The ongoing process of decarbonizing the power system and the frequent occurrence of extreme weather have led to a significant increase in operating uncertainty and greatly reduced the system controllability. An effective risk assessment and early-warning tool will significantly assist system operators in monitoring and controlling power systems. However, existing methods mainly rely on simplified analytical conditional probability models to describe the component failure probability under different operating conditions and are not suitable for low-probability risk assessment. Inspired by the idea of Bayesian statistics, a Bayesian deep learning-based probabilistic risk assessment and early-warning model considering meteorological conditions is proposed in this article. A new Bayesian neural network (BNN) is proposed which efficiently utilizes expert experience and domain knowledge as prior to model the contingency probability. And a hybrid neural network is developed to rationally utilize the multisource heterogeneous data and comprehensively analyze the historical and forecast information. Finally, a novel risk assessment and early-warning model for high-impact, low-probability extreme events is proposed, which can predict the risk for the next n time-steps. The proposed method is tested on a modified IEEE 118-bus system and a real-world provincial grid, and the results verify its effectiveness.
Demand response (DR) is proposed to address the uncertainty of distributed energy resources (DER) and energy storage systems in distribution networks. However, current research neglects the temporal relationships and the noncooperative relationship of demand response participants. A novel bi-level non-cooperative Stackelberg-Nash dynamic game-theoretic framework is proposed in this paper, specifically designed to address the issue of overlooking consumers’ interests and temporal relationships of participant actions in traditional demand response frameworks. Within this innovative framework, an asynchronous multi-agent reinforcement learning algorithm is proposed and named as Stackelberg-Nash multi-agent proximal policy optimization algorithm (SN-MAPPO). SN-MAPPO allows utility company (UC) and consumers to maximize their utilities within the proposed framework, ultimately leading to the convergence of UC and consumers’ strategies to Stackelberg-Nash equilibrium. To evaluate the effectiveness of the proposed framework, simulations are conducted using authentic data, involving one UC and eight power consumers. Simulation results confirm that the proposed demand response mechanism, while safeguarding consumers’ benefits, fosters the integration of distributed energy resource and effectively mitigates load fluctuations.
22 December 2023 in IEEE Transactions on Industrial Informatics
Transient stability assessment (TSA) is of great significance for the security of power systems. The widely studied postfault TSA based on machine learning methods relies on real-time transient response captured by phasor measurement units (PMUs), which faces difficulties when directly applied to large-scale power systems with a tremendous number of signals as inputs. In this article, we propose a complete scheme for recognizing the most discriminative PMU signals for TSA. First, the original PMU measurement trajectories are projected into uniformly distributed low-dimensional space while maintaining the inherent local structure. Then, a probabilistic dueling clustering method enhanced by a corrected Calinski-Harabaz index is proposed. It is able to divide the projected signals into discrete segments, and then the mutual information between signals and transient stability can be computed as the correlation indicator. Afterward, a signal recognition method based on Eigen decomposition of Laplacian matrix in the information domain is proposed to select the most discriminative signals, which aims to search for the global optimum of maximized relevance and minimized redundancy, and a parallel framework is adopted to improve the recognition efficiency. Key steps of the whole signal recognition scheme are strictly demonstrated in a theoretical way, and case studies on an actual power system provided by China Electric Power Research Institute also verify the effectiveness of the selected signals.
The volatility from the high penetration of distributed energy resources (DERs) makes distribution networks more susceptible to voltage violations. Besides, with the increasing coupling of cyber and physical sides in modern power systems, the risk of potential cyber contingencies (CCs) is rising, which can weaken existing voltage regulation methods. Addressing these issues, this paper proposes a novel graph-based deep reinforcement learning (DRL) framework for enhancing the tolerance of voltage regulation to CCs. Firstly, typical CCs including data missing, data noise, and time delay are modeled in a unified manner, based on the cyber-physical architecture of distribution network. The voltage regulation problem is formulated into a Markov decision process (MDP) with a pertinently designed reward function, while the partial observability exhibited in scenarios involving CCs is also described. In the proposed framework, a novel graph feature representation (GFR) algorithm aiming to mitigate the impact of CCs, which fully utilizes the graph information in the cyber-physical distribution network, is developed in detail and embedded into the proximal policy optimization (PPO) algorithm, whose implementation is specified to ensure its feasibility. Case studies on the 33-bus and 141-bus networks prove the effectiveness and tolerance of the proposed method to CCs of different severities.
30 July 2024 in IEEE Transactions on Sustainable Energy
Accurate wind power prediction (WPP) is crucial to the secure and stable operation of large-scale power systems, and data-driven WPP methods have recently been widely studied and applied. However, existing data-driven methods cannot be applied to new wind farms due to the lack of operational data. This paper presents a novel Bayesian deep learning-based adaptive wind farm power prediction (BDL-AWFPP) method, which is the first time to utilize the computational fluid dynamics (CFD) simulation results as the prior of BDL-based method, thus avoiding the problem that data-driven approaches cannot be applied to newly constructed wind farms. Firstly, a CFD-based wind farm numerical simulation database and a wind turbine power curve database are established to construct a multi-source heterogeneous prior dataset. Then, the BDL-AWFPP model is proposed to utilize the multi-source heterogeneous prior dataset, which can be updated adaptively with newly acquired operational data and saved periodically throughout the life cycle. And an auxiliary aging assessment method for wind turbines is also developed according to the periodically-saved models. Finally, a stochastic variational inference (SVI)-based parameter updating algorithm is derived for the proposed BDL-AWFPP model. Case studies on an actual wind farm validate the effectiveness of the proposed method.