Adaptive learning-based time series prediction framework for building energy management

Sustainable building energy management is inevitable in order to reduce global energy demand. For this purpose, building energy management systems need to know the expected behavior of building automation systems, energy production units, or thermal dynamics. Designing the underlying models by domain experts might be a complex and expensive task. However, the models are already inherent in the growing amount of available monitoring data. Thus, this work proposes a framework utilizing learning-based modeling for the prediction of relevant time series in order to support comfort satisfaction and resource efficiency in building energy management. A set of neural networks is generated and trained using monitoring data and building context information modeled in an ontology. Autonomous and building-independent application is achieved by continuous performance evaluation and conditional adaption of the neural networks. The evaluation presents exemplary results and discusses the major findings.