Customer
This company integrates world-leading process and energy technologies. They provide end-point to cloud integration connecting products, controls, software and services. The goal was to develop models that, based on historical data, would be able to predict the future consumption on different detail levels.
Challenges
- Significant limitation regarding the training data set:
historical training data was limited
the building metadata was of poor quality - Forcast aggregation; For certain buildings the consumption should have been foreseen for different time span. The challenge in this is how to forecast?
- Test data differed from the training data. It made reliable local model validation very difficult.
Solution
- Treating different aggregation methods as separate problems and creating separate dedicated models for these
- Maximal use of training data - one time series was used to create many data points
- Extended attributes that included, among others, selected data regarding power consumption from the past
- A custom model of neural networks that on the one hand significantly reduced the network size and made it easier to train, and at the same time allowed to capture the daily and weekly rhythms in the forecasts.
