UC Irvine

With the development of intelligent measurement systems, power grids improved reliability and efficiency according to the vast amount of collected information. Machine learning(ML) techniques are increasingly used in smart grids since they are efficient in dealing with huge amounts of data and extracting valuable information. However, large-scale deployment of ML models relies on how trustworthy the model is. While sole pursuit of overall learning performance may lead to unfair results. Specifically, the model may unintentionally discriminate different subgroups. To mitigate the unfairness, we propose accuracy parity, equal opportunity and predictive equality regularizers, which can be used for different classification tasks in power grids to mitigate the corresponding discrepancy.However, most tasks in power grids cannot be formulated as classification tasks. Instead, more practical tasks like regression and decision-making take precedence. When addressing the fairness of dynamic decision-making problems over a continuous time scale, two different fairness objectives naturally arise: the instantaneous fairness objective that aims to ensure fairness at every time slot and the long-term fairness objective that aims to sustain fairness over a period. Long-term fairness becomes increasingly crucial due to its adaptability and applicability. We formulate the problem as an online optimization problem with the long-term fairness constraint and propose an algorithm to tackle it. The proposed method analytically yields sub-linear dynamic regret and sub-linear accumulated fair violations.