medians resist outliers from outages or test runs.
Annotate, do not overdecorate. Mark shift change times, scheduled warmup periods, or control events like economizer enable. If your facility uses demand response, flag those calls. The audience should see cause and effect. Labels beat legends when time is limited in a stakeholder review.
Reading the peaks with a diagnostician’s eye
The first question is always, what do the two peaks represent physically? The answer lives in your building systems and your people.
If the lower peak hovers around base building services - elevators, emergency lighting, IT closets, pumps, and minimal ventilation - compare it to manufacturer nameplate sums or submeter baselines. I often find that the supposed base is inflated by a handful of constant volume fans left in hand mode or by electric reheat coils that fight a conservative nighttime ventilation schedule.
If the higher peak matches occupied hours but sags on Fridays, you likely have a schedule that edges toward weekend modes sooner than intended. If it stays full size on Saturdays, either the building serves weekend events or some equipment ignores the time schedule. I once tracked a stubborn Saturday peak to a gymnasium AHU whose time clock reset each time the building lost power. A bimodal chart spanning two months made the Saturday pattern impossible to dismiss.
Look at the distance between peaks. A tight gap suggests modest modulation between states, possibly good if the process requires stability, possibly wasteful if the building could afford deeper setbacks. A wide gap calls attention to the cost of being occupied. That raises questions about staging, VFD tuning, and sequences like demand-controlled ventilation. If the base sits higher than peers, the first 10 to 15 percent savings often hides in that valley.
Study the shoulder between the humps. A sag with ripples can indicate a rolling start - equipment that wakes in waves. A flat shoulder may mean continuous circulation pumps or fans that blur the boundary between states. When you see the shoulder drift later into the evening over weeks, suspect creeping schedules or temporary overrides that no one canceled.
Seasonality shows up as a moving peak. In hot months, the occupied peak creeps later into the afternoon as cooling loads ride the sun. In cold months, the morning warmup can form its own mini hump around 6 to 8 a.m. If that warmup hump grows on mild days, your heating control sequence may be overly aggressive, firing boilers to hit a tight morning setpoint even when the building would coast on residual heat.
Linking the chart to tariffs and money
Most commercial tariffs penalize high demand. The maximum kW in a billing window sets part of your bill for the entire month, sometimes the entire season. Bimodal behavior often hides the demand setter. The facility might spend most hours at a modest 600 kW, then climb to 1,100 kW for forty minutes when second shift starts and the chiller restarts. A daily line chart shows the spike once or twice, easy to miss. A value-based bimodal chart stacks those heavy observations into a distinct bulge. That bulge is a bill.
Two strategies emerge. You can reduce the height of the higher peak by better sequencing and pre-cooling or pre-heating, or you can broaden the lower peak upward deliberately by load shifting. The right choice depends on process constraints. In a fulfillment center that must move parcels on a tight window, shifting is hard. In a call center with flexible start times, staggering shifts by fifteen minutes can flatten the top by a meaningful margin. Demand charges of $10 to $30 per kW are common. Shaving 150 kW off a monthly peak saves $1,500 to $4,500 without touching total energy.
When you correlate the peaks with weather bins, the tariff conversation becomes more precise. A peak that shows up only on humid afternoons will point to latent loads and dehumidification penalties. Reheat systems that run to wring moisture from air can push electrical and gas peaks together. Controls that enable demand-controlled ventilation with humidity limits can slice that peak without suffocating the space. The bimodal chart, annotated with dew point bands, makes the case to finance in one slide.
Pitfalls and false positives
A bimodal shape can mislead when data sources mix or when the time window masks special events. Watch for:
- Mixed meters. If your main meter includes a tenant space with an independent schedule, your peaks may simply reflect two unrelated businesses. Submeter or separate the data if you can. Commissioning noise. During retrofits, technicians might run equipment in manual. Two months of such activity can mimic a second operating state. Tag commissioning periods and remove them from the distribution before drawing conclusions. Holidays and outages. A week with a holiday can shift the distribution enough to suggest pattern changes. When possible, exclude atypical days or plot them separately. Weather extremes. One heat wave or cold snap can dominate a month. Binning by temperature prevents those events from rewriting the story.
I keep a checklist by my keyboard: confirm meter scope, confirm schedule, check for overrides, check weather bins, align to tariff windows. It feels tedious until it saves you from the expensive decision that fixes the wrong problem.
Practical examples from the field
An airport terminal with three concourses struggled with chronic evening demand spikes that pushed monthly bills out of alignment with traffic growth. A value-based bimodal chart showed two clear peaks, one at 1.8 MW and a second at 2.4 MW, with the second occurring on three to four evenings per week. Time-of-day analysis pinned the spike between 6 and 8 p.m. Operations insisted this matched the departure bank. Yet deeper review of gate schedules showed the bank ended by 6:15. We dug into HVAC logs and found a cleaning protocol that boosted ventilation to 100 percent outside air during overnight sanitation, and the control program began pre-flushing at 6 p.m. to avoid staff complaints. By moving pre-flush to 8:30 p.m. and adjusting the OA fraction to target CO2 values instead of a fixed percentage, the second peak dropped by 300 to 400 kW. The airport did not buy a new chiller, and the chief engineer reclaimed political capital.
A pharmaceutical plant ran two shifts and a sanitation cycle. Gas consumption from the boilers showed a stubborn second hump at dawn, even in summer. We suspected domestic hot water. A time-of-day bimodal view by hourly gas usage revealed a trough at 2 a.m., then a sharp build from 4 to 6 a.m. When we overlaid sanitization timestamps, the overlap was exact. But another pattern emerged: the morning peak bled into occupied hours, suggesting the boilers kept shoulder loads hot. We installed smarter recirculation controls with temperature band widening during the last hour of sanitation, and added heat recovery on one process. The second peak shrank, and the area under the curve fell by roughly 9 percent month over month.
A downtown office tower swore by optimal start, yet a third hump appeared around 1 p.m. on sunny days. The bimodal shape had a small noon shoulder only on the south-facing facades. The building had installed new glazing and believed the solar control film solved heat gain. The data disagreed. Solar gain still lifted zone temperatures past cooling bands, which triggered electric reheat in perimeter zones due to a control quirk during economizer transitions. The fix required sequence changes, not new glass. The perimeter VAV boxes stopped fighting the AHUs, and the noon shoulder vanished.
Designing control strategies around two peaks
The chart is not the answer; it is the lens. Once you see two regimes clearly, align your control logic with that reality. Avoid trying to force the building into a single state that ignores how people and processes actually use it.
For HVAC, consider separate sequences for core and perimeter zones, and for day and evening occupancies. If the evening peak aligns with people returning for events or with late shift change, set ventilation to follow demand, not the clock, with verified CO2 or occupancy sensing. If you fear sensor failure, build fail-safe minimums, but keep them reasonable. Pay attention to reheat. Bimodal peaks often hide reheat penalties that grow when supply air temperature resets lag the load.
For production equipment, ask if warmup can be decoupled from startup. Machines that require preheat may be able to start in a rolling fashion. The shape of the higher peak often reflects everyone flipping the switch at the same time. Offering a small incentive or changing SOP language can nudge behavior.
On the electrical side, storage and staging can reshape the curve. Batteries sized for 30 to 60 minutes of peak shaving can clip the top hump effectively, especially when coupled with predictive control tied to weather and schedule. Thermal storage, like chilled water tanks, can do the same for cooling loads. But do the math with the chart in view. If your second peak lasts three hours, a small battery only shaves a sliver unless you combine it with sequencing changes.
Communicating the findings to executives and operators
Engineers love scatter plots and control sequences. Executives do not. The virtue of a bimodal chart is that it tells a story quickly. Two humps, two states, two bills. Put the annotated chart early in your deck. Label each peak with plain language: base and occupied, or plant idle and production. Add one callout with dollar values tied to demand charges. Then offer two or three targeted actions with expected ranges, not precise single-point estimates, and a rough payback.
Operators should see a different version. Give them the time-of-day distribution with overlays of schedules and setpoints. Mark where overrides held equipment in hand and when optimal start fired. If the second peak arises from their reality, not a misstep, respect that. Ask where sequence tweaks add work or risk. The best fixes survive because they fit the shop’s rhythm.
Edge cases where bimodal analysis breaks down
Not every two-humped chart deserves an intervention. Some facilities truly operate in two strongly defined modes with little room for change. A research lab with round-the-clock experiments and strict air change requirements may show a high base and an even higher occupied state. Chasing the base may threaten safety or compliance. In such cases, focus on equipment efficiency, not schedule.
Another edge case appears in mixed-use buildings with different tenants and legal constraints. The base peak might include a grocery store that must ventilate according to code while the rest of the tower sleeps. Unless the lease allows cost-sharing for central upgrades, you may not be able to alter the peak shape. Submetering and tenant education become the practical path.
Finally, short datasets can lie. A two-week window that includes a seasonal event or maintenance cycle might show an artificial second hump. If the stakes are high, wait for more data or simulate the building with a calibrated model to test hypotheses before committing capital.
Integrating bimodal charts into a continuous program
Treat the chart as a living report, not a one-off study. Fold it into your monthly review packs. Seasonally, refresh the distributions and compare peak positions. After major control changes, watch how the humps move. If a second peak reappears months later, you have either regression in sequences or drift in schedules.
Set alerts that do not cry wolf. Instead of triggering on single high intervals, monitor the rolling count of intervals above a threshold that defines the upper peak. If the count exceeds a learned baseline for that time of year, prompt an operator review. Tie alerts to context - after a power outage, suspend them for a day. Good alerts preserve trust.
When you invest in metering, add context tags early. Store shift schedules, holiday calendars, and weather in the same warehouse as your interval data. The richer the context, the cleaner your peaks and the faster your diagnosis.
A brief technical note on method
While a histogram or time-banded chart suffices for most decisions, those interested in statistics can fit a two-component Gaussian mixture model to the value distribution. The fit yields means and variances for the two regimes and the mixing proportion. In practice, this helps quantify whether a second peak is truly distinct or just a shoulder. If the model collapses to one component, you probably do not have a stable bimodal pattern. I do not present mixture models to non-technical audiences, but I use them behind the scenes to check my visual instincts.
Smoothing methods also matter. Kernel density estimates produce elegant curves, yet they can suggest peaks where bins would not. Use them as a secondary view. If your histogram shows two humps and the kernel shows two, confidence rises. If only the kernel shows two, rethink your bandwidth or gather more data.
Where to start if you are new to this
You do not need an enterprise platform to try it. Export hourly data from your building automation system or utility portal. In a spreadsheet or a basic analytics tool, plot a histogram of kW values for the past 60 days. If two clusters appear, break the data by weekday and weekend. Then slice by temperature if you have it. Add a simple time-of-day line for context. If the peaks persist across these cuts, brief your team and choose one test action, such as tightening night setbacks by 1 degree Celsius or staggering a start sequence by 10 minutes. Measure for two weeks. Replot. Iterate.
The first time you share a clear bimodal chart with a skeptical finance lead and connect one hump to quantifiable cost, the conversation changes. You are no longer asking for budget to “optimize controls.” You are showing a specific peak linked to a process or schedule and offering a small, testable change with a likely range of savings. That is the kind of energy management that earns permission for the next step.
Why this lens endures
Buildings and plants are machines powered by people. People do not move in straight lines. Shifts end late, events run long, seasons bend schedules. The bimodal chart respects that messiness and shapes it into something you can work with. It keeps you honest about base load and clear-eyed about occupied cost. It invites dialogue across engineering, operations, and finance because anyone can see two humps and ask why.
Over years, I have seen the same pattern surface in airports, universities, food processing, office towers, and distribution centers. The details differ. The discipline does not. Gather enough interval data to see the shape, annotate it with the story of your building, and act where the humps give you leverage. Return later, draw it again, and see how your choices changed the landscape. That rhythm, humble and iterative, beats guesswork every time.
And yes, call it what it is. A bimodal chart. The term anchors the practice in a shared vocabulary that crosses software, spreadsheets, and whiteboards. When your team uses it fluently, the conversation gets faster, the fixes get smaller and smarter, and the bills get lighter.