Risk-Informed Decision-Making in Power Systems
Modern power grids are in a time of significant change. Legacy generators, such as coal and nuclear generators, are retiring. New technologies, such as wind generators, solar generators, and distributed energy resources, are injecting significant variability to the grid and introducing new challenges to reliable grid operation. In this dissertation, we seek to develop new approaches to incorporate uncertainty and risk into operational decision-making to support reliable and cost-efficient power grid operation. Four objectives are pursued. (1) We investigate various approaches of probabilistic forecasting for characterizing the uncertainty in and learning the statistical dependence between power grid variables. We proposed a novel computationally efficient and accurate forecasting approach by combining principal component analysis with periodic regression. This approach helps to reduce the computational expense, and employs a meaningful treatment of the periodicity in the data. (2) We develop a probabilistic risk assessment methodology and associated risk metrics to score and communicate the risks associated with a particular generator scheduling decision. We also develop machine learning-based surrogate models to support fast updating of the risk estimates in real-time. We further develop a hazard aware loss function in training machine learning models for risk assessment. (3) We employ global sensitivity analysis to identify the most influential sources of uncertainty in the grid and define a novel dimension reduction technique to ease the computational burden of stochastic generator scheduling optimization algorithms. (4) We implement a risk-informed decision-making framework to assist operators in transparently, consistently, and holistically considering trade-offs between objectives when making generator scheduling decisions. We also propose an optimization-based approach to select the optimal mitigation among a set of candidates if the generator schedule is found to have unacceptable risk levels. These objectives are illustrated by simulating operational decision-making in three synthetic power grids: a 200-bus grid, a 30-bus grid, and a commercial-sized grid. Through this research, we demonstrate that the principal component analysis plus periodic regression blended the best properties of sequence-to-sequence models (slow degradation with longer forecast horizons) and recursive models (computational efficiency and high performance for short-term predictions). We find that sophisticated stochastic optimizations for making generator scheduling decisions lower the risk when compared to the schedules obtained from existing deterministic optimization approaches. However, they achieve this by increasing online generator capacity; this additional capacity is allocated even if the schedule from the deterministic optimization is already low risk. This could lead to unnecessary increases in cost and emissions. We show that dimension reduction is a promising tool to support the adoption of these stochastic optimizations by lessening their computational cost. By only considering uncertainty in significant variables, we are able to reduce the time required to resolve a stochastic optimization in half without a substantial impact on the risk and cost of the subsequent generator scheduling decision. Finally, we proposed a tool with which operators can select their preferred approach to generator scheduling, while holistically considering risk, cost, emissions, and other factors. We also demonstrate that probabilistic analysis could support making operational decisions in a consistent and transparent manner (as opposed to the ad-hoc, judgment-based approaches currently used).