< Back to previous page

Project

Machine Learning for Energy Performance Prediction in Early Design Stage of Buildings

The early building design is an iterative process. In this process, architects and engineers evaluate different design concepts to ensure the design brief is fulfilled. The rising need for a design to adhere to certain performance has introduced building performance simulation (BPS) into the design process. Early design decisions have the highest impact on building performance. Therefore, it is important to make good design decisions at this stage of the design process.

The computational effort, in terms of model development time and prediction time, is high for BPS. The computational effort combined with other factors like the need for detailed design information limits BPS in the early design stages. This research focusses on having computational methods that deliver the predictions as fast as possible. The speed to obtain a prediction is important for iterative early design, as the prediction models should be able to keep up with the thinking speed of a designer. For example, simple BPS takes about 5 minutes to make a prediction on design performance. Five minutes may appear like an insignificant amount of time; This time accumulates as more design needs to be evaluated. Moreover, during the creative process of a designer, his or her speed of iterating over design options is faster than a BPS model’s time to provide performance results. The slow nature of BPS results in limited design options evaluated for building performance and potentially missing design with ideal performance.

Furthermore, computational speed becomes a challenge when designers with different educational backgrounds need to collaborate on a design. The low computational speed of BPS makes designers rely more on rule-of-thumb knowledge. The applied rule-of-thumb knowledge may or may not be valid for the proposed design problem. It is increasing the risk of not taking the right design decisions from a performance point-of-view. To overcome the challenge of computational speed, this research evaluates machine learning (ML) as an alternative method for building performance prediction. Reason for using ML is its high computational speed and prediction accuracy. However, ML models have to overcome challenges like generalization, reusability, and interpretability.

This research evaluates generalization through two different approaches, which are component-based approach and deep learning. Both approaches model the relationship between building design parameters and design performance in hierarchies. Results indicate that ML models do generalize in unseen design cases, provided the evaluated design is similar to the nonlinearity present in the training data distribution. Within the ML algorithms, deep learning model architectures based convolutional neural networks (CNN) outperform traditional neural networks (NN). CNN is able to outperform traditional NN as it could extract features from the data in a hierarchical manner.

The reusability of the ML model is evaluated to reduce the computational effort required in developing multiple ML models. Transfer learning and multi-task learning methods are evaluated to understand ML model reusability. Developing[1] two deep learning models sequentially takes ~22 minutes. This development time accumulates as the number of models to be developed increases. Results indicate that through both transfer learning and multi-task learning computation effort for model development can be reduced without compromising on model accuracy. The development times are reduced to ~14 minutes and ~8 minutes respectively.

The interpretability of deep learning methods is evaluated through dimensionality reduction methods. Results indicate that deep learning models learn to re-organize the design space based on the design’s energy signature. Therefore, the trained model is similar to a top-down approach for predictions, in which, predictions are based on design similarity.

Finally, results show that ML models predict design performance for 201 design options in 0.9 seconds, while the same results can be obtained from BPS in ~20 minutes. Showing that ML models are significantly faster than BPS. Results are giving an indication that ML models could indeed keep-up with the speed of a designer. The thesis elaborates more on the ML methods evaluated and the outcomes of the research. 

[1] Training the ML model. This is not the same as developing BPS during design.

Date:2 Dec 2015 →  21 Feb 2020
Keywords:Machine Learning, Neural Network, Energy-efficient building design
Disciplines:Architectural engineering, Architecture, Interior architecture, Architectural design, Art studies and sciences, Electrical power engineering, Energy generation, conversion and storage engineering
Project type:PhD project