To address the challenges of carbon emission prediction in complex energy systems within industrial parks, this paper proposes a Categorical Boosting (CatBoost) ensemble learning model that integrates a node-semantic attention graph network (NSAGN) and an external attention-based bidirectional long short-term memory (EA-BiLSTM). First, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) technique is employed to preprocess the data, effectively eliminating anomalies and noise in nonstationary multi-energy load and carbon emission data, thereby providing high-quality input data for subsequent prediction models. Second, the NSAGN is constructed to enhance the ability to capture associations in heterogeneous data, enabling accurate extraction of hierarchical features related to energy consumption and carbon emissions. Subsequently, an external attention (EA) mechanism is introduced into the BiLSTM to achieve adaptive weight allocation, further improving the model’s ability to capture local and global feature variations in carbon emission data and enhancing its effectiveness in mining time-series features. Finally, CatBoost is utilized to perform weighted integration of the predictions from the two branches, fully leveraging the advantages of time-series analysis and nonlinear learning of heterogeneous data, thereby improving the accuracy of carbon emission prediction. Comparative experimental analysis demonstrates that the proposed model outperforms comparative models in terms of evaluation metrics such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination ([Formula: see text]) across different prediction horizons, enhancing the accuracy and robustness of carbon emission prediction. This provides strong support for carbon emission assessment in industrial parks.

To address the issues of loss of defect details, low brightness, and indistinct defect features on photovoltaic (PV) panel images caused by variations in electroluminescence radiation intensity during defect detection, an improved CycleGAN-based infrared image enhancement method is proposed. A generator with an Auto-Encoder network structure featuring a self-attention mechanism is designed to address the poor feature extraction capability of the original network structure. A Unet-structured network is proposed for the discriminator to replace the original structure, allowing the discriminator to train independently of the generator. The training ratio between the generator and discriminator is adjusted to 1:3 to enhance the guidance effect of the discriminator on the generator. Experimental results show that the proposed algorithm outperforms existing image enhancement algorithms on the PVEL-AD dataset in both qualitative and quantitative evaluations. It effectively enhances the quality, clarity, and defect feature details of the electroluminescence images of PV panels. In practical detection processes, it improves the mAP50, mAP50:95, and recall rate by 3.9%, 3.5%, and 4.2%, respectively, further demonstrating the effectiveness of the proposed method.

Various service requests from power consumers, including maintenance, consultations, complaints, etc., will be converted into electrical service work orders and assigned to relevant departments for processing. Although the traditional first-come-first-served strategy is fair, a large number of work orders often lead to delays for important and urgent work orders, resulting in inefficient service and causing significant losses for both users and power supply companies. To address this challenge, an electrical service work order priority evaluation method based on sentiment analysis is proposed. Its innovation lies in the creation of a dedicated lexicon that integrates both power features and user emotions. Combined with a weighted word vector model, it can effectively capture the severity of power issues and the emotional intensity of users, thereby aligning the urgency of user service requests with the priority of work order processing. Experimental results show that this new sentiment lexicon achieves satisfactory results in evaluating the priority of electrical service work orders. This study lays the foundation for enabling efficient and automated electrical service work order dispatch.

Automatic recognition of forms to efficiently utilize form data has become an important requirement in areas such as business and management. Accurately segmenting the table structure is a key step in the automatic table recognition technology. However, when facing complex forms with non-standard layouts, complex backgrounds, and distorted and fuzzy forms, the traditional table structure segmentation methods still have the problem of low accuracy. So to solve the problem, this study proposes a table structure segmentation method based on multi-scale feature enhancement. Firstly, we add a convolutional neural network (CNN) branch based on the framework of Swin-U net semantic segmentation model to extract form image features from three different resolution scales, namely, low, medium, and high; secondly, we introduce a Multi-Scale Feature Attention (MFA) mechanism in the branch of Swin-transformer to achieve the more accurate capture of the three-scale features; and finally, a multi-scale feature fusion module-CTFM is constructed to fuse the extracted features to enhance the segmentation ability of the model on the table structure. We trained our model on one of our own manually collected labeled complex table datasets (CWTD) and validated it with publicly available SciTSR and PubTabNet data, where our method achieves the best Dice_coef scores of 96.72% and 96.52% respectively, proving that our method has more accurate segmentation.

Thermal imaging is often compromised by dynamic, complex degradations caused by hardware limitations and unpredictable environmental factors. The scarcity of high-quality infrared data, coupled with the challenges of dynamic, intricate degradations, makes it difficult to recover details using existing methods. In this paper, we introduce thermal degradation simulation integrated into the training process via a mini-max optimization, by modeling these degraded factors as adversarial attacks on thermal images. The simulation is dynamic to maximize objective functions, thus capturing a broad spectrum of degraded data distributions. This approach enables training with limited data, thereby improving model performance. Additionally, we introduce a dual-interaction network that combines the benefits of spiking neural networks with scale transformation to capture degraded features with sharp spike signal intensities. This architecture ensures compact model parameters while preserving efficient feature representation. Extensive experiments demonstrate that our method not only achieves superior visual quality under diverse single and composited degradation, but also delivers a significant reduction in processing when trained on only fifty clear images, outperforming existing techniques in efficiency and accuracy. The source code will be available at https://github.com/LiuZhu-CV/DEAL.

To enhance multi-load prediction accuracy in integrated energy systems, this paper proposes a load prediction method based on CDA-SynFusion and CPO-VMD, aiming to innovate from three levels: anomaly point identification, feature mining and modal decoupling. First, a CDA-SynFusion (Cross-Domain Adaptive Synergistic Feature Fusion) preprocessing framework is constructed, where a multivariate synergistic anomaly detection algorithm combines temporal fluctuation patterns, load synchrony constraints, and holiday effects to accurately identify and correct outliers, ensuring high-quality input data. Second, a cross-domain temporal feature screening and fusion mechanism is developed to dynamically quantify the correlation between load and meteorological domains, enabling the construction of a feature matrix with strong representational capability. Finally, the CPO (Crested Porcupine Optimizer) is introduced to adaptively optimize key VMD parameters, eliminating empirical dependency and achieving precise decoupling of cross-domain multi-frequency components. Experimental results verify that the proposed CDA-SynFusion-CPO-VMD framework significantly improves prediction accuracy across electrical, cooling, and heating loads, demonstrating its effectiveness in data preprocessing and modal decomposition, as well as its strong applicability to integrated energy system forecasting.

Early-stage software defect prediction faces scarce labeled data, limiting supervised learning effectiveness. Existing semi-supervised methods suffer from low-quality pseudolabels and CoForest's homogeneous classifiers restrict pattern capture. This paper proposes Enhanced CoForest (ECoForest), featuring heterogeneous classifiers for sample adaptability, dynamic confidence pseudo-label selection, and adaptive population mechanism. Experiments on eight defect datasets show ECoForest achieves 0.8231 AUC with only 20% labeled data- 7.1% higher than CoForest, with 3.3% F1-score and 6.9% G-mean improvements. ECoForest outperforms four baseline methods under limited labeled data conditions.

The rapid evolution of malware variants and imbalanced family distributions pose significant challenges to accurate classification and system security. This paper introduces an adaptive multimodal learning framework for robust malware family classification. The key innovation lies in a Modality-Aware Feature Selection (MAFS) mechanism that dynamically selects the most discriminative features from byte-level, structural, and semantic modalities. Coupled with an adaptive fusion strategy, our approach effectively addresses feature redundancy, modality heterogeneity, and missing data scenarios. Extensive experiments on 5,841 samples from 10 families demonstrate state-of-theart performance, with 99.87% accuracy, 99.82% macro-F1, and 0.0175 log loss, significantly outperforming existing methods in handling both feature loss and class imbalance.

Aiming at the problems such as large memory occupation and slow speed caused by excessive data usage when using the graph-based representation method to detect defects in the software source code during the software update and maintenance process, a lightweight defect detection method for graph data based on the source code change position is proposed. The innovation point lies in that during the data preprocessing process, after fully extracting the dependency relationships between source code statements and the semantic information of the statements by constructing the program dependency graph through extracting source code data, the content and location of the source code changes in one submission are determined, thereby inferring the parts that have an impact on the source code in one change. Ultimately, the amount of data is reduced by pruning the dependency graph to remove the parts that do not participate in the change and are not affected by the change, thereby reducing the computational overhead. The processed data is more suitable for defect detection during software update and maintenance, overcoming the problem of excessive quantity caused by the original graph-based preprocessing method, and thus optimizing the effect of the model. Through experimental verification, this method reduces the data volume by approximately$58.91 / \%$on the dataset constructed with jmeter, lowers the runtime memory usage by approximately 35.89 %, and increases the speed by approximately 41.87 %. And verified through the log information of the libgdx software, there are obvious changes in the above three aspects.

Aiming to the problem that the present table structure recognition methods often have the problem of misrecognition and misalignment of cells when facing complex tables with complex row-column structure and abundant semantic content of cells, this study proposes a multi-modal complex table structure recognition network-MTSRN, which integrates visual, textual and positional features. The innovation of this method lies in, firstly, adding CANN image visual feature extraction branch on the basis of GNN to realize more comprehensive extraction of global and local visual features of the table and forming pixel location network by combining the relative positions of cells; secondly, we construct the integrated network called CLAT to obtain semantic features among text contents in order to accurately extract semantic connections between cell text contents; finally, we construct a node prediction module called NAPM, which interactively enhances visual, semantic, and positional features, and employs GNN to fuse the three features into graph nodes, as well as predicts and pairs the column relationships between nodes, in order to enhance the recognition ability of the table structure to eventually realize the accurate reconstruction of the table structure. The experimental results show that our method in this paper has advantages in precision, recall and F1 compared with the current popular methods for table structure recognition, proving that our method can achieve more accurate recognition results for complex tables in realistic scenarios. Meanwhile, our designed recognition model gets nearly 13% improvement in Row_prediction and Col_prediction metrics respectively compared with the baseline model.

In order to address the challenges of distributed power load forecasting with limited resources and the need to protect user privacy, this paper proposes a lightweight federated learning architecture, Deep Separable Temporal Convolutional Network (DS-TCN). The model consists of three core modules: parallel multi-branch inflated causal convolution is used to capture short-term mutations and long-term trends simultaneously; adaptive scale fusion module is used to dynamically weight the features of each branch; and the dynamically aware prediction head automatically adjusts the response strength according to the signal fluctuations through a gating mechanism. In the experiments on the public HUE dataset, all clients of DS-TCN perform 5 rounds of local iterations and 50 rounds of federated training, which dramatically reduces the communication and computation overheads compared to the 500 rounds of federated training of the LSTM model. Moreover, the DS-TCN federated global model achieves RMSE=0.1382, MAE=0.1267, and MSE=0.0194, which is equal to or even exceeds LSTM; the average RMSE at the local end is reduced from 0.0796 in LSTM to 0.0699, with the highest single-household reduction of nearly 49%, which fully proves the superiority of the proposed method.

By analyzing sentiment changes on social media, time series methods are able to capture the sentiment fluctuations of depressed patients at different time points, thus providing strong support for early diagnosis and risk prediction of depression. However, the existing models still have two major bottlenecks: insufficient global dependent modeling and lack of interpretability. To address the above problems, this paper innovatively proposes an attention-augmented dynamic time series model (GAIT), which realizes a breakthrough through a multi-stage architecture. First, multi-source feature fusion is performed to generate multivariate time series features by fusing sentiment lexicon and pre-trained model, which can effectively extract explicit and implicit features; second, dual interpretability is achieved by symptom-level similarity analysis and dynamic weight ablation; then, hierarchical feature extraction is performed to capture local patterns by multi-scale convolution of Inception module, and residual connectivity mitigates the gradient vanishing; and finally. Dynamic global modeling, embedding the global attention mechanism to achieve dynamic weighting of critical time steps, complemented by global average pooling to enhance robustness. The experimental results show that the model significantly outperforms the baseline model in the depression degree classification task, with F1-Scores of 90.0, 86.8, 84.0, and 87.7 on normal, mildly depressed, moderately depressed, and severely depressed users, respectively, which fully validates the effectiveness of the proposed method.

The issue of data sparsity poses a formidable challenge in the field of recommender systems. Encouragingly, leveraging the interactions among overlapping users in the source domain can enhance item recommendation in the target domain. The transfer of user preferences across domains is a crucial concern in the cross-domain recommendation and represents a hopeful method to address data sparsity. Most existing methods transfer users' preference information by building a preference transfer network. These methods focus on the cross-domain mapping of preference features and ignore the inherent data distribution differences between the source domain and target domain. Consequently, the mapped user embeddings do not align with the item embeddings in the target domain and the recommendation quality decreases. On this basis, we propose a new method called Adaptive Meta-Learning for Cross-Domain Recommendation (AMLCDR). The method includes a meta-learning network for fully extracting user characteristics and generating a transfer network to reduce the user preference loss, as well as a domain adaptation network to align user preference distributions. We perform comprehensive experiments to assess the efficacy of AMLCDR by utilizing a substantial real-world dataset. We validate the effectiveness of data distribution alignment in domain adaptation. For diverse cross-domain recommendation tasks under different start conditions, AMLCDR outperforms state-of-the-art models in multiple evaluation metrics.

Aiming at the problem that the traditional detection method has low efficiency and accuracy due to the fact that there are many key component targets to be inspected by unmanned aerial vehicle (UAV) during power inspection, and the shape difference is large, and the image quality is not high, a transmission line abnormal target detection method based on improved YOLOv8 and super-resolution reconstruction is proposed. Firstly, the super-resolution reconstruction algorithm is used to reconstruct the abnormal image to improve the clarity and enrich the characteristic information contained in the image. On this basis, the improved YOLOv8 network is used to detect the defects in the inspection image. The CBAM attention mechanism is fused in the Bottleneck part of the C2f module to strengthen the model's ability to locate the target; In order to further improve the detection ability of small targets in patrol inspection, a small target detection layer is added to make the network pay more attention to the detection of small targets. Finally, in order to be deployed to the edge devices, the original convolution in the network is modified to a lightweight convolution GhostConv to reduce the number of model parameters. The experimental results show that the proposed method can accurately detect abnormal defects of transmission line components on the basis of improving the quality of inspection images. mAP is improved by 3.7% and the number of model parameters and calculation amount are greatly reduced, which reflects the effectiveness of the algorithm, and it has stronger extraction ability and robustness for subtle defect targets, meeting the detection requirements of power inspection.

Worst case execution time (WCET) analysis is essential for exposing timeliness defects when developing hard real-time systems. However, it is too late to fix timeliness defects cheaply since developers generally perform WCET analysis in a final verification phase. To help developers quickly identify real timeliness defects in an early programming phase, a novel interactive WCET prediction with warning for timeout risk is proposed. The novelty is that the approach not only fast estimates WCET based on a control flow tree (CFT), but also assesses the estimated WCET with a trusted level by a lightweight false path analysis. According to the trusted levels, corresponding warnings will be triggered once the estimated WCET exceeds a preset safe threshold. Hence developers can identify real timeliness defects more timely and efficiently. To this end, we first analyze the reasons of the overestimation of CFT-based WCET calculation; then we propose a trusted level model of timeout risks; for recognizing the structural patterns of timeout risks, we develop a risk data counting algorithm; and we also give some tactics for applying our approach more effectively. Experimental results show that our approach has almost the same running speed compared with the fast and interactive WCET analysis, but it saves more time in identifying real timeliness defects.

With the tremendous amount of research publications, it has become increasingly important to provide a researcher with a rapid and accurate recommendation of a list of reference papers about a research field or topic. In this paper, we propose a unified graph model that can easily incorporate various types of useful information (e.g., content, authorship, citation and collaboration networks etc.) for efficient recommendation. The proposed model not only allows to thoroughly explore how these types of information can be better combined, but also makes personalized query-oriented reference paper recommendation possible, which as far as we know is a new issue that has not been explicitly addressed in the past. The experiments have demonstrated the clear advantages of personalized recommendation over non-personalized recommendation.