Early Stage Design Support using Machine Learning and Building Information Modelling

Buildings represent one-third of energy consumption that concerns the global community. During the early design stages, designers have an opportunity to improve the energy performance of buildings efficiently. However, they need information to assess the effect of their design decisions on the energy performance. The doctoral research developed methods to support energy-related design decisions at the early stages using machine learning (ML) and building information modelling (BIM) by providing relevant information.

The flexible representation of design at the early stages is a prime challenge to assess energy performance and the effect of design decisions on it. This challenge is addressed using a probabilistic approach that requires simulating several hundred models. Dynamic energy simulation tools are computationally expensive, prompting the development of quick metamodels using ML approaches. Further, a BIM integrated approach is developed to reuse the existing information and reduce the modelling efforts. 

The building design progresses by developing information through several levels of development (LOD). While focussing on energy efficiency, it is apt to reduce uncertainty in energy predictions through these LODs. Thus, the research identified design information in the order of its potential to cause uncertainty in the energy predictions. It has been found that geometrical parameters cause the maximum uncertainty in the energy predictions, followed by technical specifications such as U-values and window parameters such as window-to-wall ratio. These results form the basis for design information required in a multi-LOD context.

The research further developed metamodels using ML for quick energy predictions, primarily, focussing on evolving building geometry, making it difficult to develop an ML model for early stage energy predictions. It extended the component based ML (CBML) approach and proposed using a convolutional neural network (CNN) approach to develop ML models. An approach of collecting diverse samples is developed to train CBML components and improve their generalisation. A CNN approach is used to capture the building geometry information from an image instead of simple parameters such as relative compactness. The developed model has improved prediction accuracy as the proposed CNN model architecture allows learning the interactions between the building’s geometry and its energy performance.

There exists a prediction gap between ML predictions and dynamic simulations. A small prediction gap is allowed if it does not affect the comparative assessment of designs under the uncertain conditions of early design stages. Through a test case, the research demonstrated that the developed ML models are suitable to perform a comparative assessment of designs quickly. The aptness of ML models to provide quick comparative assessment of designs allows developing a BIM integrated solution and glean relevant information for supporting design interventions. In the final phase, a cloud based service was developed: p-energyanalysis.de. It implements the developed ML models in a graphical user interface for practical applications. The tool provides information for design space exploration, energy analysis of options and their comparative assessment, sensitivity analysis, and tracks progress. Contrary to generic engineering knowledge, the developed tool extracts context-specific energy performance information. This information is more relevant as it captures the effect of changing design scenario on the energy-efficiency.

The thesis facilitates the performance-oriented building design by providing relevant information using ML and BIM. The developed methods focus on informed design making from an energy perspective that can be extended to other performance evaluations. Besides developing ML models for energy prediction, this research integrates these models with BIM to extract useful information for supporting design interventions. The developed holistic approach provides a quick context-specific assessment of design interventions that enables the designer to make informed decisions.