An accurate building energy forecasting model is a key component for real-time and advanced control of building energy system and building-to-grid integration. With the fast deployment and advancement of building automation systems, data are collected by hundreds and sometimes thousands of sensors every few minutes in buildings, which provide great potential for data-driven building energy forecasting.
To develop building energy forecasting models from a large number of potential inputs, feature selection is a critical procedure to ensure model accuracy and computation efficiency. Though the theory of feature selection is well developed in statistics and machine learning fields, it is not well studied in the application of building energy modeling.
In this paper, a feature selection framework proposed in an earlier study is examined using a real campus building in Philadelphia. This feature selection framework combines domain knowledge and statistical methods and is developed for short-term data-driven building energy forecasting. In this case study, the feasibilities of using this feature selection framework in developing whole building energy forecasting model and chiller energy forecasting model are studied.
Results show that, for both whole building and chiller energy forecasting applications, the model with systematic feature selection process presents better performance (in terms of cross validation error of forecasted output) than other models including that with conventional inputs and that uses only single feature selection technique.