Abstract

In recent years, the massive data collection in buildings has paved the way for the development of accurate data-driven building models (DDBMs) for various applications. However, due to the variation in data volume of different conditions, existing DDBMs may present distinct accuracy for different users/occupants or periods/conditions. Accuracy variation among users or periods may creates unfairness problems (i.e., algorithmic biases created by data-driven models). This thesis explores and tackles this research problem called fairness-aware prediction of DDBMs.
This thesis first presents a comprehensive review of the entire process involved in developing a DDBM and emphasizes the research gap on achieving fairness in DDBMs. As the first research that introduces fairness concepts into the building engineering domain, this thesis summarizes three types of commonly used fairness definitions. Among these concepts, achieving Type I and Type II fairness in DDBMs shows the beneficial for enabling users to do authority management, achieving uniform predictive performance under different periods or situations, and preserving fairness for different users. In addition, this thesis reviews the commonly used fairness improvement methods for data-driven models.
Then, with the aim of improving fairness for DDBMs to have uniform predictive performance under different conditions and letting MPCs in buildings get optimal control signals based on fair prediction, this research proposes fairness improvement methods for both classification problems and regression problems in the building engineering domain and integrates fairness-aware DDBMs into model predictive controllers (MPCs). This work is separated into three tasks: 1) Task A: For classification problems, four kinds of pre-processing methods are proposed to balance the training dataset. 2) Task B: For regression problems, four in-processing methods, which incorporate fairness-related constraints or penalties into the optimization objective function during the training process of data-driven models, are studied. 3) Task C: The fairness improvement methods proposed in Task A and Task B will be integrated into MPCs.
Case studies are conducted to implement the proposed fairness improvement methods to DDBMs for apartments, develop and integrate the fairness-aware DDBMs into MPCs to get the optimal set-point temperature for controlling the electrically heated floor system (EHF, a heating system with energy storage ability) in a bungalow building. The results show that 1) The proposed pre-processing methods could improve the predictive accuracy of minority conditions and increase fairness in terms of the accuracy rate between different conditions. 2) The proposed in-processing methods could achieve user-defined trade-off between accuracy and fairness. The Type II fairness is achieved by increasing the predictive performance similarity between different conditions. 3) Although improving predictive fairness would decrease the overall predictive accuracy, fairness-aware data-driven based MPCs would not decrease the cost saving and peak shifting ability, compared to the traditional MPC without considering fairness.