Abstract:

Abstract: In response to the insufficient robustness of current safety clothing and helmet wearing detection algorithm in complex backgrounds, and frequent false positive rates, an improved YOLOv7-tiny detection algorithm for safety clothing and helmet wearing on construction site was proposed. Firstly, the EMA attention mechanism was introduced into the feature extraction module to enhance the networks feature extraction capability and mitigate complex background interference. Secondly, the RFEM module was integrated into the feature fusion stage to improve the networks receptive field, acquiring broader contextual information, and enhance perception for small targets. Finally, Shape-IoU was employed to replace the IoU boundary regression loss function, to improve detection accuracy. Experimental results showed that the improved model achieved a mAP@0.5 of 90.4% on the proprietary dataset, 3.0 percentage points higher than the original model. The frame rate speed reached 93 frames/s, and the model size comprised only 6.1 million parameters. Compared to YOLOv8s, YOLOv9s, and other models, the proposed algorithm demonstrated superior performance in detection accuracy, speed, and model efficiency, making it suitable for real-time detection applications on construction site.

Abstract: To address the challenge of balancing prediction accuracy and efficiency in time series forecasting, in this paper a lightweight time series forecasting model named ILformer was proposed, built upon the iTransformer architecture. As a representative variable-based model for temporal data, iTransformer effectively captured complex inter-variable dependencies. However, it was constrained by high computational complexity and a substantial parameter footprint, limiting its practical deployment in resource-constrained scenarios. To mitigate these limitations, ILformer incorporated the following enhancements:the model first introduces a Linear Attention mechanism to replace the traditional attention mechanism, allowing for more flexible input processing. By leveraging linear projection and dimension rearrangement, ILformer significantly reduced the number of parameters while better adapting to varying input shapes and structures. It achieved high computational efficiency, particularly when handling largescale datasets, and drastically lowered the computational complexity of the attention module without compromising model accuracy. Furthermore, singular value decomposition (SVD) was incorporated into the attention mechanism to achieve matrix dimensionality reduction. This approach substantially decreased the number of matrix multiplications and additions, improving computational efficiency and mitigating the risk of overfitting. Experimental results on eight diverse datasets demonstrated that ILformer achieved a 40.46% improvement in inference speed on average while maintaining the same level of accuracy. Additionally, the number of parameters was reduced by 78.75%, and the operations were halved, underscoring its superior performance and practical applicability.

Abstract: In expensive optimization problems, if the optimal solution of the problem was not unique, such problems were referred to as expensive multimodal optimization problems. However, it was extremely difficult to obtain multiple optimal solutions with limited computational resources. Moreover, existing surrogate-assisted evolutionary algorithms paid less attention to multimodal attributes. In view of this, a surrogate-assisted multi-population differential evolution algorithm based on region decomposition was proposed to solve expensive multimodal optimization problems. Firstly, in the population individual initialization stage, the correlation between inter-individual distances and objective values was used to detect potential sub-regions, and sub-populations were divided to explore multiple optimal solutions. Secondly, in the early stage of evolution, the differential evolution algorithm was used to perform global search in each sub-population to capture multiple optimal solutions. After multiple optimal individuals were obtained in the early stage of evolution, the covariance matrix adaptive evolution strategy was adopted to carry out local search on them to improve the quality of optimal solutions. In addition, an infilling criterion was proposed, which could adaptively select appropriate individuals for real evaluation according to specific parameters to improve the accuracy and generalization ability of the surrogate model. Finally, the proposed algorithm was compared with seven other algorithms on 20 test functions. The results showed that the proposed algorithm achieved optimal performance on 13 functions with the PR metric, and was slightly inferior to the comparison algorithms on at most 5 functions. Overall, the proposed algorithm exhibited excellent performance in solving expensive multimodal optimization problems.

Abstract: Backdoor attacks pose a serious threat to the security of deep neural networks. Most existing backdoor defense methods relied on partial original training data to remove backdoor from models. However, in real-world scenarios where these data access was limited, these methods performed poorly in eliminating backdoor and often significantly impact the model′s original accuracy. To address these issues, in this study proposes a data-free backdoor removal method was proposed based on pruning and backdoor unlearning (DBR-PU). Specifically, the proposed method first analyzed the pre-activation distribution differences of model neurons on a synthetic dataset to identify suspicious neurons. Then, it reduced the impact of backdoor by pruning these suspicious neurons. Finally, an adversarial backdoor unlearning strategy was employed to further eliminate the model′s internal response to any residual backdoor information. Extensive experiments on the CIFAR10 and GTSRB datasets against six mainstream backdoor attack methods demonstrated that, under data access constraints, the proposed method achieved a minimal accuracy gap compared to the best baseline defense methods and performed the best in reducing attack success rates.

Abstract: Vehicle re-identification serves as the foundation for vehicle tracking. However, pedestrian interference significantly degrade the quality of feature extraction, thereby reducing the accuracy and precision of vehicle re-identification. To address this issue, a vehicle re-identification algorithm named TRaBS was proposed, with track refinement and background suppression techniques. Firstly, ResNeXt101-IBN-a was employed to extract initial features from vehicle images, while a background suppression algorithm was applied to the camera frames to generate background-suppressed vehicle features. Secondly, to mitigate the impact of pedestrian interference on feature extraction, Euclidean distance and Gaussian kernel functions were utilized to replace image-level features with more stable trajectory-level representations. Through these techniques, the problem of pedestrian-induced interference in vehicle re-identification was effectively alleviated. To comprehensively evaluate the effectiveness of TRaBS in handling pedestrian interference and its generalization ability in interference-free scenarios, extensive comparative and ablation experiments were conducted. Experimental results demonstrated that the vehicle re-identification model integrated with trajectory optimization and background suppression achieved significant performance improvements on both benchmark datasets and derivative datasets with pedestrian interference. Specifically, on the VeRi-776 dataset, the model achieved a mean average precision (mAP) of 83.6% and a Rank-1 accuracy of 97.6%, better than all existing methods.

Abstract: To effectively solve the problem of triple overlap in the joint extraction of entities and relations in unstructured text. A Chinese entity and relation joint extraction method was proposed based on RoBERTa and pointer network. Firstly, for the entity overlap problem, an entity recognition module was based on the pointer network, and the entity recognition task was constructed as a token-pair recognition problem, which extracted designed all possible entities by recognizing the start and end positions of the entities. Secondly, for the triplet overlap problem, a relation extraction module was designed based on the multi-head attention mechanism and Ptr-Net to construct the triple (s, r, o) extraction task as a quintuple (sh, st, r, oh, ot) identification problem. Finally, extensive experiments on the Chinese information extraction dataset DuIE showed that the comprehensive performance of the proposed model was better than all baseline models, with the precision, recall and F1 values of 81.04%, 85.82% and 83.36% respectively.

Abstract: Constrained multi-objective evolutionary algorithm based on multi-tasking competition and cooperation has problems in resource allocation and collaborative optimization, resulting in low effectiveness populations wasting computational resources, and underutilized high-quality solution information. Therefore, in this study, a constrained multi-objective evolutionary algorithm based on competitive and cooperation multitasking was proposed, which included two main strategies. Firstly, a competition-based resource allocation strategy was proposed to achieve adaptive allocation of computing resources based on the historical performance of each task population. Secondly, a collaborative optimization strategy based on parent aggregation and offspring diffusion was designed to generate highquality offspring through cross-task cooperation and spread them to various task populations, achieving efficient utilization of effective information. The proposed algorithm was compared with five other advanced algorithms (CMOEA_MS, cDPEA, EMCMO, MTCMO, and CMOEMT) on 38 test functions, and the results showed that the proposed algorithm achieved optimal results on 25 and 26 functions with IGD and HV indicators, respectively, and was superior to the compared algorithms on at least 23 and 24 functions, respectively. The proposed algorithm had a feasibility rate of 100% on all functions and can effectively solve constrained multi-objective optimization problems.

Abstract: In order to improve the accuracy of short-term photovoltaic power forecasting, a model integrating an improved dung beetle optimizer, variational mode decomposition (VMD), and bidirectional long short-term memory (BiLSTM) was proposed. Firstly, a VMD-BiLSTM-based prediction framework was constructed, where time-series data were decomposed into multiple components via VMD and fed into BiLSTM for individual prediction. The final output was obtained by reconstructing the component-level results to enhance overall prediction performance. Subsequently, to address the tendency of the dung beetle optimizer (DBO) to fall into local optima, an improved DBO algorithm (IDBO) was developed through the introduction of four strategies: logistic chaotic mapping for initialization, Levy flight for global exploration, golden sine strategy for position updating, and adaptive T-distribution perturbation for local exploitation. Finally, the IDBO was utilized to optimize critical parameters, including the decomposition number K and penalty factor α in VMD, as well as the hidden layer size and Dropout ratio in BiLSTM, thereby enhancing the model′s learning capability and mitigating overfitting. The proposed model was experimentally tested using actual data from photovoltaic power stations in Shandong and Hebei provinces. Compared to the unimproved model DBO-VMD-BiLSTM, the results showed that the proposed model had the best MAE, MAPE and RMSE at two power stations.

Abstract: Aiming at the problems of increasing network losses and voltage overruns faced by power systems containing a high proportion of new energy, a two-layer planning strategy of soft open point integrated with energy storage system (E-SOP) for distribution networks considering demand response was proposed. Firstly, the typical scenarios of wind power output were generated based on Frank-Copula function considering the relevant characteristics of wind power output. Secondly, a two-layer planning model of E-SOP was established. The upper layer took the objective of the lowest annual comprehensive operating cost of the distribution network for the siting and capacity setting of ESOP, The lower layer took the demand response participation into consideration. The operation optimization was carried out with the objective of the minimum operating cost of each scenario, and the multi-strategy improved whale optimization algorithm (MIWOA) and Second-order Conic Programming (SOCP) were used to solve the model. Finally, the IEEE33-node systems were used for example analysis, and the simulation results showed that the annual integrated costs of the systems were reduced by 7.94%, respectively, which verified that the proposed scheme could effectively improve the stability and economy of distribution network operation.

Abstract: In order to improve the detection method of heating defects of electrical equipment and the recognition accuracy of the algorithm for infrared images of electrical equipment, a method of infrared image detection of electrical equipment based on data enhancement and improved YOLOv5 was proposed. Firstly, for the problem of low signal-to-noise ratio and low contrast of infrared images of electrical equipment, a fast guided filtering algorithm was used to denoise the infrared images in the data set. The CLAHE algorithm was improved by introducing Gamma correction, and then the contrast of infrared images was enhanced. The premise of detecting the heating defects of electrical equipment was to accurately identify and classify the equipment. Then, in order to improve the accuracy of the detection algorithm, the information aggregation and distribution mechanism was introduced to improve the feature fusion module based on the original YOLOv5 algorithm, which enhanced the multi-scale feature fusion ability. Meanwhile, the Focal-CIoU loss function was also introduced to make the algorithm pay more attention to high-quality samples and suppress low-quality samples, which enhanced the rate of convergence for the model. It was verified that the mAP value of the improved algorithm on the self-built data set was 93.6%, which was 4.0 percentage points higher than that before the improvement. Compared with Faster R-CNN, SSD, YOLOv3 and YOLOv7, the mAP value of the proposed algorithm was increased by 4.5 percentage points, 6.1 percentage points, 4.7 percentage points and 3.5 percentage points respectively, and the frame rate reached 32 frames per second, which could meet the real-time recognition requirements of electrical equipment.

Abstract: In the context of large-scale electric vehicle (EV) charging load of to the distribution network, the spatio-temporal uncertainty of EV charging behavior makes it difficult for traditional prediction methods to accurately describe its dynamic characteristics, which directly affects the effectiveness of distribution network operation optimization and charging scheduling decisions. In order to solve the problem of strong dependence on the timing of EV charging load and complex influencing factors, a charging load prediction method for EVs based on parameter optimization and attention mechanism (AM) was proposed. Firstly, the input feature data were screened by Pearson for feature screening (FS). Then, based on the CNN network, an improved AM network was introduced to construct the CNNAM network. Furthermore, an EV charge load CNNAM-BiLSTM combined forecasting model was proposed, which used a multi-layer convolution layer, AM, and two-way structure of BiLSTM to improve the mining of charge load characteristic data and time series, and sparrow search algorithm (SSA) to adaptively optimize the parameters in the forecasting model. Finally, based on the actual load data of charging stations in Wuhan, the SSAFS-CNNAM-BiLSTM combination forecasting model proposed was compared with the traditional deep learning forecasting model and combination forecasting model. The results showed that the proposed forecasting method achieved better forecasting results and had stronger adaptability in the complex dynamic environment.

Abstract: In order to achieve a win-win situation between micro grid and hydrogen energy storage power station, in this study a combined heat and power (CHP) supply based micro-grid was built by considering the services of hydrogen energy storage power station and demand response. The system combines electricity, heat and hydrogen energy and realized the energy conversions among them.And the waste heat recovery was included to effectively improve the whole energy utilization. Then, a double planning model with two optimization problems was presented. The upper-level model focused on optimizing hydrogen energy storage power stations and aiming to minimize the total operating cost of these stations. The lower-level model analyzed the optimization operation model of microgrid system with hydrogen energy storage power stations to achieve the goal of minimizing the total operating cost of the CHP based microgrid. Based on the Lagrange function and KKT condition, lower level target were transformed and added into upper level as constraint conditions, and then the large-M method was used to solve the problems by transforming the nonlinear programming problem into a mixed integer linear programming problem. The impacts of price type and alternative load ratio on the system were analyzed, and the appropriate ratio could be given to maximize the benefits of the system. Considering the characteristics of two typical seasons, winter and summer, three scenarios were selected for comparative analysis. The results verified the feasibility and effectiveness of the model. After reasonably adjusting the load proportion of demand response, the system had lower operational cost and better economic efficiency.

Abstract: To address the issue of bus voltage fluctuations caused by reduced system inertia and insufficient reactive power regulation capability due to the high penetration of power electronic devices in new energy bases, a voltage control strategy was proposed for high penetration sending-end systems, wherein grid-forming wind turbines were actively engaged in voltage regulation. Firstly, a grid-following static var generator was employed to provide fast reactive power response and suppress initial voltage fluctuations. Then, grid-forming wind turbines emulated the electrical and mechanical characteristics of synchronous generators to enhance system voltage support capability. Considering the capacity limitations of converters, an adaptive active power shedding control was applied in severe voltage sag conditions to improve reactive power compensation capability. Meanwhile, constant power and constant DC voltage control strategies were designed to achieve power balance between the sending and receiving ends of the flexible DC transmission system and reduce DC bus voltage fluctuations. Finally, a simulation model of a new energy base based on the IEEE 13-bus system was established in MATLAB/Simulink to validate the effectiveness of the proposed control strategy. Compared with conventional voltage regulation methods, the proposed approach could reduce voltage deviation by 18% with sudden reactive load changes and by 50% in the event of voltage sags caused by short-circuit faults. These results clearly demonstrated that the proposed strategy significantly could improve the voltage recovery capability of high-penetration sending-end systems with fault-induced disturbances.

Abstract: As a key component of pre-set arc suppression coil, the damping resistor is currently configured primarily to suppress series resonance overvoltage and prevent protection maloperation during normal operation. This leads to difficulty in the neutralpoint displacement voltage exceeding the alarm threshold during high-resistance ground faults, resulting in the failure to activate the line selection device. To improve the sensitivity of resonant grounding systems in detecting high-resistance ground faults, the system′s normal operation characteristics and the behavior during high-resistance ground were quantitatively analyzed faults. Using the conditions that the displacement voltage should remain within limits during normal operation and reliably trigger an alarm during a high-resistance fault as constraints, calculation formulas was derived for the critical values of series and shunt damping resistors. Then the influence of parameters such as detuning degree, transition resistance, system unbalance, and capacitive current on these critical values were systematically examined, and principles and methods were proposed for selecting damping resistors. Finally, to address the insufficient detection capability for high-resistance ground faults in a typical resonant grounding system, optimization method proposed in this paper was applied. The original 15 Ω series damping resistor was replaced with either a 5.05 Ω series resistor or a 346.37 Ω shunt resistor. The optimized damping resistor effectively suppressed series resonance overvoltage while ensuring reliable exceedance of the displacement voltage during high-resistance ground faults, thereby increasing the system′s tolerance to transition resistance to 2.78 kΩ. The shunt damping configuration demonstrated better performance. This approach could be implemented using only routine online monitoring data from the arc suppression coil, requiring no additional detection equipment, and offers strong feasibility and cost-effectiveness.

Abstract: Reasonable data analysis and accurate prediction of deformation monitoring data for concrete dams are key means to ensure the safe and long-term operation of dams. In response to the periodic and nonlinear characteristics of environmental variables that could affect dam deformation, as well as the shortcomings of traditional random forest model parameter optimization methods such as poor applicability and low computational efficiency, a new type of dam deformation prediction model was proposed. The model uses t-distributed random neighborhood embedding to reduce the dimensionality of eigenvalues and improve the classification performance of the model. The traditional random forest model was improved using the northern eagle optimization algorithm, which enhanced the efficiency of selecting optimal parameters for the random forest model. The parameters of the random forest model could be determined using the northern eagle optimization algorithm in the 80th iteration, and the fitness function was 0.249 3, which achieves better results compared to the Sparrow Search Algorithm and Particle Swarm Optimization Algorithm. The analysis of the 18#th and 26#th sections of a concrete dam showed that the fusion model proposed in this study had average absolute errors of 0.501 93 and 0.173 02 mm, mean square errors of 0.359 71 and 0.043 87 mm2, average absolute percentage errors of 0.819 59% and 0.113 62%, and determination coefficients of 0.914 56 and 0.892 74, respectively. Compared with other models, this model performed better in prediction accuracy and model stability, opening up new possibilities for accurate prediction of concrete dam deformation.

Abstract: To address the difficulty of obtaining underwater topography data for large water bodies with insufficient data, Danjiangkou Reservoir was selected as the study area, and a retrieval method based on Landsat remote sensing imagery and water depth zoning was proposed. The underwater topography of the shallow and deep water areas of the reservoir was reconstructed using the waterline kriging interpolation method and four water depth inversion models (single-band, dual-band ratio, BP neural network, and multi-band random forest), and the inversion accuracy was evaluated. The results showed that the underwater topography inversion in the shallow water area performed well (Root Mean Square Error, RMSE=2.553 m). In the deep water area, the multi-band random forest model performed best in the Han Reservoir area (RMSE=2.428 m), while the BP neural network model performed best in the Dan Reservoir area (RMSE=1.599 m). The accuracy of different inversion models varied across different depths and regions, with the multi-band random forest model demonstrating advantages in deep-water topography inversion.

Abstract: This study investigated the spatio-temporal characteristics and influencing mechanisms of groundwater drought in the Beijing-Tianjin-Hebei Region, aiming to provide scientific support for sustainable water resource management and promote high-quality regional development. Using Gravity Recovery and Climate Experiment (GRACE) satellite data (October 2003 to September 2023) and Global Land Data Assimilation System (GLDAS) data, groundwater storage anomalies (GWSA) were retrieved for the study area. Based on these results, a groundwater drought index (GDI) was constructed, through which groundwater drought events were identified using run theory. The occurrence frequency and spatio-temporal patterns of groundwater drought were subsequently analyzed, followed by an examination of relationships between groundwater drought and various influencing factors using meteorological data and water resource bulletins. The results indicated that: ①Higher frequencies of groundwater drought occurred in the central-eastern region, with the highest frequencies of moderate-to-severe drought concentrated along the southeastern periphery; ②Groundwater drought events primarily clustered between 2014 and 2021, characterized by high frequency, wide spatial extent, but relatively low intensity; ③Seasonally, autumn and spring droughts were most severe in southeastern cities, while summer droughts were milder, correlating with agricultural irrigation activities during March to May and October to November; ④Interannually, groundwater drought intensified after 2014 following sharp precipitation declines, reaching maximum severity in 2020 when widespread moderate-to-severe drought covered the entire region, before alleviating in 2021 due to increased precipitation; ⑤The South-to-North Water Diversion Project effectively replenished surface water resources and facilitated shifts in water supply-demand patterns, playing a crucial role in mitigating long-term groundwater deficits.