Clean-image backdoor attacks, which use only label manipulation in training datasets to compromise deep neural networks, pose a significant threat to security-critical applications. A critical flaw in existing methods is that the poison rate required for a successful attack induces a proportional, and thus noticeable, drop in Clean Accuracy (CA), undermining their stealthiness. This paper presents a new paradigm for clean-image attacks that minimizes this accuracy degradation by optimizing the trigger itself. We introduce Generative Clean-Image Backdoors (GCB), a framework that uses a conditional InfoGAN to identify naturally occurring image features that can serve as potent and stealthy triggers. By ensuring these triggers are easily separable from benign task-related features, GCB enables a victim model to learn the backdoor from an extremely small set of poisoned examples, resulting in a CA drop of less than 1%. Our experiments demonstrate GCB's remarkable versatility, successfully adapting to six datasets, five architectures, and four tasks, including the first demonstration of clean-image backdoors in regression and segmentation. GCB also exhibits resilience against most of the existing backdoor defenses.

In the global energy crisis, optimal power flow (OPF) has attracted increasing attention as a tool for power system analysis and energy conservation. However, due to the non-convex and nonlinear nature of the OPF problem, it is often challenging to solve. This paper proposes a data-driven approach based on Convolutional Neural Networks (CNN) and transformers for predicting OPF solutions. Traditional studies have typically focused on either complete topological structures or modifications within small bus system test sets. In contrast, the proposed method applies topology labels as inputs and employs one-hot encoding to represent the topology change, allowing for effective computation across networks with varying topological structures. Experiments were conducted using the IEEE 30-bus system and 118-bus system, incorporating N-1 topology changes within the bus system. Compared to recent studies, the errors for PGand VGare reduced by 57.73% and 45.12% on the IEEE 30-bus system, and by 36.26% and 38.41% on the IEEE 118-bus system, respectively.

The growing demand for energy storage systems has highlighted the importance of accurate battery State-of-Charge (SOC) estimation, which directly affects energy availability and operational safety. Existing deep learning-based methods focus on single-cycle charge or discharge estimation, and their performance degrades when applied to continuous multi-cycle scenarios. To address this, a physics-gated neural network, named SocGate, is proposed for multi-cycle battery SOC estimation. The model incorporates physical knowledge through specially designed gate functions, which constrain the estimation within realistic physical boundaries while retaining the flexibility of data-driven modeling. Experiments conducted on a public lithium-ion battery dataset demonstrate that SocGate effectively estimates SOC across sequences involving three discharge cycles, two charge cycles, and two plateau periods. The model achieves a low mean absolute error of 0.46% and root mean squared error of 0.70%.

As climate change intensifies, the electricity industry is crucial for achieving carbon neutrality, but carbon price volatility makes prediction challenging. Traditional statistical and econometric models struggle to capture the nonlinear features of carbon prices effectively. In recent years, deep learning methods, particularly Graph Neural Networks (GNNs), have made significant progress in modeling spatial and temporal dependencies. However, traditional GNNs face limitations in handling higher-order relationships and temporal dependencies among multiple entities. This paper proposes CPP-GNN, a carbon price prediction model based on Temporal Hypergraph Graph Neural Networks. The core innovation lies in using a hypergraph structure to represent complex relationships and employing temporal GNNs to capture the dynamic evolution of carbon prices. Experimental results demonstrate that CPP-GNN performs excellently across multiple carbon emission trading datasets from different regions, significantly outperforming the best benchmark models. CPP-GNN consistently outperforms baselines, achieving an average improvement of 9.3% in mean absolute error, 7.5% in root mean squared error, and 2.6% in directional accuracy. Notably, the model achieves the largest gains in highly dynamic markets such as GDEA and SHEA, highlighting its robustness and adaptability. Therefore, decisionmaking in the electricity market can be further supported by CPP-GNN.

Accurate modeling of structured spatio-temporal signals remains challenging due to complex spatial, temporal, and frequency-domain dependencies. Existing graph neural networks (GNNs) typically neglect frequency-domain anomalies and multi-scale temporal dynamics. To address these gaps, we propose STSSRI-GNN, a Spatio-Temporal Graph Neural Network integrating multi-scale temporal convolution, spectral-aware encoding, and dynamic message passing. Experiments on smart grid, industrial anomaly detection, and biomedical datasets show STSSRI-GNN significantly outperforms state-of-the-art models. Specifically, STSSRI-GNN achieves up to 27.2% lower forecasting MAE, 11% higher anomaly detection PR-AUC, and 5.6% improvement in ECG classification F1-score. Ablation studies further validate the critical role of temporal encoding, attention mechanisms, and spectral-aware features, highlighting the model’s robust generalization across diverse structured signal domains.

Accurate and efficient operational risk prediction is critical for modern power systems facing dynamic unit commitment and renewable integration. Existing machine learning and deep learning approaches often ignore grid topology or temporal evolution, limiting their ability to capture fault propagation under evolving system states. We propose TGN-RiBC, a Betweenness-Aware Temporal Graph Neural Network that incorporates temporal attention, temporal betweenness centrality (TBC), and contrastive learning to jointly model temporal dynamics and structural vulnerability. TGN-RiBC uses time-aware message passing and reachability-based supervision to identify propagation-critical nodes in real time. Experiments on three benchmark systems demonstrate that TGN-RiBC reduces prediction error by up to 3 times compared to classical models and achieves 1.5 to 2 times lower error than prior GNNs, while delivering 60 to 117 times faster inference than SCUC solvers. It maintains reliability with load shedding deviation under 3%, remains robust under up to 20% hourly generator changes, and ablation studies confirm that all core components—temporal encoding, centrality-aware attention, and contrastive learning—are essential to its performance.

Accurately predicting the State of Health (SoH) of batteries has become increasingly important with the widespread adoption of lithium-ion batteries in electric vehicles (EVs). However, existing SoH prediction methods, such as equivalent circuit models and machine learning models, often neglect the integration of machine learning with signal processing. Therefore, the prediction accuracy of SoH on certain battery datasets can be further improved. In this paper, KUformer is proposed, integrating Kolmogorov–Arnold Networks (KAN) and Unscented Kalman Filter (UKF) within a Transformer framework. This model is designed for battery SoH data input characterized by long-term complex dependencies and smooth outputs, achieving high prediction accuracy. Experimental results on two battery datasets with different charging protocols demonstrate that KU-former achieves an average Root Mean Square Percentage Error (RMSPE) of 0.183%, representing at least 38% improvement over the current models, such as iTransformer and PINN.