Machine learning (ML)-based transient stability assessment (TSA) provides extraordinary accuracy performance while limited by potential misjudgment risks. To address this issue, this letter originally develops a generic scene-dependent credibility evaluation (SCE) framework. The variance upper bound of ML model prediction error is inferred using an improved localized generalization error estimation (ILGEE) method, and the probability density of system stability is furtherly described as a Gaussian distribution incorporating Neumann boundary condition. Then the scene-dependent credibility index (SCI) is ultimately derived and defined as the information entropy implying the uncertainty of TSA results. Case studies verify the validity of the SCE framework and demonstrate the promising 100% accurate TSA performance with critical proposed SCI as 0.93.

With the increasing penetration of new energy sources in power systems, preliminary security verification by transient stability assessment (TSA) under typical operation modes gradually becomes inadequate. Hence this paper introduces an incremental dataset construction method for efficient TSA model update, which aims at recognizing the cases composing out-of-scope region and boundary region where TSA models generate incredible results. Firstly, a composite distance metric integrating value-based and shape-based similarities of transient response is put forward to identify the local sample space. The cases possessing low membership to the space are classified as outliers. Secondly, an improved localized generalization error estimation (ILGEE) algorithm is originally proposed for variance upper bound estimation of worst TSA error, and furtherly the error is modeled as a Gaussian distribution incorporating Neumann boundary condition. The scene-specific credibility index (SSCI) is then defined such that TSA conditions corresponding to high SSCI are categorized as boundary cases. Finally, the TSA model could be fine-tuned with the incredible out-of-scope and boundary cases labeled by time domain simulation. Case study on a simplified provincial power grid verifies TSA accuracy enhancement (97.58% to 98.35%) and credibility improvement (critical SSCI from 0.93 to 0.67) within 36.7% incremental time by the proposed scheme.

This paper proposes a novel multi-task learning approach based on spatial-temporal recurrent imputation network (SRIN) for power system short-term voltage stability (STVS) assessment with incomplete PMU measurements. The state-of-the-art data imputation methods are based on single and separated learning tasks, which lack optimality for fully exploiting the information in available data. They are also facing several challenges in practical applications, e.g., dependence on complete datasets for training, and performance degradation under continuous data missing scenarios. As a significant advantage, the proposed SRIN method jointly optimizes the objective of missing value imputation and stability prediction through a multi-task recurrent network model. In this way, the integrated model can fully learn from any available data in the incomplete historical database, and the performance of both tasks can benefit from knowledge sharing and transferring across tasks. Moreover, the proposed method has superior advantages in handling both spatial and temporal consecutive missing scenarios, where the imputations are derived by an intelligent combination of history-based and feature-based estimations. Numerical simulation results on two test systems show that, under any PMU missing condition, the proposed method can maintain a competitively high STVS assessment accuracy with a much less imputation error. Note to Practitioners-This paper addresses the challenge of incomplete system observations for power system real-time stability assessment. This problem is not unique to power systems but also extends to other sequential prediction problems facing severe data incompleteness. Existing approaches to solve the missing data problem either relay on complete historical data to train an imputation model, which may not always hold true during practical applications, or impute the missing data by simple statistics, which lacks optimality and adaptivity under diverse missing patterns. This paper proposed a novel, integrated approach to solve this problem by jointly optimizing the two tasks together through a new recurrent network model. In this way, the method can fully learn from seriously undermined datasets. Moreover, this method deals with consecutive missing in time and space, by the design of a trainable weighting component. Numerical simulation results on standard power systems shows that the proposed multi-task model improve the performance of both two tasks and have high adaptivity to different data missing scenarios. In the future research, we will try to address the learning efficiency of this approach for application to larger systems and exploring its adaptability in more extreme scenarios.

With numerous renewable generators and energy storage systems integrated into the power grids, the security-constrained DC optimal power flow (DCOPF) is essential for power system operation. For large-scale power grids, traditional CPU-based optimization algorithms (such as the simplex and barrier methods) have saturated in computational efficiency and are inherently difficult to parallelize. To tackle these issues, by incorporating the symmetric Gauss–Seidel (sGS) decomposition, this work develops a GPU-based Halpern Peaceman-Rachford algorithm, termed the sGS-HPR, which enjoys an O(1k) iteration complexity in terms of the KKT residual. Moreover, the closed-form solutions for all subproblems are derived, which only consist of matrix- vector multiplications and vector operations, and thus can be easily parallelized on GPUs. As a consequence, the developed sGS-HPR algorithm enjoys a O(N_L × n/ϵ) non-ergodic computational complexity in terms of floating-point operations for obtaining an ϵ-optimal solution measured by the KKT residual for large-scale DCOPF problems, where n represents the variable dimension, and N_L denotes the number of branches in the power grid. Extensive numerical tests on large-scale power grids, reaching up to the 9241- bus PEGASE system, demonstrate the scalability and superior efficiency of the developed GPU-based sGS-HPR algorithm compared to state-of-the-art methods. Notably, the proposed method achieves a 6× speedup compared with Gurobi for large-scale instances. Additionally, for ultra-large-scale cases, Gurobi throws an “out-of-memory” error, while the proposed sGS-HPR algorithm maintains its computational scalability and efficiency.

To overcome critical limitations in existing methods for battery capacity estimation, including long-term estimation, handling of irregular sampling and missing data, and transferability across different battery types, this study proposes a novel estimation method, termed batteryCDE, by integrating neural controlled differential equations (CDEs) with attention mechanisms. It first constructs a continuous data path using cubic spline interpolation, followed by neural CDEs that generate feature-wise and cycle-wise attention features to capture essential battery characteristics. These attention-weighted features are processed by another neural CDE to model differential relationships, leading to precise capacity estimations. This study also provides a theoretical analysis of the advantages of batteryCDE over conventional discrete models in terms of long-term estimation, handling of irregular sampling and missing data, and transferability. Extensive experiments evaluate batteryCDE's performance, utilizing three scenarios: varying estimation horizons, missing data, and cross-battery transfer. Results show that batteryCDE outperforms traditional models like long short-term memory (LSTM), Transformer, and neural CDE networks. Even with 50% missing data, an estimation horizon of 100 cycles, and application to different batteries, batteryCDE maintains an estimation error below 0.05. Compared to other methods, batteryCDE reduces estimation errors by 6.59% for long-term predictions, 15.09% for handling missing data, and 17.01% for cross-battery transferability.

To enhance straggler resilience in federated learning (FL) systems, a semi-decentralized approach has been recently proposed, enabling collaboration between clients. Unlike the existing semi-decentralized schemes, which adaptively adjust the collaboration weight according to the network topology, this letter proposes a deterministic coded network that leverages wireless diversity for semi-decentralized FL without requiring prior information about the entire network. Furthermore, the theoretical analyses of the outage and the convergence rate of the proposed scheme are provided. Finally, the superiority of our proposed method over benchmark methods is demonstrated through comprehensive simulations.

This work studies gradient coding (GC) in the context of distributed training problems with unreliable communication. We propose cooperative GC (CoGC), a novel gradient-sharing-based GC framework that leverages cooperative communication among clients. This approach eliminates the need for dataset replication, making it communication- and computation-efficient and suitable for federated learning (FL). By employing the standard GC decoding mechanism, CoGC yields strictly binary outcomes: the global model is either recovered exactly or the recovery is meaningless, with no intermediate outcomes. This characteristic ensures the optimality of the training and demonstrates strong resilience to client-to-server communication failures. However, due to the limited flexibility of the recovery outcomes, the decoding mechanism may also result in communication inefficiency and hinder convergence, especially when communication channels among clients are in poor condition. To overcome this limitation and further exploit the potential of GC matrices, we propose a complementary decoding mechanism, termed GC⁺, which leverages information that would otherwise be discarded during GC decoding failures. This approach significantly improves system reliability against unreliable communication, as the full recovery¹ of the global model dominates in GC⁺. To conclude, this work establishes solid theoretical frameworks for both CoGC and GC⁺. We assess the system reliability by outage analyses and convergence analyses for each decoding mechanism, along with a rigorous investigation of how outages affect the structure and performance of GC matrices. Finally, the effectiveness of CoGC and GC⁺ is validated through extensive simulations.

To address the privacy concerns for state-of-the-art cutting-edge centralized machine learning in electric vehicle (EV) charging demand forecasting applications, federated learning (FL) has been employed to transfer training processes from the cloud server to edge devices. Nevertheless, traditional FL still grapples with several challenges in terms of personalization, transferability, feature extraction, and data security. This study proposes a domain-adaptive clustered federated transfer learning (FTL) scheme for EV charging demand forecasting. This scheme combines the principles of transfer learning (TL) with FL by utilizing maximum mean discrepancy to measure the differences between local features and cluster them, weight local model parameters in the global model aggregation, and realize domain adaptation for projecting local data and new data to the trained FL model. A multi-head attention-based transformer is leveraged to construct a forecasting model to focus on the most relevant spatio-temporal features. Under multi-stage differential privacy protections, Laplace noise is injected into the local feature, model update and local model during the clustering, FL and TL processes, respectively. The case study demonstrates that the proposed domain-adaptive clustered FTL outperforms the conventional FTL and FL, local training, and other domain shift handling techniques in predictive accuracy and operational risk.

This letter proposed a hybrid credibility learning method for data-driven Dynamic Security Assessment (DSA), aiming to enhance the trustworthiness of DSA results under out-of-distribution changing system conditions. The proposed method integrates multiple credibility criteria as learning inputs, combining model consistency-reflected through the output distribution of ensemble learners-with novel indicators of data disparities, namely the Degree of Outlier (DOO) and One-sample Maximum Mean Discrepancy (O-MMD). Moreover, the credibility model benefits from training on a specifically designed out-of-distribution dataset which is further reinforced through misclassified DSA instances. This letter also provides a comprehensive comparison of existing credible DSA methods via the proposed trustworthiness evaluation index. Simulation results on the benchmark testing system show that the proposed method can achieve the best trade-off between DSA accuracy and credibility under various extreme system conditions.

In the rapidly advancing field of federated learning (FL), ensuring efficient FL task delegation while incentivizing FL client participation poses significant challenges, especially in wireless networks where FL participants' coverage is limited. Existing Contract Theory-based methods are designed under the assumption that there is only one FL server in the system (i.e., the monopoly market assumption), which in unrealistic in practice. To address this limitation, we propose Fairness-Aware Multi-Server FL task delegation approach (FAMuS), a novel framework based on Contract Theory and Lyapunov optimization to jointly address these intricate issues facing wireless multi-server FL networks (WMSFLN). Within a given WMSFLN, a task requester products multiple FL tasks and delegate them to FL servers which coordinate the training processes. To ensure fair treatment of FL servers, FAMuS establishes virtual queues to track their previous access to FL tasks, updating them in relation to the resulting FL model performance. The objective is to minimize the time-averaged cost in a WMSFLN, while ensuring all queues remain stable. This is particularly challenging given the incomplete information regarding FL clients' participation cost and the unpredictable nature of the WMSFLN state, which depends on the locations of the mobile clients. Extensive experiments comparing FAMuS against five state-of-the-art approaches based on two real-world datasets demonstrate that it achieves 6.91% higher test accuracy, 27.34% lower cost, and 0.63% higher fairness on average than the best-performing baseline.