Industry: Energy
Context: Day-ahead forecasting
Work: Predictive modelling

Day-ahead consumption forecasting with delayed data

An energy retailer needed to forecast electricity consumption 24 hours ahead for day-ahead market decisions. The latest consumption readings were not always available when each prediction had to be made, reducing the recency of the model's inputs. We reproduced that delay during testing and built forecasting models for 15-minute and hourly intervals. The approach outperformed the baseline methods used at the time.

Blue and teal field showing a forecast error below 2.5%

Client

Sector: Energy, retail
Situation: Day-ahead trading decisions depended on forecasts produced before the latest consumption readings became available.

Problem

The latest consumption reading was unavailable at prediction time, leaving the models with less recent inputs. To reproduce that delay, each test used only the observations available when the prediction would have been made.

What was built

Models designed for delayed readings

The forecasting process works from the observations available at prediction time, without assuming the latest interval has already arrived.

Two time intervals

The models forecast the next 24 hours in both 15-minute intervals and hourly intervals.

Testing that reproduces the data delay

We shifted the prediction threshold during evaluation so each test used only the observations that would have been available at that time.

Result

The forecasting approach outperformed the baseline methods in tests that reproduced real data delays
Under 2.5% average percentage error in the same delayed-data tests

Have something similar in mind?

Facing a similar challenge? Tell us how your situation differs.

Get in touch