[Objective] The influence of photovoltaics (PV)-assisted current and extraction current on conventional relay protection hinders the effective functioning of relay protection equipment. A multi-objective setting method for relay protection in distribution networks suitable for conditions with high permeability of distributed PV was proposed to address this problem, which is aimed at enhancing the rapidity, sensitivity, and selectivity of protection, and ensuring economic viability and practicality, thus effectively safeguarding the power grid security and supporting the widespread access of distributed PV. [Methods] The influence of PV-assisted current and extraction current on the protection configuration of the distribution networks was analyzed, and the problem of unwanted operation and refuse operation of distribution network protection caused by PV access was avoided by introducing distance protection and instantaneous current protection as the protection criteria. A multi-objective optimization model with the optimal parameters of protection rapidity, sensitivity and selectivity was built, and the particle swarm optimization (PSO) algorithm was improved by adopting the dynamic splitting operator to make the solution of the protection setting meet the practical application requirements. [Results] High-permeability distributed PV results in unwanted operation or refuse operation of distribution network protection, which is effectively avoided by introducing distance protection and instantaneous current protection as the protection criteria. The multi-objective optimization protection configuration model was built, and the evaluation indexes of the overall protection effect of a certain area were formed, with the solution of the protection setting completed based on PSO algorithm. Finally, the overall evaluation of the protection effect under high-permeability PV access was realized, with the rapidity, sensitivity, and selectivity of protection improved. [Conclusions] The results show that the combination of distance protection and instantaneous current protection can effectively avoid the influence of the PV-assisted effect on the conventional instantaneous current protection. The protection performance can be effectively improved by the proposed multi-objective optimization scheme. Under the equilibrium strategy, the rapidity, sensitivity, and selectivity increase by about 82.2％, about 3.8％, and about 33.1％, respectively. The innovation of this study is that the combination of distance protection and instantaneous current protection was adopted to form the protection criteria, thus avoiding the problem of unwanted operation and refuse operation of the distribution network protection due to PV access. Additionally, a multi-objective optimization scheme for protection settings was constructed, and PSO algorithm was improved by employing the dynamic splitting operator, thereby avoiding the limitations of PSO algorithm and improving the reliability and applicability of protection settings.

[Objective] With the development of artificial intelligence (AI) and 5G technology, data centers are called important infrastructure facilities for future development. However, there is a prominent contradiction between the high energy consumption characteristics of data centers and current low-carbon development needs. The optimization of relying solely on clean energy power supply faces bottlenecks, such as difficult waste heat recovery and low efficiency. In response, a large-scale hydrogen production capacity configuration method was proposed to study waste heat recovery in data centers, offering a new approach for the green low-carbon development of data center energy supply systems and enhancing green hydrogen production efficiency. [Methods] Firstly, the energy consumption structure of the data center was analyzed. The mathematical model of the output heat energy of the water-cooled data center was constructed, and the mathematical model of the influence of the electrolyte temperature change on the efficiency of electrolytic hydrogen production was established. The electrolyte temperature rise coefficient was proposed as the coupling node of the data center and electrolytic hydrogen production, which laid the foundation for the subsequent establishment of the waste heat water coordination mechanism. Secondly, the data center′s energy consumption characteristics and hydrogen needs were analyzed. Using waste heat recovery and heating electrolyte theory, the data center+clean energy+green hydrogen operation architecture and the matched mode were built, and a dynamic supply-demand balance model was established. Based on clean energy output and data center operations, various electrolytic hydrogen production modes were formulated. Finally, considering the data center load characteristics, clean energy output fluctuations, hydrogen energy market demand, and other factors, a hydrogen production capacity configuration model was constructed with system economy, carbon emissions, and renewable energy consumption rate as optimization objectives. The multi-objective optimization method based on the improved timing difference algorithm and particle swarm optimization algorithm was designed, and simulation analysis was carried out with Matlab. [Results] The data center+clean energy+green hydrogen coordinated operation mode can reduce the annual electricity energy consumption by 2.59％ under the typical day scenario. The pricing method of the auxiliary peak shaving market can guide the system to operate according to different objectives and can deal with the economy of the system to varying degrees and low carbon demand. This study achieves the structural transformation of the data center energy system through multi-dimensional technological innovation. [Conclusions] The proposed large-scale electrolytic hydrogen production capacity configuration method meets waste heat utilization needs of data centers and reduces reliance degree of the energy supply system on fossil energy. It offers a new technical path for creating a new energy system of “adjustable load-energy storage-energy supply” in data centers, supporting digital transformation of China′s economy and the steady advancement of its “dual carbon” goals.

[Objective] During verifying settings of relay protection equipment in substations, traditional methods mainly rely on manual verification or simple program verification. The manual verification accuracy varies, with relatively low verification efficiency. Simple program verification improves its efficiency to some extent, but there is room for further accuracy enhancement. To this end, a setting verification method for relay protection of intelligent substations based on deep learning was proposed. [Methods] Firstly, an improved convolutional recurrent neural network (CRNN) was employed to identify relay protection settings. Specifically, the convolutional neural network (CNN) was adopted to convert text images into feature sequences, followed by leveraging the recurrent neural network (RNN) to identify the feature sequences. Finally, the identification results were transcribed by adopting a dictionary-based connectionist temporal classification (CTC) loss function to obtain the setting text information. On this basis, the RNN module was enhanced by utilizing a convert gate unit, thus building a bidirectional convert gate long short-term memory (Bi-CGLSTM) model to achieve adaptive adjustment of data weights. Then, the setting verification was carried out by combining Chinese word segmentation technology. A complete dictionary of setting names was constructed, with the Levenshtein distance algorithm adopted to calculate the similarity between the text to be verified and the standard text. Additionally, an improved forward maximum matching algorithm was applied to match the setting text, thus completing the one-by-one setting verification of relay protection equipment in substations. [Results] 240 relay protection setting sheets from a power supply company that cover ten common equipment models were selected as experimental samples to validate the feasibility and effectiveness of the proposed method. The training parameter setting of the deep learning model was as follows:the iteration count of 100, learning rate of 0.001, and the Adam optimizer for adjusting weights and biases. The experimental results show that the identification accuracy of the improved CRNN model exceeds 97％, while the verification accuracy of the proposed method reaches 97.07％, with relatively shorter verification time and better overall performance than that of other comparative methods. [Conclusions] The identification accuracy of the setting text of substation relay protection in the context of big data can be effectively enhanced by the improved deep neural network. Additionally, verification accuracy can be ensured and verification efficiency can be significantly enhanced by the combination of the Levenshtein distance algorithm and the improved forward maximum matching algorithm. Powerful technical support is provided by the proposed method for the intelligent operation and maintenance of intelligent substations.

[Objective] With the high proportion integration of renewable energy sources such as wind and solar power into the grid, their inherent intermittency and volatility pose significant challenges to the voltage stability at the distribution end of the power grid. In particular, at the end of regional power grids, the uncertainty in wind-solar-load output increases the risk of rapid voltage drops, which may lead to equipment damage or even cascading failures. Existing studies have notable shortcomings in areas such as handling prediction errors in wind-solar-load output and multi-objective collaborative optimization. For example, the full-pure embedding sensitivity analysis method fails to adequately consider the influence of prediction errors, while the source-grid-load coordinated control framework ignores the interference of prediction errors on collaborative outcomes. To address these issues, this paper proposed a novel fast regulation algorithm for low voltage. By quantifying the uncertainty in wind-solar-load output, a multi-objective optimization model was developed that balances safety, performance, and cost. The model aims to achieve rapid and stable voltage regulation at the distribution end of the grid, thereby improving the reliability and adaptability of high-proportion renewable energy integration into the power grid. [Methods] The Collaborative Genetic Algorithm (CGA) was used as the core solution method. Firstly, precise probability density function models were established to account for the randomness in the output of wind power, photovoltaics, and load output. The output of wind power was quantified by combining the Weibull distribution of wind speed with the normal distribution of prediction errors. Photovoltaic output was associated with light intensity and photoelectric conversion efficiency, incorporating prediction errors. Load output was represented by a probability density function reflecting its volatility. Based on this, a low-voltage regulation model was developed with optimization objectives balancing safety, performance, and cost. The safety indicator quantified the total power loss at the distribution end of the grid, the performance indicator included the overall network loss and voltage deviation, while the cost indicator calculated the total lifecycle cost. Through integer-based mixed coding schemes and dynamically adjusted crossover and mutation probabilities, the algorithm effectively optimized the population and output the optimal solution that satisfied voltage stability margin requirements. [Results] Based on actual grid data from a region in Guangzhou, simulation experiments validate the effectiveness of the proposed algorithm. In terms of uncertainty handling, the proposed algorithm shows a significantly higher correlation between wind and photovoltaic output predictions and actual data compared to other traditional methods. This is due to the algorithm modeling output power prediction errors as random variables, which more accurately reflects the uncertainties in real-world systems. Regarding voltage regulation, when fluctuations in wind-solar-load output and increased load lead to voltage drops, the algorithm quickly and effectively restores node voltages to normal levels. Its performance outperforms traditional methods, such as those based on steady-state grid models and the double-loop voltage-current control algorithm. In terms of static voltage stability margin, the proposed algorithm maintains a high voltage stability margin of over 0.8 across various test scenarios, demonstrating strong voltage regulation capability. Furthermore, while ensuring voltage stability, the algorithm also considers the economic and performance efficiency of grid operations. [Conclusions] The fast regulation algorithm for low voltage effectively addresses the issue of low-voltage instability at the distribution end of power grids with high integration of renewable energy by deeply combining the wind-solar-load output uncertainty modeling and multi-objective optimization. This algorithm innovatively introduces probability density functions to quantify prediction errors, which significantly improves the accuracy of wind-solar-load output predictions. By using CGA for coordinated optimization of safety, performance, and cost targets, the algorithm achieves rapid dynamic voltage regulation. Experimental results show that the proposed algorithm outperforms traditional methods in terms of regulation speed, stability margin, and economic efficiency, which provides reliable technical support for intelligent grid control with high proportions of renewable energy integration. The research results not only have significant theoretical value but also demonstrate great potential in practical engineering applications. Future exploration will focus on voltage coordination control strategies across multiple time scales to continuously enhance the stability and economic efficiency of grid operations.

[Objective] In the operation and management of power system, the medium and low-voltage distribution network serves as a key link between power sources and users. Its operating efficiency and stability are directly related to the safety and reliability of the entire power system. Three-phase line loss, as an important indicator of the distribution network′s operational efficiency, not only reflects the energy loss during the power transmission process but also directly affects the voltage quality, power consumption, and safe operation of the power grid. However, the three-phase line loss data in the distribution network exhibit complex distribution characteristics, such as multi-modal and asymmetric features. During dynamic changes, it is difficult to accurately capture the inherent patterns and structures in the data, which reduces the accuracy of anomaly detection. Therefore, this paper proposed an intelligent method for the anomaly detection of three-phase line loss in medium and low-voltage distribution networks. [Methods] During the data collection process of the distribution network′s three-phase line loss, the data can be influenced by multiple factors such as electromagnetic interference and equipment errors, leading to the presence of significant noise and outliers. These noises not only reduce the signal-to-noise ratio but also obscure the true features of the data, thereby affecting the accuracy of subsequent analysis. Therefore, a radial basis function (RBF) neural network was used to extract features from the collected three-phase line loss data. By performing nonlinear mapping of the input data, the method effectively suppressed the interference from noise, enhancing the signal-to-noise ratio. The preprocessed data were then normalized, which further improved the completeness and accuracy of the data collection. A loop current-based method was employed to decompose the circuits in the distribution network into multiple independent loops. In each loop, the real and imaginary parts of the voltage and current were calculated. By analyzing the temporal and phase variations of these values in detail, the operating status of the circuit was thoroughly understood, and potential anomaly patterns were accurately identified. Based on the real and imaginary part values of the voltage and current on the three-phase branch circuits, a Gaussian mixture distribution model was constructed. This model used multiple Gaussian distributions to describe the complex distribution features of the three-phase line loss data, allowing for more accurate capture of the inherent patterns and structures in the data. The maximum expectation algorithm was then used to fit the normalized line loss rate and construct a hybrid Gaussian model consisting of multiple Gaussian mixture distributions. The likelihood probability function of the eigenvector was calculated, and based on a preset probability threshold, the data were determined to be anomalous or normal. If the likelihood probability was below the threshold, it was classified as anomalous; otherwise, it was considered normal. This approach enabled the identification of line loss anomaly data. [Results] Experimental results show that the proposed method can accurately identify three-phase line loss buses, reducing the risk of misjudgment and missed detections. [Conclusions] This method can promptly detect and address faults in the distribution network, which is of significant importance in improving the operational efficiency and reliability of power systems.

[Objective] To assess the security of unmanned aerial vehicle (UAV) communication environments in power applications, an effective security architecture and design scheme was proposed to address electromagnetic interference, data security, and other challenges UAV communications face during power inspections, ensuring efficient collaborative operations and secure data transmission of UAVs in complex electromagnetic environments. [Methods] Firstly, the structure of UAV collaborative wireless networks and security threats they face were analyzed, especially the influence of strong electromagnetic interference near power lines on UAV communications. Subsequently, an identity cryptography-based solution to UAV collaborative wireless networks was proposed. Additionally, by designing enhanced communication protocols and anti-interference mechanisms, stable transmission of critical data under strong electromagnetic interference was ensured. Meanwhile, the mutual authentication, signature, and identity verification mechanisms for communication data were introduced to enhance the overall security of UAV communications. [Results] Experimental results demonstrate that the proposed security assessment architecture and design scheme exhibits high data recovery rates and low resource consumption under varying numbers of UAVs, communication failure rates, electromagnetic interference intensities, and data packet sizes. In particular, the system maintains high data recovery rates and fault tolerance capabilities even under high electromagnetic interference, effectively resisting potential network intrusions and data tampering threats. [Conclusions] The proposed security assessment architecture and design scheme significantly enhances the security and reliability of UAV communications during power inspections, reducing the information leakage risk and enabling efficient collaborative operations in complex electromagnetic environments. The innovation of this study lies in combining identity cryptography and public key mechanisms to design a lightweight, efficient, and safe solution, providing effective security guarantees for UAV communication networks in power inspection tasks.

[Objective] High-voltage circuit breakers are key control and protection devices in the power system, and their reliable operation is crucial for the safety and stability of power grids. However, during long-term operation, high-voltage circuit breakers may trigger various faults due to mechanical wear, component aging, and other problems. Currently, the detection of high-voltage circuit breakers faces challenges such as diverse detection signals, great difficulty in fault detection, and low accuracy. Therefore, studying an efficient and accurate mechanical status detection method for high-voltage circuit breakers is of great significance for ensuring the safe and stable operation of the power system. [Methods] This study proposed a mechanical status detection model for high-voltage circuit breakers based on multi-modal efficient Transformer. In the data acquisition stage, vibration sensors, current sensors, and displacement sensors were comprehensively employed to synchronously acquire vibration signals, current signals, and displacement signals during the operation of high-voltage circuit breakers, thus constructing a multi-modal signal dataset. In the signal preprocessing stage, wavelet transform technology was adopted to process the acquired multi-modal signals and decompose the signals into different frequency scales, thus effectively removing noise components in the signals, enhancing fault feature signals, and significantly improving signal quality. In terms of model building, an efficient Transformer module was introduced. With its powerful self-attention mechanism, the module could effectively capture long-distance dependency relationships in signal sequences and dig deeply into complex features in multi-modal signals. Additionally, by classifying the operation status of high-voltage circuit breakers into six categories, including normal operation, failure to maintain closing, loose soft connection, single-phase contact wear, loose insulating tie rod, and opening spring fracture, accurate diagnosis of the mechanical status of circuit breakers was realized. [Results] In the simulation experiments, simulation models of different fault types of high-voltage circuit breakers were built to simulate various working conditions during actual operation and generate multi-modal signal data. Inputting the data into the proposed detection model for testing shows that the model can accurately identify different fault types. In the actual experiments, multiple high-voltage circuit breakers were selected as test objects, and multi-modal signal data were collected under their normal operation and different fault settings. The experimental results reveal that the proposed method significantly improves the detection accuracy compared with traditional detection methods while ensuring the detection speed. [Conclusions] The proposed mechanical status detection model for high-voltage circuit breakers based on multi-modal efficient Transformer effectively solves the problems of complex detection signals and severe noise interference. By leveraging the powerful feature extraction and classification capabilities of the efficient Transformer model, accurate identification of multiple mechanical faults in high-voltage circuit breakers is realized. Simulation analysis and experimental results fully demonstrate that this method performs well in both detection accuracy and speed, providing reliable technical support for the status monitoring and fault diagnosis of high-voltage circuit breakers in the power system. It helps to timely detect potential faults of the equipment, and holds application significance and broad promotion prospects for ensuring the safe and stable operation of the power system.

[Objective] With the increasing importance of power transmission and transformation projects in distribution networks, traditional cost estimation methods face challenges such as large errors and excessive time consumption, making them inadequate for modern engineering management. To enhance the execution efficiency and estimation accuracy of power transmission and transformation projects, this study proposed a novel cost estimation method based on radial basis function neural network (RBFNN) and significance cost theory. This approach aims to address the limitations of traditional methods in complex cost estimation scenarios while enhancing the robustness and adaptability of the model. [Methods] This study applied significance cost theory to screen historical project data and identify the main factors affecting cost estimation for power transmission and transformation projects. These factors were used as input features for the neural network. The method introduced radial basis functions (RBFs) to restructure the traditional artificial neural network (ANN) architecture, creating a cost estimation model specifically for power transmission and substation projects. The model processed input data using Gaussian functions, initialized the hidden layer centers with the K-means clustering algorithm, and used least squares and gradient descent methods to train the output and hidden layers. To validate the model′s effectiveness, 100 sets of data from power transmission and transformation projects were used to compare the cumulative absolute error rate and average execution time of traditional methods (unit cost method and index estimation method) with the proposed model. Moreover, SHAP value analysis was employed to quantify the impact of key factors on estimation error rates. [Results] Simulation results demonstrate that the RBFNN-based cost estimation method outperforms traditional methods in both cumulative absolute error rate and execution time. When the test sample size increases to 20, the cumulative error rate for the unit cost method reaches 440％, while the index estimation method reaches 180％, and the proposed model maintains an error rate below 110％. In terms of execution time, traditional methods require an average of 5 s, while the proposed model reduces the time to just 0.5 s. In addition, SHAP value analysis reveals that factors such as wire cross-sectional area, steel pipe poles, and the number of circuits have the greatest influence on estimation error rates, with their SHAP values significantly higher than those of other factors. This finding provides critical insights for model optimization and cost control. [Conclusions] The cost estimation method proposed in this paper, based on RBFNN and significance cost theory, effectively improves the accuracy and efficiency of cost estimation for power transmission and transformation projects. Although the method still exhibits some errors in complex construction environments, it outperforms traditional methods in overall performance, making it highly practical with strong potential for widespread application. Future research will focus on integrating regression analysis, support vector machine (SVM), and other machine learning algorithms to further optimize model precision and better handle the complexity and variability of power transmission and transformation projects.

[Objective] With the increasing proportion of new energy, traditional grid-following (GFL) control based on phase-locked loop (PLL) synchronization gradually exhibits inherent stability limitations in weak grid conditions. Meanwhile, grid-forming (GFM) control with self-synchronizing source characteristics has emerged as a hot solution. However, existing research predominantly focuses on the voltage regulation or synchronization stability of GFM control, with less attention to its frequency modulation capability and characteristics. [Methods] This paper systematically reviewed four mainstream GFM control methods, including droop control, virtual synchronous generator (VSG) control, matching control, and virtual oscillator control (VOC), explained their frequency modulation principles, and analyzed their advantages and disadvantages from the aspects of the control loop and application scenarios. On this basis, a grid-connected simulation model for new energy systems was built to conduct a simulation-based analysis of the frequency modulation response characteristics of different kinds of frequency modulation control across diverse scenarios. Finally, this study summarized challenges of GFM control in strategy optimization, parameter tuning, and multi-unit coordination, with the future development prospects pointed out. [Results] Droop control regulates the active power of generating units by responding to system frequency deviations, featuring advantages of the simple structure and strong grid strength adaptability. However, its lack of inertia support results in relatively weaker frequency modulation performance. On the basis of droop control, VSG control simulates the inertia response characteristics of conventional synchronous machines and can better suppress the change performance of system frequency. However, it faces challenges in parameter tuning, fault ride-through, and multi-unit coordination. Matching control utilizes the dynamic characteristics of DC capacitors to simulate the inertia properties of traditional synchronous machines and thus restrain change performance of system frequency, but it fails to provide sustained support in the frequency quasi-steady state. VOC generates frequency responses similar to droop control via oscillator dynamic equations that directly govern amplitude and frequency. However, it is difficult for its high output harmonics to satisfy grid connection requirements. [Conclusion] Virtual synchronous machine control has become the most promising research direction in GFM control due to its technical advantages of balancing frequency modulation performance and strong grid strength adaptability in participating in system frequency modulation. However, technical challenges including synchronization stability, fault ride-through, and coordinated control need to be tackled. In the future, in-depth research should be conducted on control strategies and parameter optimization, and multi-unit collaborated control to facilitate the large-scale application of GFM control.

[Objective] The cost of distribution network engineering is influenced by multidimensional factors such as scale and capacity, equipment and material costs, and geographical conditions. Traditional statistical methods (e.g., linear regression) struggle to handle high-dimensional nonlinear data effectively, while existing machine learning approaches, despite incorporating feature reduction techniques, still exhibit limitations. For instance, principal component analysis (PCA) sacrifices prediction accuracy for dimensionality reduction, and grey relational analysis (GRA) ignores feature interactions. Therefore, there is an urgent need for a prediction method that retains critical feature information while accounting for complex inter-feature relationships. This study integrated recursive feature elimination (RFE) with the random forest (RF) algorithm to develop a RFE-RF prediction model, aiming to resolve feature redundancy and nonlinear modeling challenges. [Methods] A technical framework of “feature selection-model construction-experimental validation” was adopted. For feature selection, the recursive feature elimination (RFE) method was employed, which iterated training models to gradually eliminate features with minimal predictive contributions, retaining an optimal feature subset. For model construction, the RF algorithm was utilized. Based on ensemble learning principles, RF constructed multiple decision trees and averaged their outputs, effectively mitigating overfitting and enhancing model robustness. RF was insensitive to noisy data and quantified feature importance, providing reliable feature ranking criteria for RFE. By embedding RFE into the RF training process, a closed-loop optimization workflow was established. [Results] Experimental validation used data from 190 distribution network engineering projects provided by a power grid company, covering 21 initial features such as voltage level, line length, and equipment costs. Categorical features were numerically encoded while preserving their original distribution characteristics. Through five-fold cross-validation and root mean square error (RMSE) optimization, the optimal feature subset was identified as 12 optimal feature subsets, including such key factors as line length, comprehensive cable price, and voltage level. Compared with traditional linear regression (LR), RF, and mutual information-based RF (MI-RF) algorithms, the RFE-RF algorithm achieves a mean absolute error (MAE) of 8.6579 and a mean absolute percentage error (MAPE) of 6.97％ on the test set, significantly outperforming other algorithms. The MAE of RFE-RF on the test set increases by only about 4.5％ compared to the training set, indicating lower overfitting risks and demonstrating that feature selection effectively enhances model stability. [Conclusion] Feature selection is pivotal for improving the accuracy of distribution network cost prediction. RFE dynamically eliminates redundant features through iterative processes, substantially reducing data dimensionality and noise interference. The RFE-RF model combines high precision with strong interpretability, reduces MAE significantly compared to traditional models, and clearly quantifies the impact weights of individual features on costs. This study marks the application of combining RFE and RF in cost prediction for distribution network engineering, addressing challenges in feature interaction and redundancy filtering and providing a new paradigm for data modeling in complex engineering systems. The model serves as a precise cost prediction tool for power grid enterprises, aiding investment decisions and cost control, thus advancing intelligent and refined construction of distribution networks. Moreover, it reveals the impact mechanism of feature selection on the generalization capability of machine learning models, offering practical references for feature optimization in high-dimensional nonlinear datasets.

[Objective] With the global energy transition and the rapid development of clean energy, the penetration rate of high-penetration photovoltaic (PV) sources in distribution networks is increasing. However, PV output power exhibits significant fluctuations and uncertainty due to factors such as solar irradiance and temperature. When a large number of such sources are integrated into distribution networks, they can cause voltage fluctuations, frequency variations, and other issues, presenting significant challenges for power outage fault prediction. Traditional fault prediction methods struggle to accurately capture fault characteristics in complex distribution networks with high PV penetration, leading to reduced prediction accuracy and efficiency, which fails to meet the stability requirements for distribution network operation. [Methods] To improve prediction accuracy and efficiency, this study proposed a fault prediction method for distribution networks with high PV penetration. First, a PV-integrated grid model was built to analyze the impact of PV sources on fault current characteristics in distribution networks. This model clarified how PV sources influence fault current magnitude and distribution under different operating conditions, providing a theoretical basis for subsequent fault zone identification. Next, potential outage zones were inferred by combining grid topology and load imbalance features. The grid topology reflected the connectivity of components, while load imbalance indicated regional load variations. By integrating these factors, the method more accurately localized the fault zone. In addition, power flow entropy was introduced to assess whether circuit loads were in a critical state. Key fault-related power flow features were then extracted from the identified zones. These features were fed into an optimized SA-SAE for training, allowing the system to automatically learn underlying patterns from large datasets and achieve precise outage prediction. [Results] Experimental results demonstrate that the proposed method achieves high prediction accuracy in fault localization for distribution networks with high PV penetration, correctly identifying fault zones (sections 3-6 of the K5-K8 lines) and fault types. Moreover, the average prediction time is only 2.236 seconds, significantly outperforming comparative methods in both accuracy and efficiency. [Conclusion] By comprehensively considering PV integration effects, grid topology, load characteristics, and leveraging power flow entropy and SA-SAE, the proposed method enables high-precision and high-efficiency outage prediction in distribution networks. This method not only enhances prediction accuracy and timeliness, reducing outage risks and economic losses, but also provides robust support for grid planning, operation, and maintenance. It ensures stable distribution network operation and facilitates large-scale integration of clean energy.

[Objective] With the increasing demand for application of unmanned aerial vehicles (UAVs) in complex scenarios such as power grid inspection and emergency rescue, the limitations of single UAV in task execution have become increasingly prominent. Multi-UAV formation can effectively improve inspection efficiency and expand operation coverage, but significant challenges remain in formation maintenance, collaborative trajectory optimization, and environmental adaptability to complex environments during practical application. An optimal control method that integrated virtual spring forces with the hp-adaptive pseudospectral method was proposed to address the difficulties of formation maintenance and path optimization during large planar maneuvers of UAV swarms, thus enhancing the stability, flexibility, and disturbance resistance of collaborative flight of UAV formations and providing technical support for high-demand scenarios for UAVs such as power grid inspection. [Methods] First, a multi-UAV system dynamics model was built, and a virtual spring mechanism was incorporated into the formation control system to realize flexible constraints and elastic self-adjustment between UAVs. By combining the virtual spring method with the traditional leader-follower method, a formation strategy that could balance rigid support and adaptive adjustment ability of formations was designed. On this basis, the hp-adaptive pseudospectral method was then applied to solve the optimal control problem of UAV formations. By discretizing state and control variables at Legendre-Gauss nodes and constructing global interpolation polynomials, the trajectory optimization problem was transformed into a nonlinear programming (NLP) problem, with constraints such as dynamics, energy consumption, and velocity combined to conduct a high-precision numerical solution. In simulation experiments, a typical four-UAV diamond formation was set up, and the algorithm′s adaptability to different terrains, wind disturbances, and mission requirements was comprehensively explored. [Results] Simulation results show that the proposed virtual spring-based hp-adaptive pseudospectral method can realize smooth formation turning and velocity control. During a 90° large maneuver, UAVs can not only satisfy multiple constraints such as path deflection and speed change, but also maintain a stable formation. Compared with traditional leader-follower and artificial potential field methods, the new method demonstrates significant advantages in position error, formation maintenance, and wind resistance. Under 10m/s strong wind, the formation stability of the proposed method exceeds 70％, showing significant advantages over its competing algorithms. 3D terrain simulations and real flight tests further validate the algorithm′s adaptability and robustness, and the method still maintains lower formation deformation rates and trajectory tracking error under multiple terrains such as hills, mountains, and canyons, with the features of reasonable energy consumption control and strong engineering practicability. [Conclusion] By innovatively integrating the virtual spring elastic constraint with the hp-adaptive pseudospectral method, an optimal control technique for UAV formation trajectory planning in complex environments was proposed. The rigidity constraint limitations of traditional formation methods are overcome, flexible maintenance and adaptive adjustment of formations are realized, and the accuracy and efficiency of collaborative trajectory optimization are significantly improved by the method. The research results provide an efficient and reliable technical path for collaborative flight of UAV swarms in demanding tasks such as power grid inspection and emergency rescue. Future studies may further increase the method′s application potential in multi-formation collaboration and complex obstacle environments, promoting the intelligent and practical development of UAV formations.

[Objective] Interior permanent magnet (IPM) motors are widely adopted as submersible motors in oil well applications due to their structural stability, high efficiency, and superior power factor. However, the constrained wellbore diameter and elevated ambient temperatures in deep-well environments impose stringent requirements for enhanced torque density and anti-demagnetization capability of IPM motors. To address these challenges, a novel sine-shaped permanent magnet (PM), synthesized from flat and arched PM, was adopted in this study. The adopted PM optimizes rotor space utilization above conventional flat PM, thereby increasing permanent magnet volume and d-axis permanent magnet thickness, improving torque density and anti-demagnetization capability of IPM motors. [Methods] Maintaining constant dimensions of the flat PM, the finite element analysis (FEA) was employed to systematically evaluate the effects of arched PM sagitta on short-circuit current, anti-demagnetization capability, and no-load and on-load electromagnetic performance. Moreover, the equivalent ring method was implemented to quantify the maximum stress of the rotor core under various sagittas, ensuring the mechanical integrity of the optimized rotor structure. [Results] Although the short-circuit current of the IPM motor will increase as the sagitta increases, its permanent magnets exhibit stronger anti-demagnetization capability. While increasing the sagitta will increase the maximum stress of the rotor core, it is much lower than the yield stress of the core material, meeting practical needs sufficiently. Moreover, the effect of a smaller sagitta on the reluctance torque of the IPM motor can be ignored. When the sagitta is greater than 3 mm, the maximum reluctance torque decreases significantly as the sagitta increases, while the total output torque keeps increasing and torque ripple decreases. [Conclusion] Taking into account the overall influence of the sagitta on the motor′s performance, a suitable sine-shaped PM size was selected, and a prototype was manufactured and tested. The experimental results are in good agreement with the 2D FEA simulation results, verifying the accuracy of the simulation analysis, which provides a new approach for the rotor design of the IPM motors.

[Objective] Conducting a state evaluation of secondary equipment in power systems is crucial for mitigating the operational risks of the power grid and improving grid reliability. To address the logical issues in the analytic hierarchy process (AHP) and the limitations of subjective judgment in the entropy weight method, this study proposed an improved entropy weight method for assessing the health status of secondary equipment. [Methods] Based on the fundamental characteristics of secondary equipment in the power system, both technical and management indicators for the state evaluation were developed. The study utilized forward, reverse, and trapezoidal mapping relationships to standardize the indicator parameters. In addition, a membership function based on normal distribution was adopted. This function retained valid information from high membership intervals, incorporating information from low membership ranges, and avoided misjudgment caused by an overemphasis on the low membership range. The weight principle of AHP was used to establish the judgment matrix of different evaluation indicators, and the coefficient of variation in the entropy weight method was introduced based on the arithmetic mean and standard deviation of each indicator, enabling an objective representation of the evaluation indicator weight. Subsequently, a comprehensive model for the state evaluation of secondary equipment was established. The model was validated using 36 protection devices in a substation. [Results] The verification results demonstrate that the evaluation results aligns with the actual operation status of the protection device. The membership value for “good” is 0.890 1, and for “fair” it is 0.097 9, which allows for the determination that the 220 kV main transformer protection device operates normally and consistents with the actual operational state of the protection device. The range of the membership function for evaluation indices obtained using AHP is 0.321 0, while the entropy weight method yields a range of 0.341 4, which may lead to misjudgments. Longitudinal comparisons of the AHP-entropy weight method and the entropy method′s weighting algorithms show that the membership degree ranges are 0.125 0 and 0.184 9, with minimum values of 0.806 5 and 0.708 8, respectively, with no misjudgments. In this method, the difference in the range values of the membership degree is 0.048 1, with maximum and minimum values of 0.900 0 and 0.851 9, resulting in a narrower fluctuation range and higher judgment reliability. [Conclusion] The innovation of this study lies in the combination of AHP and the entropy weight method, which effectively avoids the interference of human factors in subjective weighting. Moreover, by incorporating objective factors into equipment evaluation and introducing an entropy weight calculation method based on the coefficient of variation, the weight calculation accurately reflects the equipment′s actual operation status. The actual calculation results show that the proposed method more effectively reflects the actual state of equipment operation and provides crucial support for equipment operation and maintenance.

[Objective] With the continuous expansion of the power grid scale and the increasing complexity of its structure, traditional power grid modeling and visualization methods have gradually revealed many issues. For example, modeling accuracy often fails to meet the refined display requirements of complex grid structures. Application scenarios are limited and unable to effectively address diverse business needs. Moreover, there is a lack of scientific and effective verification mechanisms to ensure the accuracy and reliability of modeling results. To address these challenges, this study proposed a 3D power grid modeling and verification method integrating GIS-GIM and DETR networks, to achieve high-precision 3D grid modeling and establish an effective verification system, providing a solid data foundation and reliable decision support for grid planning, operation and maintenance, and management. [Methods] The first step involved integrating the grid information model (GIM) into the geographic information system (GIS). By leveraging GIS′s powerful geospatial analysis and display capabilities, and combining GIM′s detailed descriptions of grid equipment and topological structures, a more comprehensive 3D grid modeling approach was achieved, visually presenting the grid′s overall layout and equipment distribution from a geospatial perspective. Second, the DETR network was improved by optimizing its structure, adjusting parameter settings, and employing more effective training strategies, enabling it to more accurately detect and classify 3D grid equipment. During training, a large volume of 3D grid equipment data was collected to build a rich and diverse dataset. The data were then annotated and preprocessed to improve the model′s generalization ability. Last, the improved DETR network was applied to the 3D grid modeling process to detect and classify equipment in the modeling results individually, ensuring the accuracy of equipment information and the overall accuracy of the modeling results. [Results] To validate the effectiveness of the proposed method, experimental analyses were conducted on 100 sets of equipment data from three newly built substations. The results show that, compared to traditional modeling methods, the proposed 3D grid modeling method that integrates GIS-GIM and DETR networks significantly improves modeling accuracy, enabling more precise restoration of the spatial positions, structural forms of grid equipment, and connection relationships between equipment. Regarding the verification of the modeling results, the verification network demonstrates a good performance with an accuracy rate of 93.14％, indicating that the method can effectively detect potential errors and deviations in the modeling process and ensure the reliability of modeling results. [Conclusion] The proposed 3D power grid modeling and verification method, integrating GIS-GIM and DETR networks, performs excellently in improving grid modeling accuracy and establishing an effective verification mechanism, meeting the high-precision requirements of actual grid modeling. The method contributes significantly to improving the scientific basis for grid planning and provides intuitive 3D visualization for daily operations, maintenance, fault diagnosis, and repair, supporting reliable decision-making in grid management. It holds important theoretical significance and broad application prospects.

[Objective] Substations, as the core hubs of power transmission and distribution, play a crucial role in ensuring the safe and stable operation of power systems, which is essential for efficient and reliable power supply. However, traditional substation monitoring methods face challenges such as limited automatic monitoring capabilities and inadequate target detection accuracy, making it difficult to meet the increasing safety demands of modern power systems. This study aimed to develop substation target recognition and safety monitoring technology based on the regional fully convolutional network (R-FCN) to overcome the shortcomings of traditional monitoring methods, significantly enhance substation safety assurance, and establish a solid foundation for the stable operation of the power system. [Methods] This method combined the unique advantages of region extraction and fully convolutional networks to construct an efficient and intelligent monitoring system. High-definition video surveillance cameras were deployed in the data collection phase to continuously capture real-time image data of the substation from multiple angles, providing massive and precise raw data for subsequent in-depth analysis. The advanced R-FCN model was applied for object detection based on the collected images. Due to its fully convolutional nature, R-FCN could effectively maintain high-resolution feature maps when processing images of different sizes, avoiding the information loss commonly encountered during downsampling with traditional methods, thus significantly improving object detection accuracy. A specially designed area extraction module, resembling an intelligent navigation system, accurately located key facilities such as transformers, switchgear, and insulators within the complex substation environment, ensuring real-time and precise monitoring of equipment status. Moreover, abnormal behaviors, such as unauthorized personnel entering hazardous areas or equipment suddenly emitting smoke or catching fire, were detected promptly, allowing for valuable time to be saved for emergency responses. [Results] Extensive simulation experiments and practical testing in real substation monitoring scenarios demonstrate the system′s excellent performance. In comparison with traditional object detection methods, the system significantly improves detection accuracy, enhances monitoring reliability, and reduces unnecessary manpower and resource expenditure. [Conclusion] The R-FCN-based substation target recognition and safety monitoring technology combines efficient real-time processing and precise target positioning capabilities. When handling massive monitoring data, it can quickly and accurately identify various targets and abnormal situations, providing robust technical support for the safe and stable operation of the power system. This technology has profound implications for enhancing substation monitoring levels and ensuring the reliable power supply of the power system.

[Objective] Due to the long-term exposure of overhead transmission lines to the natural environment and the significant impact of environmental factors, timely monitoring of their operating status plays a key role in the safe operation of the power grid. With the development of UAV flight control technology and the widespread use of detection technologies such as infrared, ultraviolet, and LiDAR, these methods are increasingly used in the inspection of power transmission lines. However, traditional methods currently only show optimal results in single-scenario line inspection. In more complex environments, such as mixed transmission line inspections, it is challenging to quickly and accurately analyze transmission line inspection data. Therefore, this study proposed an optimization method for point cloud data processing in hybrid transmission line inspection. [Methods] First, a transmission line inspection point cloud data processing platform was constructed. LiDAR mounted on the UAV platform collected the mixed point cloud data of the transmission line and processed it in four stages: data management, preprocessing, classification, and intelligent inspection. The mixed point cloud data were thinned using the octree method to reduce redundant data and ensure the accuracy and quality of the data. Finally, a neural network model was designed to optimize the sparse data, consisting of three main parts: the feature learning layer, the convolutional layer, and the classification layer. The feature learning layer avoided the impact of the disorder in 3D point cloud data on feature extraction through multiple projections and maximum pooling. The convolutional layer extracted common features from voxel grids and surrounding entities while incorporating traditional transmission line feature extraction algorithms to extract voxel grid features. The classification layer included a fully connected layer with a ReLU activation function, using the Softmax model as the classification function to obtain the classification results of the mixed point cloud data. [Results] In the experiment, the LDLRS3100 LiDAR was selected to collect point cloud data of a transmission line channel in a certain area. The UAV LiDAR system has a range of 360 m, a flight speed of 20 km/h, and a flight altitude of 150 m. The proposed method was analyzed based on the Pytorch platform, and the results show that it can effectively identify the differences between transmission lines and ground objects, and obtain clear information on the tower and its surrounding environment. The overall accuracy reaches 92.71％, which is significantly better than other comparative methods. To balance the highest sparsity rate and the best visual effect of the point cloud data, the sparsity density is set to 0.02 m. [Conclusion] By optimizing the point cloud data of mixed transmission line inspections using the octree sparsity method and a neural network model, various types of point cloud data can be quickly and accurately classified, thus improving the reliability of intelligent transmission line inspections.

[Objective] Traditional motor optimization design methods involve establishing analytical models for motor volume, loss, and cost, selecting optimization algorithms to refine them, and deriving optimal design variables. However, since motor models are complex, analytical models fail to precisely describe partial variables. Stator magnetic density is an important variable of slotless permanent magnet direct-current motors, whereas the accuracy of its analytical formula is low. The particle swarm algorithm is widely used in optimization design, but its optimization ability is poor. [Methods] To solve the above problems, an optimization design method of the slotless permanent magnet direct-current motors based on adaptive improved particle swarm algorithm was proposed. An analytical model of the slotless permanent magnet direct-current motors was established, and an objective function was constructed with motor volume, loss, and cost as optimization goals. The Sobol method was employed to identify high-sensitivity variables of the motors, thereby reducing the number of design variables. Subsequently, a magnetic circuit model was developed using finite element simulations, and magnetic density data were extracted after design variable parameters were adjusted. The response surface method was then applied to re-fit the magnetic density data, and a stator magnetic density response surface model was established to replace the original analytical formula. The particle swarm algorithm was improved. The updating modes of inertial weight and learning factor were selected through comparisons between fitness values of individual particles and the average fitness value of global particles during iteration, which enhanced algorithmic precision. Finally, both the original and improved algorithms were utilized to optimize the objective function. The optimal motor design parameters were achieved by comparison. [Results] Comparative analysis of stator magnetic density calculations between the analytical formula and the response surface model reveals that the latter exhibits significantly reduced computational errors. When the adaptive particle swarm algorithm, original particle swarm algorithm, and other classical algorithms were applied to optimize the objective function, the improved particle swarm algorithm achieves the most optimal results. [Conclusion] The experimental results demonstrate that replacing the analytical formula for stator magnetic density with a response surface model effectively mitigates the significant calculation errors associated with the analytical approach. Meanwhile, the particle swarm algorithm incorporating adaptive updates of inertia weight and learning factor exhibits an enhanced optimization capability. Comparative analysis with classical algorithms confirms its superior optimization capability.

[Objective] In an alternating current (AC) power system, the distributed control system (DCS) controller serves as a core component responsible for real-time acquisition and processing of various critical data, which is vital for the stable operation and fault prediction of the system. However, in practical applications, the data acquisition process of DCS controllers often encounters issues such as data loss or anomalies due to external electromagnetic interference, hardware failures, and other factors, which makes it difficult to determine data density and thus affects system reliability and accuracy. In view of this, a high-speed multi-channel synchronous sampling method for DCS controller data in AC power systems was proposed to address interference and data missing during data acquisition, thereby enhancing data quality and system performance. [Methods] The signal conditioning circuits preprocessed analog data signals from different channels to ensure that the signal quality met the analog-digital converter (ADC) conversion requirements. Field-programmable gate array (FPGA), serving as the control center, leveraged its parallel processing capabilities and programming flexibility to precisely control the ADC conversion process for each channel, achieving high-precision, low-latency synchronous sampling and effectively addressing the issues of phase deviation and data inconsistency caused by asynchronous sampling. For data missing, the Clearbout theory was adopted for data interpolation, intelligently estimating and filling missing data based on the time-frequency characteristics of the signal and the correlation of known data points and thereby ensuring data continuity and integrity. Additionally, the synchronous sampling algorithm was optimized using the ant colony algorithm, which dynamically adjusted sampling parameters by simulating the pheromone update mechanism of ants searching for food to enhance sampling efficiency and accuracy. [Results] Experimental results demonstrate that the proposed multi-channel synchronous sampling method significantly enhances the data acquisition performance of DCS controllers. The frequency spectrum diagram of the acquired DCS data is highly consistent with the actual data frequency spectrum diagram, which verifies the accuracy and reliability of the sampling method. The sampling speed is significantly increased, meeting the high real-time requirements of AC power systems. [Conclusion] In summary, the proposed method incorporates FPGA control to achieve high-precision, low-latency multi-channel synchronous sampling, solving phase deviation and data inconsistency issues. Introducing the Clearbout theory and ant colony algorithm effectively guarantees data integrity and optimizes the sampling algorithm. The designed multi-channel data upload mechanism avoids conflicts during data upload, ensuring smooth data transmission. These innovations not only improve the data acquisition capability of DCS controllers in AC power systems but also provide useful reference for the design and optimization of similar systems. Therefore, the application of the proposed method helps enhance the stability and reliability of entire AC power systems, reduces the risk of system failures caused by data anomalies, and is of great significance for ensuring the safe operation of power systems.

[Objective] Agricultural parks, characterized by abundant renewable energy resources, play a pivotal role in advancing green and low-carbon transformation under the carbon peaking and carbon neutrality goals. However, current agricultural parks face challenges such as low energy utilization efficiency, imbalanced multi-energy distribution, and insufficient local renewable energy accommodation capacity, which hinder agricultural productivity and sustainable development. To address these issues, this study proposed a deep learning-based optimization method for constructing a more economical and low-carbon integrated energy system (IES) in agricultural parks. [Methods] First, a multi-objective optimal scheduling model for agricultural park IES was established, integrating economic objectives such as fuel costs of gas turbines, grid interaction costs, and equipment maintenance costs, while formulating mathematical constraints for multi-energy coupling systems. Second, an improved long short-term memory (LSTM) neural network was employed to predict photovoltaic/wind power outputs and load demands. The hyperparameters of the LSTM model, including hidden layer units and learning rates, were dynamically optimized using quantum particle swarm optimization (QPSO) to enhance prediction accuracy. Finally, to mitigate premature convergence in the traditional golden sine algorithm (GSA), an enhanced GSA algorithm was proposed by incorporating Lévy flight strategies to expand the search space and designing dynamic weight mechanisms to balance global exploration and local exploitation capabilities. [Results] Case studies demonstrate that the errors of improved QPSO-LSTM prediction model are controlled within 5％, outperforming traditional optimization algorithms in accuracy and robustness against local optima. For scheduling optimization, the enhanced GSA algorithm achieves a 69.7％ reduction in daily operational costs and a 27.9％ improvement in local renewable energy accommodation rates compared to unscheduled scenarios, significantly surpassing conventional GSA and other methods. These results validate the algorithm′s effectiveness in balancing economic efficiency and low-carbon requirements for multi-energy coordination. [Conclusion] The proposed deep learning-based optimization framework enables high-precision power prediction and cost-effective scheduling for agricultural park IES. It significantly reduces operational costs while enhancing renewable energy utilization, demonstrating superior performance in synergizing economic and low-carbon objectives. This study provides a reliable technical pathway for the efficient and sustainable operation of agricultural park IES.

[Objective] To address the problems of traditional load forecasting methods, such as low information utilization efficiency, large errors, and difficulties in adapting to the diversity and randomness of actual power load changes, a power system load forecasting method based on land spatial information perception was proposed. This method could improve the accuracy and reliability of load forecasting, provide key data support for power system planning and construction, and meet the needs of economic and social development. [Methods] This algorithm adopted the strategy of classification by area with the use of urban land spatial information and power grid load data. For developed areas (with loads known), historical load data were used for curve fitting to carry out load forecasting. As for newly developed areas (with loads unknown), the average load density of the same land type in developed areas was used for equivalentprocessing to form the basic information of load forecasting, and then load forecasting was carried out. At the same time, this algorithm subdivided the historical load data and equivalent load data. The algorithm integratedthe exponential model, the growth curve model, and the elastic coefficient model and mainly used the combination of dynamic weights to form the best fitting scheme. Using the above methods, a power system load forecasting technology based on land spatial information perception was formed. The historical data from 2014 to 2020 were used as the benchmark for parameter fitting, and the historical load data of industrial power, residential power, commercial power, public facilities power, and other power types were sorted out. The load data of 2021 were taken as the forecasting object. The differences of the exponential model, the growth curve model, and the elastic coefficient model in total forecast and forecast based on classification by area were compared and analyzed through experiments. The results show that the forecasting accuracy based on classification by area is about 33% higher than that of total forecasting. On this basis, the best fitting effects based on dynamic weights and mean weights were compared and analyzed. The calculation results show that the best fitting forecasting error is only 1.12% when dynamic weights are used, which is 12% smaller than the forecasting error in the case of using mean weights. In conclusion, the scheme proposed can significantly improve the accuracy and reliability of load forecasting. [Results] The results of this paper show that the accuracy of the load forecasting model can be improved by using the method of classification by area to classify spatial information according to the land and load types. The algorithm has better dynamic adaptability by dynamically adjusting parameter weights and integrating single forecasting models, able to achieve the best fitting results and enhance forecasting accuracy. [Conclusion] The highlights of this paper are as follows. Firstly, the data processing method of classification by area is adopted to improve the utilization rate of spatial information and load information of the urban power grid. Secondly, the traditional load forecasting models are integrated by using dynamic weights, which breaks the limitations of single models. This algorithm further improves the accuracy and reliability of urban power grid load forecasting through the above two approaches.

[Objective] Power engineering projects have long construction cycles and are affected by various uncertainties, which may lead to significant cost increase. Therefore, accurately estimating contingency costs is crucial for project management. However, traditional experience-based methods have large errors and are difficult to adapt to complex engineering environments. Existing methods based on fuzzy expert systems and machine learning have improved performance but still face challenges such as parameter optimization difficulties and severe overfitting. To address these issues, a new contingency cost estimation method was proposed. It leverages the advantages of adaptive network-based fuzzy inference system (ANFIS) in handling uncertainty problems and introduces a principal component analysis (PCA) module to mitigate overfitting and improve prediction accuracy. [Methods] A contingency cost estimation method integrating risk analysis and ANFIS was proposed. First, the relationship between contingency costs and 13 key risk factors affecting power engineering costs was modeled. Then, ANFIS was introduced to utilize fuzzy logic for handling uncertainty problems. ANFIS processed input variables through fuzzification and leveraged neural networks for inference, avoiding the dependency on fuzzy rule sets in traditional fuzzy expert systems. To further enhance prediction accuracy, a PCA module was incorporated into ANFIS, which eliminated redundant information through dimensionality reduction and alleviated overfitting issues associated with small datasets. [Results] In the experiments, 210 power engineering contingency cost data samples were selected, with 80% randomly assigned to the training set and 20% to the test set. The performance of four methods was compared:a Mamdani fuzzy inference based method, a support vector machine (SVM) based method, an ANFIS based method, and an improved ANFIS based method. The experimental results indicate that while the ANFIS based method outperforms the existing two methods on the training set, it suffers from severe overfitting on the test set. After incorporating the PCA module, the improved ANFIS based method achieves significantly better test performance, demonstrating stronger generalization ability and faster convergence. [Conclusion] The proposed contingency cost estimation method based on the improved ANFIS combines the advantages of fuzzy inference and neural networks, enhancing prediction accuracy for power engineering contingency costs. The key innovations of this study are as follows. A contingency cost estimation method was proposed, effectively addressing the uncertainty problem by combining the advantages of fuzzy logic and neural networks. In addition, a PCA module is introduced into ANFIS, which reduces redundant information through dimensionality reduction, effectively avoids model overfitting, and improves generalization capability. The proposed method provides an intelligent solution for power engineering budget management and can be extended to other fields involving uncertainty cost estimation.

[Objective] With the widespread integration of distributed energy, the complex topology and exponentially increased monitoring data of distribution networks pose new challenges to fault diagnosis. Traditional fault diagnosis methods mainly rely on monitoring data and human experience. With the rapid development of cloud computing and communication technology, artificial intelligence methods are widely applied in the field of fault diagnosis. However, existing artificial intelligence methods have a high dependence on training data, requiring a large number of basic data as support. Therefore, an intelligent fault diagnosis method for distribution networks was proposed by leveraging digital twin technology to improve the efficiency and accuracy of fault diagnosis. [Methods] A digital twin of the distribution network was constructed using digital twin technology, and virtual diagnosis results were used to guide actual system operation. Additionally, wavelet packet decomposition was utilized to obtain the energy of each frequency band of the signal to construct feature vectors, which were input into the improved convolutional autoencoder (CAE) model for learning to identify the fault type. The digital twin system included a physical layer, a data layer, a model layer, and a service layer, achieving virtual-real mapping, with the virtual twin reflecting the state of the physical entity in real time. In the simulation experiment, the three-port ring network structure of a 10 kV distribution network in an area was used as the basis, and a complete experimental dataset was constructed, including 7 520 pieces of normal and fault sample data. [Results] The performance analysis results of the proposed model show that after 100 iterations of training, the diagnostic accuracy of the improved CAE model is close to 0.98. Moreover, the intelligent diagnosis results of the digital twin system demonstrate that the fault types diagnosed by the proposed method are basically consistent with the actual fault types, and for five common fault types, it maintains an ideal diagnostic accuracy. The average accuracy reaches 0.95, and the diagnosis time is only 5.39 s. A comparison of diagnoses using different methods indicates that the proposed method has a higher diagnostic accuracy. [Conclusion] The application of digital twin technology to the intelligent fault diagnosis of distribution networks, by adopting the approach of virtual-real integration, further improves the accuracy and real-time performance of fault diagnosis, thus providing a new technical means for the intelligent fault diagnosis of distribution networks. This contributes to enhancing the reliability and safety of distribution networks and holds important theoretical and practical value for the development of smart grids. Furthermore, future research will focus on how to cope with the changes in the structure of distribution networks to improve the applicability of the proposed fault diagnosis method.

[Objective] Power engineering projects are typically characterized by high costs and long durations, and their construction processes are influenced by various factors such as climate conditions and material costs. Traditional methods for cost and duration prediction are mainly based on experience, which can lead to underestimated or excessive cost estimates, resulting in project delays or resource waste. With the rapid development of machine learning techniques, data-driven methods have been introduced in cost and duration prediction. However, due to the small size of datasets in power engineering, traditional machine learning models often suffer from overfitting, which limits their predictive performance. Thus, a hybrid model combining support vector regression (SVR), classification and regression trees (CART), multivariate linear regression (MLR), and grey wolf optimization (GWO) was proposed to improve prediction accuracy and generalization ability on small datasets by enhancing the update strategy and parameter search method. [Methods] The main approach of this paper was to combine machine learning models with an improved GWO (iGWO) algorithm to develop an efficient framework for predicting the cost and duration of power engineering projects. SVR, CART, and MLR models were used as baseline machine learning methods. GWO was employed to search for optimal parameters to prevent overfitting, with two improvements introduced: using chaotic sequences to initialize the wolf pack positions to ensure population diversity, and optimizing the update strategy of the grey wolves′ positions and enhancing the search ability by sharing information within the surrounding pack. [Results] Experimental results show that the proposed hybrid model outperforms traditional methods in cost and duration prediction. Performance comparisons on the training and testing sets indicate that traditional machine learning models are prone to overfitting, which results in poor generalization. In contrast, the model combined with GWO improves this issue. The MLR+GWO hybrid model performs better than the other models on both the training and testing sets. Further experimental results reveal that the convergence speed of the hybrid model is significantly accelerated by the iGWO algorithm. It reaches optimal fitness within 6 to 8 iterations, while the traditional GWO algorithm requires 11 to 12 iterations to achieve similar results. Additionally, the improved algorithm effectively avoids the issue that the traditional GWO algorithm is prone to falling into local optima. [Conclusion] The hybrid model based on linear regression and iGWO demonstrates performance advantages in predicting costs and durations of power engineering projects. The iGWO algorithm enhances the global search capability and convergence speed through optimized initialization sequences and update strategies. The proposed hybrid model exhibits better generalization performance, compared to traditional machine learning models. Compared with traditional methods, this approach performs better in terms of prediction accuracy and training efficiency.

[Objective] The cost prediction of electric power engineering is of important significance in resource optimization, financial stability, risk management, efficiency improvement, project decision-making, policy formulation, market order maintenance, and investor decision-making for power grid enterprises. To address the problem of poor comprehensive performance of traditional cost prediction methods for electric power engineering, a prediction model based on a gradient boosting decision tree (GBDT) was proposed, with the small sample characteristics of power engineering cost data taken into account. The accuracy of cost prediction was improved significantly by optimizing the residuals generated during the training process. [Methods] The influencing factors of electric power engineering cost were analyzed in depth from both natural environment and technology perspectives. Eleven key variables influencing electric power engineering cost were found, and feature engineering conforming to the GBDT model was constructed through data cleaning, feature encoding, and logarithmic transformation. The hyperparameters were optimized using the Optuna framework, and the model was evaluated using a 5-fold cross validation method. The evaluation index was selected as the goodness of fit in model optimization, and the optimal hyperparameters were iteratively searched for until the prediction accuracy reached the required level or the maximum number of iterations was reached. Finally, a GBDT prediction model combined with the Optuna framework was established. The cost data of substation engineering in an area were taken as an example for experimentation, with 90％ of the data samples used as training and validation sets, and the remaining 10％ as test samples. The prediction results of random forest, neural network, GBDT algorithm, and the GBDT model combined with Optuna were compared, and their performances were evaluated by goodness of fit and root mean square error. [Results] The experimental results show that the prediction performance of the GBDT model combined with Optuna is better than those of random forest, neural network, and GBDT algorithm. The predicted values fluctuate within the range of ±10 CNY/kVA on the basis of the true values, and the goodness of fit reaches 0.892 3, and the root mean square error is 8.01 on the validation set. The goodness of fit reaches 0.886 6, and the root mean square error is 8.09 on the test set. [Conclusion] The cost prediction model of electric power engineering based on a GBDT algorithm can predict electric power engineering cost accurately. Compared with traditional prediction methods, it has higher prediction accuracy, especially suitable for small sample datasets such as electric power engineering cost. Combined with the Optuna framework for hyperparameter optimization, it further improves the accuracy of cost prediction. Subsequent research will collect more sample data and integrate neural network algorithms to achieve better prediction results and promote the efficient operation and healthy development of power grid enterprises.

[Objective] Accurate electric load forecasting is the key to the smooth operation and effective management of power systems, which can enable power companies to effectively dispatch power generation equipment, thereby improving their operational efficiency and economic benefits. However, electric load data are affected by a variety of external factors and have significant time dependence, which makes their accurate prediction difficult. Therefore, an electric load forecasting model combining multi-factor modeling and time series analysis was proposed, which taking into account the analysis of the complex influences of multiple factors and the time dependence characteristics of electric load, so as to realize accurate electric load forecasting. [Methods] To break through the respective limitations of multi-factor analysis methods and time series forecasting modeling methods, an improved electric load forecasting model combining long short-term memory (LSTM) network and Bayesian optimization algorithm was proposed with the help of deep learning and a multi-factor analysis method. Firstly, a comprehensive multi-factor feature pool was constructed, including the historical time series features of electric load and a variety of external factors to fully capture the complex relationships between electric load data and multiple influencing factors. Secondly, the LSTM network was used as the core model, and its unique gating mechanism and memory unit were used to capture the time dependence of electric load data and the complex association between multiple factors. The Bayesian optimization algorithm was introduced to tune the hyperparameters of the LSTM model, and the Gaussian process was used as the surrogate model to make full use of the prior information to improve the training efficiency and prediction performance of the model. [Results] Five real transformer datasets were used to train and test the model. The effectiveness of the model was verified by several evaluation indicators. The proposed electric load forecasting method based on multi-factor feature engineering modeling has significantly better prediction performance than the model using only a single factor for forecasting on five different transformer datasets, which further highlights the effectiveness of the multi-factor feature pool. The maximum coefficient of determination of the LSTM model is 0.920 7, and the minimum mean square error and the minimum mean absolute error are 0.042 and 0.024, respectively. The results demonstrate the superior performance of the proposed method in complex electric load forecasting tasks. [Conclusion] The electric load forecasting model combining multi-factor modeling and time series analysis fully considers the complexity of external factors and the time dependence characteristics of electric load and innovatively introduces a comprehensive feature pool to participate in LSTM model training and testing. The LSTM network combined with multi-factor feature pool modeling has high prediction accuracy and robustness, which provides a new technical idea for electric load forecasting, has important reference value for the planning and dispatch of smart grid, and lays a foundation for further development of accurate load forecasting technology.

[Objective] Traditional cost prediction methods for power grid substation engineering often rely on single influencing factors or linear assumption models, which fail to comprehensively capture the complex non-linear relationships among multiple factors, resulting in low prediction accuracy. Furthermore, existing methods face challenges such as dimensionality explosion or information loss when handling high-dimensional categorical variables, and especially, overfitting is prone to occur in small-sample datasets. Therefore, this study aims to develop a robust cost prediction model for substation engineering that effectively integrates multi-source influencing factors, adapts to non-linear relationships, and performs well in small-sample scenarios, thereby providing more accurate technical support for investment decisions in power grid enterprises. [Methods] To address these issues, a substation engineering cost prediction model (ME-XGB) based on the fusion of mean encoding (ME) and the extreme gradient boosting (XGBoost) framework was proposed. First, 13 key influencing factors were extracted from dimensions such as equipment and materials, construction techniques, construction scale, geographical environment, and design standards, covering both categorical and continuous variables. For categorical variables exhibiting non-linear relationships with cost, ME was applied for feature engineering. This method converted categorical variables into continuous features by calculating the mean of the target variable (cost per unit capacity) within each category and combining with a smoothing factor to retain category information while avoiding dimensionality explosion. Second, the XGBoost algorithm was utilized to construct the prediction model. The generalization ability of the model was enhanced by integrating multiple decision trees to iteratively correct residuals and incorporating regularization terms and hyperparameter tuning. Experiments were conducted using 200 substation engineering samples from a power grid company, which were randomly divided into a training set (80％) and a test set (20％). The performance of ME-XGB was compared with MK-TESM based on a Mann-Kendall (MK) trend test method and a three exponential smoothing method (TESM), backpropagation (BP) neural network, and the original XGBoost model by using mean absolute error (M_AE) and goodness of fit (R²) as evaluation metrics. [Results] Experimental results demonstrate that the ME-XGB model significantly outperforms comparative models in prediction accuracy on the test set. Specifically, the median and mean M_AE values of ME-XGB are 5 and 6.875, where are lower than those of MK-TESM, BP neural network, and the original XGBoost. Additionally, the R² value of ME-XGB reaches 0.857 9, significantly higher than those of the other models, indicating stronger explanatory power for data variations. Boxplot analysis further reveals that ME-XGB has the narrowest distribution range of prediction errors, confirming its greater stability. Hyperparameter tuning results show that settings of hyperparameters such as tree depth and learning rate effectively balance model complexity and overfitting risks. [Conclusion] The proposed ME-XGB model addresses the challenges of non-linear representation and dimensionality control for categorical variables through ME, while leveraging the ensemble learning capability of XGBoost to significantly enhance prediction performance in small-sample scenarios. ME-XGB outperforms traditional models in terms of M_AE, R², and error stability, providing a more reliable cost prediction tool for power grid enterprises. Future research can further explore the modeling of dynamic influencing factors and extend the application of the model to cross-regional projects through transfer learning.

[Objective]In the field of traditional permanent magnet motor fault monitoring, while contact signals are widely used, they usually only reflect one operational state of motors, leading to insufficient information and difficulty in comprehensively identifying the operational state of permanent magnet synchronous motors. To enrich the amount of information, additional sensors are needed, which not only increases the complexity of the system but is also difficult to be practically applied. Therefore, improving the accuracy and convenience of permanent magnet motor state monitoring has become an important research objective. With the development of intelligent monitoring technology, the application of non-contact signals has received increasing attention. The audio signals generated by the operation of permanent magnet motors contain rich state information, providing a new direction for fault diagnosis. Compared with contact signals, audio signals can reflect in real time such characteristics as motor vibration and noise caused by faults, which have significant research value. However, these signals are easily interfered by environmental noise, which results in poor signal quality and unclear feature information and is thereby not conducive to the state monitoring of permanent magnet synchronous motors. Therefore, a deep learning model based on voiceprint recognition was proposed for permanent magnet synchronous motors, aiming to efficiently monitor and diagnose operational states of motors through deep learning technology. [Methods]Firstly, the wavelet denoising algorithm was used to reduce noise interference, improve signal quality, and thus enhance the signal-to-noise ratio, ensuring that the model can more clearly extract Mel cepstral features and laying the foundation for fault identification and classification. However, direct use of convolutional neural networks (CNNs) to extract Mel cepstral features may weaken the correlation between features, affecting the accuracy of fault identification. To address this, a spatial attention mechanism was introduced, which enhanced the spatial position correlation of features through weighting, leading the model to focus on the most critical parts and thus improving the effectiveness of feature extraction. To boost the recognition accuracy of the model, normalization of Mel cepstral features was performed, and the AAM-softmax loss function was employed. This function strengthened inter-class constraints, improving the distinguishing capability of the model between different categories, thereby enhancing the recognition accuracy and generalization ability, and optimizing the training process, so that the model was enabled to better adapt to different operating conditions. [Results]Simulation test results indicate that the proposed model performs excellently on the training set, accurately identifying the different operational states of the motor, and demonstrates strong generalization ability on the test set. The experimental results confirm that the deep learning-based voiceprint recognition method can effectively monitor the various operational states of permanent magnet motors with high accuracy and practicality. [Conclusion]In summary, the proposed deep learning model based on voiceprint recognition for permanent magnet synchronous motors can effectively eliminate noise and extract key features. By introducing the spatial attention mechanism and the AAM-softmax loss function, the model significantly enhances the recognition accuracy and generalization ability. With broad prospects for development, this model can be widely applied in state monitoring and fault diagnosis of permanent magnet motors and promote the development of intelligent maintenance technology of motors.

[Objective]Transmission lines are an important link in the transmission and use of electrical energy, and their safety and stability play a crucial role in the normal operation of the power system. Therefore, daily inspections of transmission lines are of great importance. Major accidents usually develop from small defects and hidden dangers. Daily inspections usually use manual, unmanned aerial vehicle, visualization channels, and other means. Regardless of the method, a large number of visualization, infrared, or ultraviolet photos need to be processed. However, due to the particularity of transmission lines, the installation conditions involve multiple environments, and the inspection image background is usually complex. Although the manual review method has high accuracy, it relies heavily on experience and has extremely low efficiency. Therefore, how to quickly and accurately identify inspection images of overhead transmission lines is the key to identifying defects in overhead transmission lines. The traditional image recognition method for transmission line inspection is prone to low defect recognition accuracy under complex background interference. [Methods]Therefore, to enhance the recognition accuracy of detection images of overhead transmission lines under complex backgrounds, a defect detection method that balances recognition efficiency and accuracy was proposed. The proposed method was based on compressed image technology combined with the YOLOv5 model. Firstly, an asymmetric feature aggregation compression algorithm based on sparse convolution was designed. The original image was encoded to reduce the space required by image storage data for storage and transmission. After being transmitted to the decryptor through the information channel, the compressed image was decoded and restored to improve the learning efficiency of local set features. At the same time, by the integration of the channel-spatial attention module (CSAM), the attention channel weight matrix and spatial weight matrix were obtained from the feature map, and the importance of the feature map region was determined through the weight matrix. In this way, the processing efficiency of the YOLOv5 model was improved. [Results]The compressed and restored image was input into the improved YOLOv5 model. The channel attention module (CAM) and the spatial attention module (SAM) were used to process the attention data on the channel and space of the image, respectively. The features of the target area were enhanced through global average pooling and maximum pooling, and the SAM was introduced to enhance the attention of channel attention to feature position information, so as to detect defective devices. The effectiveness of the proposed method was verified experimentally. [Conclusion]The inspection image data set of an overhead line was used as the basis for training and testing the proposed detection method. The results show that the sizes of the detection images are significantly reduced after compression using the proposed technique, and the sizes of the restored images are reduced by about 3 MB, compared to those of the original images, without distortion. The improved YOLOv5 model has high detection precision, with detection accuracy reaching 0.91 and detection time being as short as 0.87 s. The algorithm ensures detection accuracy while reducing image size and improving detection speed.

[Objective]Transmission lines are an important component of the power system, and most line faults are caused by lightning strikes. Lightning interference identification is an important basis for ensuring the correctness of traveling wave fault analysis. To quickly identify lightning strike faults in 500 kV ultra-high-voltage direct current transmission lines and ensure the stability of the power system, an automatic identification method for lightning strike faults was proposed. [Methods]The dictionary learning algorithm was used to denoise the transmission line signal, and the minimum error objective function was established for signal amplitude fluctuation. The dictionary matrix was optimized by dictionary update through hot start and Newton iteration to obtain the denoised lightning strike fault signals of transmission lines. This effectively reduced noise interference and improved identification accuracy. The wavelet time entropy method was used to extract key features from the denoised lightning strike fault signals of transmission lines. The wavelet coefficients formed by wavelet transform were used to reconstruct the coefficients in a specific layer. A sliding time window was defined to calculate entropy and information content, and features were extracted from the transient signal of lightning current in transmission lines to provide data support for lightning strike fault identification. Different characteristic signals of lightning strikes were collected, and features were trained using ensemble learning algorithms. Multiple weak classifiers were generated and fused into a strong classifier through weights, which was used to classify each transient signal sample of lightning current. The generalization ability of the classifier was improved, and it was enabled to cope with different types of lightning strike fault signals. The classifier was optimized using the sparrow algorithm, and the optimal parameters of the classifier were obtained by randomly initializing the sparrow population, calculating fitness values, screening sparrows, updating sparrow discoverers and joiners, and performing mutation operations. The optimal parameters were input into the optimized classifier to achieve automatic identification of lightning strike faults in 500 kV ultra-high-voltage direct current transmission lines. The sparrow algorithm, as a heuristic optimization algorithm, has the characteristics of adaptability and a strong global search ability. It can quickly find the optimal parameters of the classifier in a complex search space, improving optimization efficiency and identification speed. [Results]The experimental results show that the proposed method has a signal-to-noise ratio (SNR) of over 40 dB, a mean square error (MSE) of identification to be less than 1.5, an identification efficiency of over 90％, and identification time of about 2.5 s after denoising. It can accurately and efficiently identify lightning strike faults in 500 kV ultra-high-voltage direct current transmission lines. [Conclusion]This method provides a new technical means for automatic identification of lightning strike faults in 500 kV ultra-high-voltage direct current transmission lines. It significantly enhances identification accuracy and efficiency, providing strong support for the safe and stable operation of the power system. At the same time, this method can also be extended to identification of other fault types of transmission lines, which has a wide application value.