Data-driven prediction and emergency control of transient stability in power systems towards a risk-based optimal power flow operation

Abstract

Cost-efficient and reliable operation of power systems is one of the main concerns of the utilities. The large disturbances and major blackouts occurred in last two decades such as the blackout that took place on 14 August 2003 in the Midwest and Northeast US have ruinous and costly effect for millions of customers. The development of a proper stability prediction and control scheme for an emergency condition is the main objective of this study. In this study, a novel framework using two different approaches is proposed and investigated for real-time transient stability prediction (TSP) in power systems where the signals obtained from PMUs are utilized. The first proposed method is based on signal processing and machine learning approaches which take the computed energy of PMU signals in a window of measurements as an input to a classifier to predict the stability of the system. Several types of classifiers, which are multi-layered perceptrons (MLPs), decision trees (DT), and Naïve Bayes (NB) classifiers, are employed. Two alternative approaches of choosing the window of measurements used for TSP are developed, where an MLP-based fault detection process is also proposed to form the proper window of measurements. One approach is to use a fixed window of only post-fault measurements, whereas the other approach is to use an expanding window of measurements covering pre-fault, fault-on and post-fault stages. Utilization of the energy concept in TSP gives the flexibility to process signals in different sizes while providing predictions that are robust to measurement noises and missing data. It also makes feature selection methods directly applicable, making the TSP possible with fewer PMUs. The proposed methods are applied to two different test systems and a large-scale model of the Turkish power system. In the second approach, a novel methodology based on Koopman mode analysis is proposed to predict the transient stability of a power system in real-time. The method assesses the stability of the system based on a sliding sampling window of PMU measurements, and it detects the evolving instabilities by predicting future samples and investigating the computed Koopman eigenvalues. This approach is also able to identify alarm conditions, which include slowly evolving instabilities that may not be detected by predicting future samples in a limited time horizon. Identifying these conditions provides additional time to prepare a proper set of emergency control actions to be performed when necessary. Using the proposed method, groups of coherent generators that play a role in the evolving instabilities can also be identified, contributing to the design of a defensive islanding scheme for unstable cases. The efficacy of the proposed approach is demonstrated by simulating its performance with three test systems of different scales. Economical operation condition of the power system and its reliability are two contradicting issues. Reliable operation of the power system can lead to a high-cost operation, while economical operation of the power system might result in an unreliable operation of the power system. In this thesis, a novel methodology for the optimal power flow in a power system is proposed to ensure its reliable and cost-effective operation. The methodology adopts a risk-constrained optimal power flow and develops an efficient procedure to design corrective control actions including load shedding and mechanical torque reduction of generators in emergency conditions using reinforcement learning (RL). Reinforcement learning is a type of decision making tool which enables us to determine a set of proper control actions for different operating conditions and contingencies and to implement them in real-time. Since the training process of the RL-based agent is excessively time-consuming for large power systems, because of the enormity of their actions' spaces, an approach based on dynamic mode decomposition which limits the action space during the training process of agent is proposed. The proposed scheme is implemented on two test systems including a small-sized two-area power system and the 127-bus WSCC test system. A considerable amount of operating costs of the power systems corresponds to the fuel cost of the generation units. Therefore, fuel-cost minimization of the power system plays a crucial role in the economic operation of the power system. Furthermore, various faults and contingencies on the power systems might cause irrecoverable results such as widespread blackouts and following loss of money. Considering both fuel cost and reliability level of the system, it can be concluded that it is crucial to provide an optimal power flow solution with acceptable reliability for a given loading condition. Accordingly, the risk level of the system's operating points should be investigated properly. In this study, instead of rotor angle trajectory-based severity indices, the cost of the emergency control action is taken as a severity of the contingency. Using the cost of emergency control actions provided by the trained reinforcement learning-based agent as risk of the operating point, a risk-based optimization problem has been formulated. Two optimization techniques are employed to find the solution of the formulated optimization problem. The first one is Genetic Algorithm, GA, which is one of the well-known populated-based optimization techniques and the second one is Hooke–Jeeves method which is one of the well-known examples of pattern search local approaches. In these algorithms, the candidate solutions are evaluated with both cost function and constraints. The optimum operating points with and without risk constraints has been obtained for the two area and 127-bus test systems using both Genetic algorithm and Hooke-Jeeves method and the results are discussed.