< Back to previous page
Examining and Improving Accuracy in a Deep Learning-based Pipeline for the Prediction of Building Energy Demand
Book Contribution - Book Chapter Conference Contribution
It requires energy prediction of several hundreds of building design combinations to explore the design space and evaluate the effect of design parameters which lead to the development of quick energy prediction tool. In the past, researchers have suggested the use of data-driven approaches such as machine learning (ML) to make quick energy prediction. However, the challenge is to develop an ML model which is generalisable enough to accommodate the variations of the early stage, such as different shapes. Component-based ML (CBML) is a solution which we are developing further in this research. This ML model uses several ML components composed in a system entering way to predict building energy demand. Hence, accuracy relies on the performance of each component. In this paper, we are using two ML technique to understand the accuracy of CBML model. First, ceiling analysis serves to examine the improvement in accuracy by examining and exchanging internal model structure. Second, the feature importance exercise delivers information on the importance of a feature in prediction accuracy. By these methods, we identify ML components which have a higher potential to improve the accuracy of overall predictions. The inclusion of new features is examined as one way of prediction improvement. The test case applies CBML on 300 random shapes. In the test case, infiltration, heating energy, and energy demand components have been identified to improve the accuracy more than other ML components. We introduced additional features to improve the accuracy of these models. The accuracy has increased from 0.971 to 0.981 after including new features.
Book: EG-ICE 2020 Workshop on Intelligent Computing in Engineering
Pages: 62 - 72