The building sector is undergoing a deep transformation to contribute to meeting the climate neutrality goals set by policymakers worldwide. This process entails the transition towards smart energy-aware buildings that have lower consumptions and better efficiency performance. Digitalization is a key part of this process. A huge amount of data is currently generated by sensors, smart meters and a multitude of other devices and data sources, and this trend is expected to exponentially increase in the near future. Exploiting these data for different use cases spanning multiple application scenarios is of utmost importance to capture their full value and build smart and innovative building services. In this context, this paper presents a high-level architecture for big data management in the building domain which aims to foster data sharing, interoperability and the seamless integration of advanced services based on data-driven techniques. This work focuses on the functional description of the architecture, underlining the requirements and specifications to be addressed as well as the design principles to be followed. Moreover, a concrete example of the instantiation of such an architecture, based on open source software technologies, is presented and discussed.

The increase in heterogeneous data in the building energy domain creates a difficult challenge for data integration. Schema matching, which maps the raw data from the building energy domain to a generic data model, is the necessary step in data integration and provides a unique representation. Only a small amount of labeled data for schema matching exists and it is time-consuming and labor-intensive to manually label data. This paper applies semantic-similarity methods to the automatic schema-mapping process by combining knowledge from natural language processing, which reduces the manual effort in heterogeneous data integration. The active-learning method is applied to solve the lack-of-labeled-data problem in schema matching. The results of the schema matching with building-energy-domain data show the pre-trained language model provides a massive improvement in the accuracy of schema matching and the active-learning method greatly reduces the amount of labeled data required.

No abstract