Energy Usage Forecasting for Buildings and Sites

Most energy data answers the wrong question

Smart meters and building management systems have made half-hourly consumption data cheap. The typical result is a dashboard showing that last month cost more than the month before. Everyone nods, blames the weather or the tariff, and nothing changes.

The useful question is not how much did we use. It is how much should we have used, given the weather, the opening hours and the number of people in the building. Answering that needs a forecast, and once you have it, every half hour of real consumption has something to be compared against.

What drives energy use in a building

A good energy forecasting model is built from the handful of things that genuinely move consumption. For most commercial buildings the list is short.

Gradient-boosted tree models handle these inputs well and are easy to explain. For sites with strong daily cycles, a model that also sees the last few days of consumption improves short-term forecasts.

Using the gap between forecast and actual

This is where the savings usually are. When a model trained on normal operation predicts 180 kWh overnight and the meter reads 310, something is on that should not be.

• Heating or air conditioning running outside occupied hours after a BMS schedule was changed
• Base load creeping up over months as equipment is added and never switched off
• A refrigeration unit working harder as a door seal fails
• Simultaneous heating and cooling fighting each other in the same zone
• Weekend consumption that looks like a weekday because a timer was overridden

Take a retailer with 40 shops. Nobody can watch 40 load profiles. A nightly job that ranks sites by unexplained consumption, with a chart of expected versus actual, gives the facilities manager a short list each morning. That turns data into site visits. It is closely related to the approach in our post on anomaly detection for business operations, applied to meters.

Other things the forecast is good for

1. Budgeting. Next year's consumption under typical weather, so the energy budget is not last year's bill plus a guess.
2. Measuring savings. After an LED retrofit or a new heat pump, compare actual use with what the pre-change model predicts for the same weather. This is a standard measurement and verification approach and far more honest than comparing year on year.
3. Purchasing and flexible tariffs. Day-ahead forecasts help with time-of-use tariffs and demand-side response schemes.
4. Solar and battery sizing. Half-hourly demand profiles show how much on-site generation would actually be self-consumed.
5. Reporting. Weather-normalised figures make sustainability reporting more defensible.

Where energy forecasting falls apart

Forecasts trained on history assume the building still works the way it did. Several common situations break that.

• A site changes use, extends its hours or adds a server room
• Meter data has gaps, estimated reads or a meter swap that resets the series
• Occupancy changed permanently, as many offices found after hybrid working settled
• The only data is monthly bills, which are too coarse for anything beyond budgeting

Monthly bills are the most common blocker. You can build a simple degree-day regression from them, and it is worth doing for budgeting, but finding overnight waste needs at least half-hourly data.

A sensible build for a multi-site business

At SpiderHunts we would begin with the data, not the model. Pull half-hourly readings from the supplier or meter operator, clean the gaps, and join weather by postcode. Plotting a typical week per site often reveals the first round of waste on its own.

1. Collect twelve months of half-hourly data per meter
2. Join weather and site calendars, and flag meter swaps and estimated reads
3. Train one model per site, or a shared model with site features if sites are similar
4. Score each night and rank sites by unexplained consumption
5. Send a short morning report to facilities with a chart per flagged site
6. Retrain quarterly and whenever a site changes how it operates

None of this needs an expensive platform. A scheduled Python job, a PostgreSQL database and a simple report are enough for dozens of sites. Our data science services cover the modelling, and if the reporting should trigger work orders automatically, that is automation territory.

What it will not do

A forecast tells you where to look. It does not tell you why, and it does not switch anything off. Someone still has to visit the site, check the BMS and fix the schedule. If there is no one with the time or authority to act, spend the budget on a facilities contractor's hours instead of a model.

Frequently asked questions