Towards Automated Building Metadata Collection
Automatic collection of Building Metadata
Arka Bhattacharya, David Culler
- TWO MAJOR TYPES OF UNCERTAINTY
- VALUE UNCERTAINTY
- Uncertainty in attribute values due to imperfect information
- Modeled by
- distribution and distribution parameters; or
- PDF .
- RELATIONSHIP UNCERTAINTY
- Uncertainty in relationships between physical objects (e.g VAV ‘X’ maybe fed by AirHandler ‘Y’ or AirHandler ‘Z’)
- Sets of relationships in a building should be captured by a m x n probability lattice, which shrinks in size as we have more certainty.
- Perform some algebra when new information is available to update uncertainty
- Each commercial building is unique and has 1000s of sensors in custom layouts.
- Currently Building metadata (HVAC, zone information, sensor locations, etc) information is fractured and do not exist in a standard representation.
- This prevents scalable deployment of building application services
- Automated collection of building metadata from various information sources into a common representation format
- Reduce uncertainty in representation with increase in ingested information.
Uncertainty Reduction through feedback
Automation : Example 1
Automation : Example 2
EMPERICALLY FIGURING OUT ORIENTATION OF ROOMS:
- Return maximum likelihood actuator i for the VAV for the zone j.
- If returned mapping is correct, then we have nailed down the mapping
- If returned mapping is incorrect, then remove all the mappings of Ai -> Zj , normalize the remaining probabilities
- If Zone extent boundaries are perfect , then it will require O(n2) actuations to get the mapping right.
East / South
- To separate North & West from all other rooms:
- Use June evening temp. rise
- To separate West from North:
- Temp. gradient at 5am in June
- To separate East & South from Interior:
- August morning temp. rise
If zone boundary information is imperfect
1. GbXML is weak in representing relationships. Explore other ways of representing building metadata.
2. How to represent incompleteness of data in the representation
3. How useful are PL “learning-by-example” techninques , and where can they be applied ?
Calculate most likely zone and most likely VAV for that zone.
Given the user feedback, update the probabilty matrices using Bayes Rule.
Classification accuracy :
86% for all rooms on 6th and 7th floors
81% for all rooms on 5th, 6th and 7th floors