Manaz Afzal

Energy & Utilities · Facilities

Each building’s load for tomorrow, by the hour, before the plant starts

Next-day energy forecasts for a facilities portfolio, tested on 3,326,535 real metered building-hours against the same-day-last-week profile.

7.3%hourly error for the forecast
12.2%for the same day last week
100%of buildings closer

Executive summary

A next-day forecast cuts hourly error to 7.3% from 12.2% for the same day last week across 400 real metered buildings, and alerts on 4.6% of building-days instead of 10.2%.

7.3%hourly error
12.2%same day last week
4.6%of building-days alerted

Situation

BMS alerts compare today with last week, so they fire on ordinary days.

Approach

One portfolio model forecasting each building’s next day by the hour, tested on 3,326,535 building-hours.

Result

Closer in 100% of buildings; overspend alerts cut by more than half.

The problem

A facilities team running campuses, offices and residences sets each building’s expected load for tomorrow, for HVAC start times, demand charges and the overnight alert that something has been left running. Most building management systems compare today with the same day last week, which drifts with weather, term dates and holidays, so alerts fire on ordinary days and miss real ones.

I measured how much closer a next-day forecast gets to each building’s hourly load than the same-day-last-week profile, and how many false overspend alerts it removes.

Team
Head of facilities, energy manager, building services engineers
Scale
Hundreds of metered buildings across several sites
Systems
Interval meters, BMS such as Siemens Desigo or Johnson Controls Metasys, energy management platforms, weather feeds
Measure
Hourly forecast error; false overspend alerts per building-month; peak forecast error

What I built

A next-day forecast of every building’s hourly electricity, issued at midnight. It reads the building’s own last week and last four weeks, yesterday’s level and peak, the hour, the day of the week, the season, the outdoor temperature and the building’s use type. One model covers the portfolio, scaled to each building’s level, so a new building is forecast once it has a month of readings.

Interval metersHourly electricityper buildingWeatherOutdoor temperaturefor the siteForecastNext day by hourfor every buildingAlertsOverspend andpeak warnings
Alerts fire when the day’s metered use runs more than 20% over the forecast.

Results

I tested on the Building Data Genome 2 project: real hourly electricity meters from 1,578 non-residential and residential buildings across North America and Europe. 400 buildings with near-complete metering were chosen at random, the forecast was fitted on 2016 and tested on 3,326,535 building-hours in 2017.

Next-day forecast7.3%Yesterday’s profile10.4%Average of the last seven days11.7%Average of the same weekday, four weeks11.7%Same day last week12.2%
The next-day forecast misses by 7.3% of metered use each hour, against 12.2% for the same day last week, and beats that profile in 99.5% of buildings. Data: Building Data Genome 2, Miller and others, Scientific Data 2020
Next-day estimateHourly errorDaily peak errorBuilding-days over by 20%
Next-day forecast7.3%5.3%4.6%
Same day last week12.2%7.1%10.2%
Same weekday, four-week average11.7%8.2%
Same day last weekNext-day forecastWarehouse / storage10.0%Entertainment / public assembly10.9%Office9.8%Public services7.8%Lodging / residential6.4%Healthcare6.9%Education5.6%
Every use type gains; warehouses, public services and offices gain most.

Overspend alerts

With the same-day-last-week profile, 10% of building-days ran more than 20% over the expectation and would raise an alert. With the forecast, 4.6% did, so far fewer ordinary days raise an alert.

Data and method

Data
Building Data Genome 2 (CC BY 4.0): cleaned hourly electricity meter readings for 1,578 buildings at 19 sites, 2016 and 2017, with building use, floor area and site weather. Buildings are identified only by pseudonyms in the release and are not named here.
Selection
874 buildings with readings for at least 97% of hours and few zero readings; 400 chosen at random
Forecast
Gradient-boosted regression of each hour’s load relative to the building’s last-week average, on its own lagged profiles, hour, weekday, day of year, outdoor temperature, use type and floor area; fitted on 2016, tested on 2017
Comparisons
Same day last week, yesterday, the last seven days, and the same weekday over four weeks
Measures
Hourly absolute error as a share of metered use; median error on each day’s peak; share of building-days over the expectation by more than 20%

Limitations

How I would take it live

Financial impact

Fewer alerts to chase

The gain is facilities time on alerts that matter. Portfolio size, time per alert and labour rate are assumptions; alert rates are measured.

InputValueSource
Metered buildings150
Building-days alerted, last-week profile10.2%Measured
Building-days alerted, forecast4.6%Measured
Minutes per alert20
Labour rate per hourA$120

Calculation

  1. Alerts avoided a year3,032
  2. Hours saved1,011

A$121,275 of facilities time a year

Energy savings from faults found sooner come on top.

Implementation and rollout

Six weeks from meters to live alerts

  1. Weeks 1 to 2

    Connect

    A year of interval data and BMS calendars.

    Energy manager, Manaz

  2. Week 3

    Fit

    Forecast fitted; thresholds agreed.

    Manaz

  3. Weeks 4 to 6

    Side by side

    Alerts run beside the BMS comparison.

    Building services

  4. After a month

    Adopt

    False alerts and confirmed faults counted.

    Head of facilities

Tools

Interval meters, Siemens Desigo or Metasys BMS, Energy management platform, Python, LightGBM

Team

  • Head of facilities owns the alerts
  • Energy manager thresholds and reporting
  • Building services site checks
  • Manaz forecast and refresh

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
Weather forecast errorForecast weather used live; error tracked.
Meter gapsGaps filled or flagged before scoring.
Calendar changesTerm dates and holidays fed in.
Alert fatigueThresholds reviewed monthly.

Key takeaways

Key takeaways

  • Last week is a poor guide to tomorrow.
  • One model covers the portfolio.
  • Alerts fall by more than half.
  • Weather and calendar drive the gain.