Component-Based Machine Learning Modelling Approach For Design Stage Building Energy Prediction: Weather Conditions And Size

Building energy predictions are playing an important role in steering the design towards the required sustainability regulations. Time-consuming nature of detailed Building Energy Modelling (BEM) has introduced simplified BEM and metamodels within the design process. The paper further elaborates the limitations of this method and proposes a component-based Machine Learning Modelling (MLM) approach which could potentially overcome the current limitations. The paper proposes a methodology for developing component-based MLM that generalise well. Generalisation, in this paper, refers to the reusability of an MLM developed with data from a specific situation in similar circumstances. As a first step in ongoing research on component-based MLM, a model is developed with data from a simple box building with weather data of Amsterdam, Brussels and Paris and two occupancy profiles. It is shown that the MLM is able to predict the annual energy for (1) same box building under different weather conditions not included in the training data (2) different dimensions of the box building for one case weather data and occupancy. The prediction error for annual heating demand is lower than 10% for all evaluated cases while the prediction error for annual cooling demand ranges -3.4% to 28.3%. Good generalisation is observed for all heating energy predictions whereas only for a few cooling energy predictions. Possibilities for model improvement and next steps of the research project are described.status: publishe