Abstract:The power grid is one of the largest man-made critical infrastructures. It has been designed to distribute electric power from generating units to residential, commercial and industrial end-users. Due to the continuous increasing of electrical penetration, the availability and reliability of network is of paramount importance. In addition, the continuous increasing of renewable generators posed a further challenges to the stability of the network due to their dependencies on environmental changes, which are drifting weather scenarios towards extremes. Hence, resilience is becoming a major concern for the future power grid. In order to respond promptly to those important changes, the resilience of the such critical infrastructure has to be augmented. This can only be achieved with the availability of robust computational models that allow to design a better network, robustly validated and updated the results. Ideally, a computational framework for the assessment of power grid resilience should capture all the relevant physical interactions between components, subsystems and the system as a whole. Furthermore, uncertain and heterogeneous environmental factors have to be accounted for and their effect on safety-related metrics explicitly modelled and quantified. This is necessary to reveal power grid risks, hazards and identity situation for which an immediate safety and resilience enhancement is necessary. In this thesis, the existing power grid safety-related concepts (i.e. reliability, risk, vulnerability and resilience) and ancillary uncertainty quantification methods are analysed. The major weakness in existing quantification frameworks has been identified as the way a lack of data required by the frameworks and the treatment of such imprecise information. To overcome this limitation, a novel and robust methods for the uncertainty quantification in power grid safety-critical evaluations has been developed. The main contributions of this dissertation are a set of novel tools for the assessment of power grid reliability, vulnerability and resilience and accounting for a rigorous treatment of lack of data uncertainty. These methods have a limited need for artificial model assumptions, which might alter the quality of the available information and, with it, the validity of safety-critical decisions. One of the key elements for a resilient grid is the system ability to learn from past events, improving the grid structure, operations and policies. For this reason, a Reinforcement Learning framework for optimal decision-making under uncertainty has been investigated. This allows to equip the systems with learning capabilities, which is a fundamental component of the resilience concept, and it optimizes operation and maintenance decisions. The developed frameworks can be used to investigate the effect of threatening scenarios (such as extreme weather conditions, multiple contingencies and cascading events) on the grid safety performance. The validity of the approaches has been tested on scaled-down power grids and prognostic health management as well on realistic models of existing systems (e.g. the IEEE reliability test system). These tools provide a valuable contribution to the research community and industrial practitioners as they can help to discern whether the available information suffices to answer a reliability, vulnerability or resilience related question. If the information is limited and additional data has to be gathered, the method informs the decision-maker with the most relevant and sensitive factors, i.e. a basic indication on where to start collecting data so that an expected reduction in uncertainty is maximised.

Machine Learning for the Uncertainty Quantification of Power Networks

Evaluating Composite Approaches to Modelling High-Dimensional Stochastic Variables in Power Systems

Uncertainty Analysis of Time-domain Simulation for Power System Based on Generalized Polynomial Chaos Method

End-to-End Topology-Aware Machine Learning for Power System Reliability Assessment

Efficient uncertainty propagation for network multiphysics systems

Nonintrusive Uncertainty Quantification of Dynamic Power Systems Subject to Stochastic Excitations

Probabilistic Neural Network to Quantify Uncertainty of Wind Power Estimation

Topological Machine Learning Methods for Power System Responses to Contingencies

PINNs-Based Uncertainty Quantification for Transient Stability Analysis

Neural Network Certification Informed Power System Transient Stability Preventive Control with Renewable Energy

Uncertainty Quantification For Turbulent Flows with Machine Learning

Uncertainty Quantification in LV State Estimation Under High Shares of Flexible Resources

Optimization via Statistical Emulation and Uncertainty Quantification: Hosting Capacity Analysis of Distribution Networks

Evaluating Uncertainties in Electricity Markets via Machine Learning and Quantum Computing

Propagating Parameter Uncertainty in Power System Nonlinear Dynamic Simulations Using a Koopman Operator-Based Surrogate Model

Uncertainty Management in Power System Operation Decision Making

Neural State-Space Models: Empirical Evaluation of Uncertainty Quantification

A Method to Analyze the Voltage Levels of Uncertainty in Distribution Network With Elastic Net

A Data-Driven Linear Robust Optimal Power Flow Model

An Efficient Approach to Power System Uncertainty Analysis With High-Dimensional Dependencies

Robust Computational Frameworks for Power Grid Reliability, Vulnerability and Resilience Analysis