Graph Neural Networks (GNNs) have demonstrated remarkable performance in modeling complex relational data. However, their deployment in privacy-sensitive domains remains challenging due to the inherent tension between privacy protection and model interpretability. This paper introduces a novel Multi-Branch Dendritic Computing framework for Federated Graph Neural Networks (MBD-FGNN) that addresses this fundamental challenge. Our approach integrates biologically-inspired dendritic computing with federated learning paradigms, enabling distributed training across decentralized graph data while preserving both privacy and interpretability. The multi-branch architecture mimics dendritic computation in biological neurons,where each branch processes patterns at different receptive field scales (1-hop, 2-hop, and 4-hop neighborhoods), providing natural model explanations through branch contribution analysis.To ensure rigorous privacy protection, we develop a branch-adaptive differential privacy mechanism with gradient clipping and calibrated Gaussian noise injection, satisfying (ϵ, δ)-differential privacy. We theoretically analyze the framework’s ability to maintain interpretation stability under differential privacy constraints and derive formal guarantees on the privacy-interpretability trade-off. Extensive experiments on benchmark graph datasets demonstrate that MBD-FGNN outperforms state-of-the-art federated GNN approaches in terms of accuracy, privacy preservation, and explanation quality. Notably, our method achieves robust explanation stability even under privacy constraints, maintaining higher explanation consistency compared to conventional approaches where explanations significantly degrade as privacy protection increases. The proposed framework opens new avenues for deploying interpretable graph learning systems in privacy-critical applications such as healthcare networks, financial transaction graphs, and social network analysis. The code and dataset are publicly available at https://github.com/ZhaoGuan-nuist/MBD-FGNN.

Traditional deep learning methods have achieved remarkable success by leveraging large-scale labeled datasets. However, in real-world applications, acquiring labeled data is often expensive, and data privacy concerns restrict centralized data collection. Federated Semi-Supervised Learning (FSSL) has emerged to address these limitations by combining Federated Learning (FL) and Semi-Supervised Learning (SSL), enabling the utilization of unlabeled data while preserving data privacy. Despite its potential, FSSL faces two critical challenges: (1) limited supervision and model uncertainty caused by the absence of accurate pseudo-label estimation, and (2) model performance degradation arising from Non-IID data distributions. To address these challenges, this paper proposes FedGDS, a novel FSSL framework that integrates Evidence Theory-based Pseudo-Labeling (ETPL) and Graph-based Neighbor Aggregation (GNA). Specifically, ETPL employs evidence theory to fuse predictions from multiple models, reducing reliance on any single model. It further introduces a client consistency constraint to enhance the stability and accuracy of pseudo-label estimation. In addition, GNA regards IoT client models as nodes in a graph and performs layer-wise aggregation from neighboring nodes to enrich feature representations, thereby alleviating the adverse effects of Non-IID data. Extensive experiments on standard datasets such as CIFAR10 and SVHN, as well as IoT-specific datasets including BloodMNIST and RT-IOT2022, show that the proposed FedGDS framework significantly improves model accuracy and effectively narrows the performance gap between Non-IID and baseline IID settings.

Silent packet loss (SPL) poses a significant challenge for state estimation, due to the lack of information regarding the packet loss status (PLS). This issue is particularly prominent in event-based systems, where the interplay between event-triggered feedback mechanisms and SPL complicates the estimation process. Existing approaches often employ detectors, but achieving 100% detection accuracy remains nearly impossible, and false detections further complicate the probability distribution of system states. In this article, we propose a silent message-passing mechanism (SMPM) to address the SPL of measurements, which may be coupled with an event-based scheduler, as the feedback channel is also affected by the SPL. Besides, a dynamic Chi-square ( $d$ - $\mathcal {X}^{2}$ ) detector is proposed, whose detection accuracy is proven to converge to 100% with time. Subsequently, an estimator based on the $d$ - $\mathcal {X}^{2}$ detector under the SMPM is designed for unstable systems with the SPL of measurements. More importantly, a stability condition is established, revealing the relationship between the SPL rate and estimation performance. Both simulations and experiments validate the effectiveness of the theoretical findings presented in this article.

Accurate and efficient multienergy load forecasting is fundamental for the secure and economic operation of integrated energy systems (IESs). However, current methods face several critical challenges as follows. 1) Insufficient perception of low-energy transient features leads to significant deviations in marginal scenario predictions. 2) Existing architectures struggle to simultaneously capture long-term trends and short-term dynamics efficiently. 3) Many high-accuracy models suffer from high computational complexity, limiting their practical deployment. To overcome these limitations, we propose the frequency-enhanced seasonal-trend recurrent neural network (FESTRNN), which incorporates two newly designed modules: the energy-enhanced spectral module to amplify critical frequency signals, and the dual-stream collaborative temporal forecasting module to efficiently model multiscale temporal dependencies. Experimental results on the ASU dataset demonstrate that FESTRNN achieves state-of-the-art accuracy across various forecasting horizons with significantly lower computational costs. This study provides a reliable basis for real-time scheduling in IES, offering strong practical value.

With the rapid advancement of remote sensing technology, change detection using bitemporal remote sensing images has significant applications in land use planning and environmental monitoring. The emergence of convolutional neural networks (CNNs) has accelerated the development of deep learning-based change detection. However, existing deep learning algorithms exhibit limitations in understanding bitemporal feature relationships and accurately identifying change region boundaries. Moreover, they inadequately explore feature interactions between bitemporal images before extracting differential features. To address these issues, this article proposes a novel interactive and supervised dual-mode attention network (ISDANet). In the feature encoding stage, we employ the lightweight MobileNetV2 as the backbone to extract bitemporal features. Additionally, we design the neighbor feature aggregation module (NFAM) to aggregate semantic features from adjacent scales within the dual-branch backbone, enhancing the representation of temporal features. We further introduce the interactive attention enhancement module (IAEM), which effectively integrates self-attention and cross-attention mechanisms. This establishes deep interactions between bitemporal features, suppresses irrelevant noise, and ensures precise focus on true change regions. In the feature decoding stage, the supervised attention module (SAM) reweights differential features and leverages supervisory signals to guide the learning of attention mechanisms, significantly improving boundary detection accuracy. SAM dynamically aggregates multilevel features, balancing high-level semantics and low-level details to capture subtle changes in complex scenes. The proposed model achieves F1 scores that are 0.28%, 1.6%, and 0.76% higher than the best comparative method, spatiotemporal enhancement and interlevel fusion network (SEIFNet), on three CD datasets [LEVIR-CD, Guangzhou dataset (GZ-CD), and Sun Yat-sen University dataset (SYSU-CD)], respectively, while maintaining a lightweight design with only 6.93 M parameters and 3.46G floating-point operations (FLOPs). The code is available athttps://github.com/RenHongjin6/ISDANet.

Deep learning has significantly reshaped the landscape of remote sensing change detection (RSCD). However, detecting subtle changes remains a formidable challenge due to the simultaneous presence of three critical issues: domain shift from seasonal and illumination variations, severe class imbalance between change and background areas, and the need for high deployment efficiency. To address these multifaceted problems in a unified manner, this paper introduces SDTNet, a comprehensive framework featuring three key innovations. First, to counter domain shift, we propose the Style-Unified Temporal Difference Contrastive Learning Strategy (STDCL), a novel, fine-grained contrastive learning method that guides the model to decouple real from pseudo changes without incurring additional inference costs. Second, to mitigate severe class imbalance, we design the Adaptive Dice Focal Loss (ADFLoss), which, unlike static loss functions, introduces two novel adaptive factors to dynamically balance sample weights and implement a curriculum learning strategy. Third, for efficient feature enhancement, we present the Dual-Branch Temporal Difference Transformer (DTDFormer), an efficiency-aware architecture that adapts the differential attention mechanism for bi-temporal inputs and incorporates a sparse attention module to reduce computational complexity. Experimental results demonstrate that our integrated SDTNet framework achieves state-of-the-art (SOTA) performance across five large-scale change detection datasets. The code is publicly available at https://github.com/Baobaon/SDTNet.

Unit commitment (UC) is a critical component for the power system dispatching departments. Current methodologies for solving UC problems predominantly rely on mixed-integer linear programming and are supplemented by data-driven approaches. These methodologies have two primary limitations: first, as the scale of the power grid expands, the complexity of algorithms increases sharply. Second, they fail to fully exploit grid topology information and historical data trends. To address these limitations, this article proposes a constrained graph-temporal convolutional network, which addresses the UC problem by directly predicting power output with constraints. The algorithm takes historical load data from grid nodes as input, utilizes graph convolutional networks to capture the physical grid topology information, and employs temporal convolutional networks to extract temporal features. Ultimately, the output of the graph-temporal convolutional network is projected into the feasible domain through a linear constraint activation layer, achieving accurate power prediction. Experiments conducted on the IEEE 30-BUS and IEEE 118-BUS systems validate the feasibility and superiority of our method in terms of accuracy and computational efficiency.

In the context of vertical federated learning (VFL), agents utilize multimodal data on their edge devices to corporately train and inference with the deep learning models. However, in classical VFL, there exists three problems from the perspective of embeddings. 1) the utilization of oversimplified embedding fusion mechanism may result in suboptimal performance of the models; 2) the exchange of embeddings and their gradients poses a potential risk of private information leakage, as they inherently contain sensitive information; 3) finally, the withdrawal of some agents from cooperation disrupts the collaborative inference capabilities of the remaining agents. To mitigate these problems, this article introduces a novel VFL algorithm grounded in embedding alignment. It includes two distinct schemes: 1) performance-oriented scheme (POS) and 2) privacy-respecting scheme (PRS). Within POS, this article employs contrastive loss and joint fine-tuning to augment the expressiveness and the overall performance of models. While the PRS incorporates homomorphic-encryption-based contrastive loss and individual fine-tuning to safeguard the data security. In addition, the PRS eliminates the necessity of collaborative inference. In this article, comprehensive security analysis and proofs are conducted for PRS. Moreover, experiments demonstrate the superior performance of the proposed POS over classical VFL, showcasing a substantial performance improvement. Simultaneously, the PRS surpasses the performance of training alone, even under stringent security constraints.

The segmentation task of cloud and cloud shadow has always been one of the important tasks in remote sensing image processing. At present, cloud detection based on deep learning methods lacks generalization, which is easy to cause the loss of space and detail information, and missed detection and false detection occur from time to time. Aiming at the above problems, this paper proposes a multi-path, multi-scale attention network (MMA). In this network, Muti-scale overlapping patch embedding (MSP) is introduced to extract multi-scale semantic information with multi-path PVT, and strip convolution is used to supplement the spatial detail information of the image, so as to realize the effective aggregation of fine and rough features at the same feature level. In order to aggregate the global information, the multi-scale global aggregation module (MGAM) is used for deep feature extraction to supplement the high semantic information. In the decoding stage, aiming at the problem of small target cloud detection, the attention guided fusion module (AGFM) is proposed to focus on the important information of the image, remove the network noise and increase the detection accuracy of small targets. Contextual information fusion (CIF) decoding method is proposed for the coarse segmentation boundary problem, which fully integrates the context information and effectively helps to restore the image. The experimental results on Biome 8 Dataset, HRC-WHU Dataset and SPARCS Dataset confirm that our method is superior to the current cutting-edge cloud and cloud shadow detection technology.

Accurate forecasting of multienergy loads is essential for designing, operating, scheduling, and managing integrated energy systems (IESs). Recent research suggests that transformer models have the potential to improve long-sequence predictions. However, existing transformer models often emphasize capturing temporal dependencies while neglecting crucial dependencies among different variables necessary for multienergy load forecasting. Moreover, transformer models encounter challenges related to quadratic time complexity and significant memory usage, which hinder their direct applicability to tasks involving long-sequence, multienergy load forecasting. To tackle these challenges, we propose a model called DTformer and apply it to the task of multihorizon, multienergy load forecasting in IES. Within DTformer, we employ patch embedding to convert the input multienergy load sequences into a 3-D vector array, preserving both temporal and variable information. Subsequently, we propose the temporal top windowed attention (TWA) module and the dual variable attention module to handle extended temporal dependencies and intervariable dependencies. Importantly, the computational complexity and memory requirements of the TWA model are regulated at a level of$O(N^\frac{4}{3})$. Through extensive experimentation, we found that our DTformer surpasses baseline models in terms of performance using the IES dataset sourced from Arizona State University's Tempe campus.

Cloud and cloud shadow segmentation is one of the most important problems in remote sensing image processing. Due to the vulnerability to ground object interference, noise interference and other factors, and the lack of generalization ability, traditional deep learning network would inevitably lose details and spatial information, resulting in imprecise segmentation of cloud and cloud shadow boundaries, missing detection and false detection. In order to solve these problems, a multi-scale attention feature aggregation network is proposed, which extracts semantic information at different levels based on residual network. A multi-scale strip pooling attention module is designed to extract multi-scale context information and deep spatial channel information. In order to improve the feature extraction of the model, the deep multi-head feedforward transfer attention module is introduced to enable the adjacent two layers of the backbone network to guide each other for feature mining. Then, a bilateral feature fusion module is used to guide the fusion of low-level semantic information and high-level detail information. Finally, in order to enhance the repair of boundary details of cloud and cloud shadow, a boundary refinement boosting model is designed. The experimental results show that our model can handle complex cloud cover scenes, and has excellent performance on cloud, cloud shadow dataset and CSWV dataset. Its segmentation accuracy is superior to the existing methods, which is of great significance to the related work of cloud detection.

Traditional building and water segmentation methods are vulnerable to noise interference, and hence they could not avoid missed and false detections in the detection process. Excessive deep learning downsampling would lead to significant loss of feature map information, and image location information offset, and the overall effect of falling apart. To address these issues, a Multi-Scale Location Attention Network (MSLA) is proposed. Location-spatial information and channel information are particularly important for edge detail segmentation in building and water cover. The network includes a Location Channel Attention Unit (LCA) to focus on tributary details of rivers and segmentation of building edge eaves. Moreover, this paper builds a Dual-Branch Multi-Scale Aggregation Unit (DBMSA) to obtain deeper multi-scale semantic information. Finally, the Multi-Scale Fusion Unit (MSF) is used to guide the information merging of multiple stages, and the boundary information is improved by splicing the acquired deep multi-scale information with the information of the relevant feature extraction layer in the downsampling. The experimental results on several datasets show that the proposed approach outperforms other methodologies in segmentation accuracy.

Parameter identification plays an important role in power system. The existing parameter identification methods usually have two limitations: 1) the existing methods only consider the information of single branch and ignore the influence of adjacent branch information, and do not effectively use the topological structure constraints of the power grid; 2) the measurement data has the problems of poor numerical stability, numerical divergence and noise interference. Data contamination has become an important factor restricting the prediction accuracy of branch parameters. In order to solve these problems, this work proposes an automatic weighted loss-graph convolution networks model combined with the spatial structure of power grid. Based on the graphical modeling, the correlation between power grid branches is effectively used, and the adjacent branches are used to provide more effective information for the parameter identification of a branch, which improves the identification accuracy. In addition, the model enhances the local characteristics through the power grid structure information constraints, realizes the local optimal fitting, and constructs a self tradeoff loss function to weaken the impact of pollution data (data noise and data loss) on the model. The test results show that compared with the traditional method, this method has higher accuracy and stronger robustness.

Cloud and cloud shadow segmentation is one of the most important issues in remote sensing image processing. Most of the remote sensing images are very complicated. In this work, a dual-branch model composed of Transformer and convolution network is proposed to extract semantic and spatial detail information of the image respectively to solve the problems of false detection and missed detection. To improve the model’s feature extraction, a Mutual Guidance Module is introduced so that the Transformer Branch and the Convolution Branch can guide each other for feature mining. Finally, in view of the problem of rough segmentation boundary, this work uses different features extracted by the Transformer Branch and the Convolution Branch for decoding, and repairs the rough segmentation boundary in the decoding part to make the segmentation boundary clearer. Experimental results on the Landsat-8, Sentinel-2 data, the public dataset HRC_WHU of Wuhan University and the public dataset SPARCS demonstrate the effectiveness of our method and its superiority to existing state-of-the-art Cloud and Cloud Shadow Segmentation approaches.

A weighted densely connected convolution network (W-DenseNet) is proposed for reinforcement learning in this work. The W-DenseNet can maximize the information flow between all layers in the network by cross layer connection, which can reduce the phenomenon of gradient vanishing and degradation, and greatly improves the speed of training convergence. The weight coefficient introduced in W-DenseNet, the current layer received all the previous layers’ feature maps with different initial weights, which can extract feature information of different layers more effectively according to tasks. According to the weight adjusted by learning, the cross-layer connection is pruned to remove the cross-layer connection with smaller weight, so as to reduce the number of cross-layer. In this work, GridWorld and FlappyBird games are used for simulation. The simulation results of deep reinforcement learning based on W-DenseNet are compared with the traditional deep reinforcement learning algorithm and reinforcement learning algorithm based on DenseNet. The simulation results show that the proposed W-DenseNet method can make the results more convergent, reduce the training time, and obtain more stable results.

The biometric information contained in a face image is affected by many factors such as living environment, racial differences, and genetic diversity, this complexity leads to the nonstationary of the age estimation. In order to reduce the overlap of face features between adjacent ages and improve the accuracy of age prediction, a multi-stage feature constraints learning method is proposed for face age estimation. The proposed method gradually refines the feature through three feature constraint stages. In each stage, the algorithm continuously updates the feature center of its corresponding age range, and minimizes the distance between each age feature and feature center of the corresponding age range through feature constraint. Feature constraint makes the feature distances between different individuals in the same age feature space smaller and decrease the overlap areas between adjacent age range feature spaces. Meanwhile, the feature distance of different age range feature space is enlarged. The proposed network efficiently merges the features of three stages and optimizes the mapping of feature maps to an ordered binary comparison space. Experiments show that the proposed method is able to effectively improve the discrimination between different age features, and hence to improve the accuracy of face age estimation. In addition, the proposed algorithm is simple enough to achieve fast face age estimation.

Convolutional Neural Networks(CNNs) have achieved great success for object recognition in still images. However, CNNs can't make evident improvement for action recognition in videos, one reason is that many current network architectures are relatively shallow compared with deep models in image domain, and the other reason is that CNNs can't capture effective long-term motion information from videos. Encouraged by the good performance of Residual Network-s(ResNets) for training extremely deep models, and Long-term Recurrent Convolutional Networks(LSTMs) for dealing with tasks involving sequences, we presented an action recognition method based on a two-stream architecture, with 2D ResNets with LSTMs in one stream and designed 3D ResNets with LSTMs in the other stream, which can combine appearance and motion information better. Especially, our proposed method first learns spatiotemporal features of videos through the Residual networks, then models complex temporal dynamics by the Long-term Recurrent Convolutional networks, and with a softmax layer on the top of two streams, the final classification results can be predicted by fusing scores of each stream with weights on score distribution. Furthermore, for better reducing the influence of redundant background information in videos for recognition results, we also applied a center extraction method to generate central regions of videos instead of an entire video into a visual representation. On two video action benchmarks of UCF101 and HMDB51, our method achieved promising performance compared with state-of-the-art.

Sequence information processing, for instance, the sequence memory, plays an important role on many functions of brain. In the workings of the human brain, the steady-state period is alterable. However, in the existing sequence memory models using heteroassociations, the steady-state period cannot be changed in the sequence recall. In this work, a novel neural network model for sequence memory with controllable steady-state period based on coherent spininteraction is proposed. In the proposed model, neurons fire collectively in a phase-coherent manner, which lets a neuron group respond differently to different patterns and also lets different neuron groups respond differently to one pattern. The simulation results demonstrating the performance of the sequence memory are presented. By introducing a new coherent spin-interaction sequence memory model, the steady-state period can be controlled by dimension parameters and the overlap between the input pattern and the stored patterns. The sequence storage capacity is enlarged by coherent spin interaction compared with the existing sequence memory models. Furthermore, the sequence storage capacity has an exponential relationship to the dimension of the neural network.