Improving Building Operations – Using Machine Learning to Facilitate Monitoring of Building Portfolio

Abstract

Climatic events are intensifying due to anthropogenic emissions, with buildings producing a significant portion of greenhouse gases. New constructions are designed to consume energy in an efficient way and according to their specifications, but uncertainties in their operation during their entire lifecycle lead to significant deviations. Manually detecting potential issues is costly and requires significant human capital from building operators, especially in the case of diverse portfolios.

This thesis aims at leveraging hourly consumption data generated from energy meters of a portfolio of buildings to derive insights about their operation. Raw data is cleaned and extracted from Siemens’ Navigator platform. Features related to statistics and patterns for electricity, heating and water are extracted, aggregated and normalized for all buildings. A selection of these features is made based on their influence on energy consumption. These features are then used to cluster buildings based on how similarly they operate.

Weekly energy diversity ratios, along with daily and longer-term consistency in electricity, have the largest influence on energy consumption. Clusters include a mix of residential, retail and office buildings, which suggests that the category assigned to a building does not capture its unique operational patterns. Some clusters contain primarily one type, in which case typical aspects of behavior are distinguished. The change of clusters over the years indicates a shift in the building’s behavior or potential malfunction in its systems. Outliers are buildings that operate very differently from all others.

The information derived from the grouping of buildings can be used to facilitate the energy commissioning process by tailoring buildings to a cluster with similar operational behavior. Buildings with mixed activities are more effectively labelled by being matched with peers that behave similarly. Building operators are alerted about buildings with potentially problematic behavior. Knowledge from typical operation patterns can inform the design of new buildings of specific types. Insights derived from the clustering helps researchers more easily explain and quantify the performance deviations between the planning and operational phase of buildings.