AI4Buildings - Artificial Intelligence for Digital Planning of Buildings

Given that the building stock represents the largest potential for energy savings and the integration of renewable energy sources, two factors have significant leverage when it comes to the energy performance of buildings:

1. Renewable energy and energy efficiency measures are subsidised in Austria. The amount of subsidies is a decisive factor when it comes to implementing these systems and measures, but requires a thorough understanding of the subsidies available for the specific location, building type and other parameters. Our project aims to develop a tool that makes it easier for future building owners to search for subsidies. This requires an AI-based information retrieval system which automatically extracts all subsidies available for the planned project, together with other relevant parameters (funding amount, deadlines, conditions etc.) in real time, via a web-based interface.
2. BIM (building information modeling) is the method of choice in digital planning. The established open standard for BIM is IFC (Industry Foundation Classes). The quality of IFC-based BIM models, however, varies significantly due to different processes. Definitions of important elements such as IFC spaces or relationships between elements (e.g., to which storey a wall belongs) are often lacking, and some relationships are not defined in IFC at all (e.g., which room is directly accessible from another). This information is, however, important for automatically verifying specific parameters, such as the prescribed length of escape routes (OIB Fire Safety Regulations). Our aim is to improve the quality of IFC-BIM models using geometric algorithms and machine learning approaches.

The AI4Buildings exploratory project provides the basis for developing an application for a follow-up project, for which we are constantly looking for potential project partnerswerden.

First results

We have developed a workflow based on classical geometric algorithms which extracts wall elements from IFC files and calculates the spaces from the geometric parameters. Our prototype is able to write the recognised spaces back to the IFC file, thus enhancing the IFC model with additional information.

We have developed a graphical user interface (GUI) to demonstrate the program prototype, see Figure.

To solve the information retrieval problem concerning funding options, we have written a Python script for importing relevant websites of the individual federal provinces into a local database. We initially focused on the document relevance calculation. The classical TF-IDF algorithm was implemented as a reference and used for testing several vector embedding techniques.

These methods map texts extracted from websites to a vector space and subsequently calculate the text similarities of algebraic vectors (using Euclidean distance or cosine similarity) in this vector space.

Two neural NLP methods were analysed for vector space mapping (embedding): 

1. Fasttext and 
2. Google's BERT (Bidirectional Encoder Representations from Transformers). 

To enable a fast search for similar vectors, we save the embedding vectors to a Milvus database developed specifically for the scalable storage of embedding vectors and fast (nearest neighbour) search functionalities.

Consortium lead

AIT - Austrian Institute of Technology GmbH

AIT - Austrian Institute of Technology GmbH
Dr. Techn. Adam Buruzs
Giefinggasse 4
A-1210 Vienna