Scaling Smart Cities with Federated Learning

Abstract The building energy sector is a significant contributor to carbon emissions, thereby playing a crucial role in driving global sustainability efforts to achieve the net-zero targets outlined in the Paris Climate Agreement. Precise predictions of building energy performance are imperative for effective planning and investment decisions aimed at enhancing energy efficiency. While data-driven methods, primarily leveraging machine learning techniques, offer promising predictive capabilities, they heavily rely on large datasets for accurate assessments. However, a prevalent challenge arises as energy consultants and agencies often lack expansive datasets, and if they do, they are reluctant to share their data. To overcome these hurdles, the study implements a decentralized, privacy-preserving machine learning approach known as federated learning. This approach was applied to a dataset encompassing over 25,000 residential buildings featuring diverse construction attributes and energy sources. The simulation involved mimicking different energy agencies by segmenting geographic regions. The study compared the prediction performance of federated learning with that of a model accessing the entire dataset and a fully isolated local model. The findings demonstrate that federated learning achieves a 12% improvement in prediction performance compared to the isolated model. This outcome underscores federated learning’s capacity to leverage the full potential of scaling data-driven methodologies, providing a pathway to unlock new business models in both research and practice, while aligning with net-zero aspirations.