Machine learning for early stage building energy prediction: Increment and enrichment

Collecting data for machine learning (ML) development is a resource-intensive task that necessitates identifying an efficient data collection approach. This study focuses on ML models that provide quick energy results by dramatically reducing computational demand. The generalisation of such models for multiple building shapes is vital to early-stage energy prediction. Therefore, this article examines which approach of collecting new training samples improves generalisation more - increment of samples in a similar data range or enrichment with samples exhibiting novelty in shape. The first training dataset collects samples from a box-shaped building energy model (BEM). Distribution analysis suggests that they fill only a small portion of the design space. Using the same BEM, the increment approach collects samples that fill the same portion. In contrast, using three differently shaped BEMs, the enrichment approach collects samples well-distributed in the design space. The distribution of samples in a training dataset is quantified to assess their potential to improve generalisation. Using the same number of training samples, the enrichment approach fills the design space better than the increment, reducing the generalisation error (root-mean-square-error) by 58%, compared to 38% after the increment. Hence, the article suggests analysing the distribution of existing and prospective samples to identify an efficient data collection approach having a higher potential to improve generalisation. The developed method will be useful to save expensive data collection resources by focussing on a limited number of samples.