Qiang QIN1 Yongjiao YANG2, Jiaxin LIN 3 Hanye HUANG4 Yuetian HUANG5
Affiliation: Information Center, Guangdong Power Grid Co., Ltd., Guangzhou, Guangdong 510000, China
1、Email:qin201709092025@163.com
2、Email:meigui10203040@163.com
3、Email：yu20041128fjzdf@163.com
4、Email:qun19860616@163.com
5、 Email:Hyt24@pku.org.cn
Abstract: A real-time fault detection method combining deep learning detection model and adaptive human-computer interaction is proposed to address the problems of poor real-time performance and high cognitive load of operators in power transmission network fault detection. Create a system consisting of four parts: data collection and preprocessing, deep learning detection, adaptive interaction, and operator feedback. The CNN-LSTM Transformer combined model is used in the algorithm to achieve accurate identification and rapid localization of various faults; From the perspective of human-computer interaction, reducing the workload of operators through user profiling, dynamic alarm priority, and interpretable visualization. The experimental basis of this article is based on IEEE standard power systems and data from actual power grids. The results show that the method proposed in this paper has a detection accuracy 98.9% higher than traditional impedance methods and non adaptive deep learning systems, with a delay of less than 86ms and a 22% reduction in NASA-TLX psychological load index. Research has found that adaptive human-computer interaction combined with intelligent detection can improve the stability and intelligence of power transmission networks, providing a new solution for future safe and efficient management methods.
Key words: power transmission network; Fault detection; Adaptive human computer interaction
1 Introduction
With the expansion of the power system and the integration of a large number of renewable energy sources into the grid, the risk, faults, and operation of the power transmission network are becoming increasingly complex. According to the latest market report, the global market for intelligent systems used for power transmission fault detection is estimated to have reached $3.54 billion by 2024, and is expected to grow at a compound annual growth rate of approximately 9.82% to $8.45 billion by 2034[1]. In this context, how to make correct judgments on faults as early as possible and transmit the results to operators has become an important issue to ensure the reliability and response speed of the power grid.
Common faults in transmission lines include inter turn short circuit, ground short circuit, and wire breakage. At present, the application of modern machine learning and deep learning techniques has achieved a high level of accuracy in fault detection and qualitative analysis. For example, a fault identification accuracy based on this method in high-voltage direct current systems can reach 99.97% [2]. In addition, scholars have proposed a coherence analysis method based on phase current waveform to quickly identify the type and location of line faults [3]. These technological means have pushed fault detection to a new high level.
Most existing research has focused on how to improve the accuracy of algorithm models and enhance signal processing capabilities. However, there is still little attention paid to how to present detection results in a way that is suitable for operators’ cognition and decision-making. Even if the detection model itself is accurate and fast, if the interface design or alarm strategy cannot interact with the operator, there may still be too much and too much information, errors, and delays, which may make the system less useful. On the other hand, Human Machine Interaction (HMI) has developed rapidly in recent years and has also begun to receive attention in industrial systems. HMI technology can improve the way information is transmitted between humans and systems, allowing operators to have a more direct and faster understanding of the system’s condition and abnormal situations[4].
Therefore, the purpose of this article is to propose an adaptive human-computer interaction method and combine it with a real-time fault detection model to form a collaborative system. Attempt to fill the gap between power system fault detection and HMI research, and construct a scalable adaptive human-machine collaboration framework that can be implemented.
2 Theoretical Basis and Related Work
2.1 Fault detection methods for power systems
The power system belongs to the infrastructure of national economy and social operation. In this field, its fault diagnosis technology has always been the core of research, which can be roughly divided into three categories.
(1) Method based on signal processing
Traditional power system fault detection mostly relies on the analysis of primary signals such as current and voltage. Common methods include Fourier transform, wavelet transform, coherence analysis, and so on. The Fourier method performs well in processing periodic signals, but has low sensitivity to non-stationary signals [5]; Wavelet analysis can have both time-domain and frequency-domain characteristics, and has great advantages in transient signal analysis. However, in complex scenarios, it is easily affected by noise and prone to misjudgment when the types of faults are varied[5].
(2) Methods based on machine learning and deep learning
In recent years, artificial intelligence has been applied in the field of power systems, and research has shown that convolutional neural networks, long short-term memory networks, and transformer models have an accuracy rate of over 95% in identifying faults in power transmission lines [6]. The deep learning based fault classification and localization method proposed by Ö zdemir et al. (2025) [7] achieved an accuracy of 99.97% on the IEEE 39 bus system. However, deep learning models are often “black boxes” that are unexplainable to operators.
(3) Method based on digital twin
With the maturity of virtual simulation technology, digital twins are widely used in the state monitoring and predictive maintenance of power networks. Digital twins can achieve “virtual real mapping”, utilizing real-time updates of virtual power grids to anticipate potential problem nodes in advance and improve predictability. However, it requires a large amount of real-time data and high-performance computing platforms, and its application in cross regional power grids is still relatively difficult.
2.2 Application of human-computer interaction in electromechanical/power systems
(1) SCADA system
Supervisory Control and Data Acquisition (SCADA) system plays a very important role in the power system. The SCADA interface can collect and process power grid data and provide dispatchers with an operational interface, but the interface is mainly composed of tables and graphics. Although there is a lot of information, the interactivity is not good enough [8].
(2) Visual monitoring platform
With the development of graphic technology, visual monitoring systems have begun to emerge. Research has shown that 3D visualization platforms and augmented reality technology can enhance operators’ awareness of the operational status of the power grid. However, these platforms often only display information without intelligent adaptation capabilities [9].
(3) Intelligent HMI (Human Machine Interaction)
In recent years, intelligent human-computer interaction technology has been applied to the field of electric transportation, improving efficiency and reliability. For example, HMI technology is used in the fault detection process of highway electromechanical equipment, which presents the results of circuit fault diagnosis to maintenance personnel on an intuitive interface, greatly reducing the time required for detection and relieving maintenance personnel of excessive workload [10]. In the power system, similar mechanisms can also be applied to grid scheduling and operation: using dynamic information filtering and improved interaction to enable operators to detect faults earlier and respond quickly [11].
3 Methodology
3.1 Overall System Architecture Design
In order to enable real-time detection of power transmission network faults and achieve intelligent interaction, this paper proposes a four layer system architecture including data acquisition and preprocessing layer, fault detection layer, adaptive human-computer interaction layer, and operator feedback layer. The design purpose of this architecture is to build a closed loop process system, so that the entire system can quickly and accurately identify various faults. In this process, human-machine interaction can further enhance the decision-making efficiency and correctness of the operation and maintenance personnel themselves.
At the level of data acquisition and preprocessing, the system relies on synchronous phasor measurement units (PMUs) and supervisory control and data acquisition systems (SCADA) to obtain data from multiple signal sources such as voltage, current, phase angle, frequency, etc. Filtering, normalization, and integrity verification are used to process these raw data, thereby removing noise and outlier interference. The caching mechanism ensures the temporal continuity and transmission stability of the data. In the fault detection layer, the system uses a hybrid deep learning model to complete local feature extraction, time series correlation modeling, and global pattern recognition, achieving high-precision and low latency detection performance in large-scale power systems.
At the level of adaptive human-computer interaction, the information display method and alarm level settings are changed based on operator portraits and workload conditions, and interpretable artificial intelligence components are added. The judgment thinking logic is reflected using heat maps and feature contribution degrees, reducing the cognitive pressure on operators and enhancing their trust in their detection results. The operator feedback layer allows users to confirm or correct the detection results output by the system. The feedback information will be stored in real time and applied to the iterative updates of the subsequent model, enabling the system to have adaptive improvement capabilities. Overall, this architecture achieves a cyclic operation of data collection, intelligent detection, automatic adjustment of communication, and feedback, laying the methodological foundation for subsequent experimental confirmation and application.
Figure 1 Architecture of Real time Fault Detection System for Power Transmission Network
3.2 Adaptive Human Computer Interaction Mechanism
Traditional power system monitoring and scheduling interfaces generally use fixed tables and static graphics, which can provide all data but cannot be changed according to specific situations, and are prone to generating too much information in the event of sudden failures. In response to this drawback, this article proposes an automatic adaptive human-machine interaction form, aiming to flexibly improve information presentation and communication processes based on the characteristics of the operator and their immediate responsibilities, thereby optimizing the speed of fault handling and the reliability of decision-making.
(1) User profiling modeling.
The system will collect the operator’s past operational data, error rate, experience level, and subjective load evaluation based on the NASA-TLX scale to form a dynamic user profile. According to different types of operators, the system will automatically select the appropriate form of information presentation. Novices are given a simplified alarm interface, while experienced users can obtain a view of all parameters to achieve personalized interaction.
(2) Dynamic alarm priority.
When a fault occurs, the system will automatically arrange and filter out alarm information according to the severity of the fault and the operation status of the power grid. Only the most important information will be pushed to the main interface, while the remaining information will be displayed in a graded manner in the form of color, sound, or location prompts, so as not to disrupt the operator’s operation rhythm due to the “alarm avalanche” phenomenon.
(3) Interaction design.
This article selects three different modes in the interaction mechanism, namely automatic mode, semi-automatic mode, and auxiliary mode. In automatic mode, the system directly provides the results of fault diagnosis and localization; In semi-automatic mode, the operator needs to confirm the candidate results of the system; In auxiliary mode, operators can manually intervene and annotate the system’s output to achieve human-machine collaboration.
(4) Interpretability and auxiliary decision-making.
With the help of explainable artificial intelligence (XAI), the judgment criteria of the detection model are visualized as heat maps or feature contribution distributions, so that operators can understand at a glance how the system makes decisions and make more appropriate interventions when necessary. Through the above design, the adaptive human-machine interaction mechanism proposed in this article not only reduces the cognitive pressure caused by information redundancy, but also achieves “human-machine co intelligence” in various forms of interaction, thus providing support for the intelligent operation and maintenance of power transmission networks.
3.3 Fault detection algorithm design
In the design of fault detection algorithms, this article adopts the design concept of “traditional methods+deep learning methods”, aiming to make the results both accurate and real-time, and also relatively easy to interpret and use. The traditional fault location method in power systems has laid a certain theoretical foundation for the emergence of intelligent algorithms in the future, and the use of deep learning technology can also improve detection results in complex power grid environments.
(1) Input features.
This article comprehensively considers primary parameters such as voltage amplitude, current amplitude, phase angle difference, and frequency deviation, and also incorporates power quality indicators such as harmonic distortion rate to enhance the sensitivity of different types of faults. Using multiple features as inputs can ensure that the model can capture abnormal signals even in complex operating environments.
(2) The traditional method of benchmarking.
This article uses the single-phase grounding fault location formula of impedance method as the benchmark model for traditional methods:
Among them, D is the distance to the fault point, V is the measured voltage, I is the fault current, and Z ‘is the impedance value per unit length of the line. This method is widely used in engineering practice, but it is more susceptible to interference in noisy environments. This article will use it as a reference standard for comparative experiments.
(3) Deep learning models.
This article proposes a hybrid CNN-LSTM Transformer model structure. We use Convolutional Neural Networks (CNN) to extract local features of power signals, Long Short Term Memory Networks (LSTM) to capture temporal dependencies of power signals, and Transformer modules to achieve global pattern recognition modeling with multi head attention. The combination of these three points can ensure the sensitivity of transient faults and also make the model have strong stability and robustness over long time series.
(4) Loss function and optimization.
This article uses the cross entropy loss function to supervise the classification results during the training and optimization process of the model:
The true label of the sample is represented by, and the predicted probability distribution of the model is represented by. In order to alleviate the problem of class imbalance in power fault data, the FocalLoss mechanism is added to the loss function to improve the ability to identify faults with few samples. Using AdamW as an optimizer to ensure stable convergence during the training process.
Table 1 Experimental Parameter Setting Table
| Parameter Category | Parameter Name | Setting Value |
| Data Sampling Rate | PMU Sampling Frequency | 50 Hz / 60 Hz |
| Model Structure | Number of CNN Layers | 3 layers |
| Number of LSTM Units | 128 | |
| Number of Transformer Heads | 8 | |
| Training Parameters | Batch Size | 64 |
| Learning Rate | 0.001 (AdamW) | |
| Loss Function | Cross Entropy + Focal Loss | |
| Output Content | Fault Types | Single-phase Grounding, Line-to-line Short Circuit, Three-phase Short Circuit |
| Output Dimensions | Confidence Score, Fault Location, Feature Contribution Weights |
3.4 System Implementation Process
The true label of the sample is represented by, and the predicted probability distribution of the model is represented by. In order to alleviate the problem of class imbalance in power fault data, the FocalLoss mechanism is added to the loss function to improve the ability to identify faults with few samples. Using AdamW as an optimizer to ensure stable convergence during the training process. This paper designs a complete system implementation process to ensure that the proposed method can efficiently run in practical scenarios of power transmission networks. Firstly, in the data collection and preprocessing stage, PMU and SCADA systems are used to obtain various signals such as voltage, current, phase angle, and frequency. Before entering the detection model, all data must undergo filtering, normalization, and integrity checks to ensure the accuracy and continuity of the signal, and caching methods are used to reduce the impact of communication jitter on the signal. Subsequently, in the process of model inference, the deep learning detection model will perform further operations on the input data to quickly provide fault types, locations, and credibility scores, and also generate some feature contribution information for later analysis.
On this basis, enter the adaptive human-computer interaction process. Based on the profile of the operator and real-time load status, the detection results are judged. Different results are displayed in different modes, and important information is emphasized through graded alarms to prevent excessive alarms from affecting the speed of decision-making. When the system detects a serious fault, it will immediately activate the fault confirmation and protection functions. On the interface, operators can perform confirmation or correction work, and at this time, the system will start to perform tripping, isolation, or recovery operations, forming a closed-loop protection. Finally, the operator’s interactive feedback will be stored in the database, which is both the running log and the basis for later adjustment of model parameters. Through this process of “detection interaction feedback optimization”, the system can continuously possess adaptive capabilities in a dynamic environment and improve the performance of fault detection and decision assistance as a whole.
4 Experiments and Results
4.1 Data Sources and Experimental Platforms
In order to comprehensively verify the real-time fault detection method based on adaptive human-computer interaction proposed in this article, simulation data and actual power grid operation data were used as two sources in the experiment. In the simulation section, IEEE39 bus and IEEE118 bus systems were selected as the simulation experimental objects, and typical fault situations such as single-phase ground, phase to phase short circuit, and three-phase short circuit were simulated using Matlab/Simulink and PowerFactory platforms, and corresponding timing signals of voltage, current, and frequency were output. In terms of actual data, the operation logs and PMU measurement data of a certain power grid dispatch center were collected, including records of various disturbances and faults. This combination of “standard simulation actual power grid” makes the experiment representative and robust. When running the experiment, an NVIDIA A100 GPU high-performance server with Intel Xeon processor and 128GB memory was used. The algorithm used was Python 3.10 and PyTorch, and the signal processing was completed in Matlab.
Table 2 provides a detailed description of the experimental dataset, which includes both classical IEEE testing systems and real complex operating conditions of the power grid, laying a solid foundation for model training and validation.
Table 2 Experimental Dataset Explanation Table
| Dataset Type | Node Scale | Sampling Rate | Fault Types | Data Volume | Remarks |
| IEEE 39-bus | 39 nodes | 50 Hz | Single-phase Grounding, Line-to-line Short Circuit, Three-phase Short Circuit | 12,000 records | Generated by Matlab/Simulink |
| IEEE 118-bus | 118 nodes | 60 Hz | Single-phase Grounding, Line-to-line Short Circuit, Three-phase Short Circuit | 35,000 records | Generated by PowerFactory |
| Real Grid Data | Regional Grid, 42 nodes | 50 Hz | Disturbances and Multiple Faults | 8,000 records | PMU data from a regional dispatch center |
4.2 Experimental Design
In order to comprehensively verify the method proposed in this article, three comparative schemes were experimentally designed: (1) using traditional fault location methods based on impedance as the baseline for engineering applications; (2) Use deep learning to detect non adaptive systems with fixed interfaces to test algorithm contributions; (3) The system combining deep learning detection and adaptive HMI in this article serves as the final solution. Use the same experimental dataset and environment in all schemes to maintain fairness.
In terms of performance evaluation, accuracy, recall, f1 score, and detection delay are used at the algorithm level to measure the detection performance of the model; From the perspective of human-computer interaction, the NASA-TLX six factor psychological load scale is used in conjunction with user satisfaction to evaluate the effect of interaction design on improving operator decision-making ability. The experimental design and evaluation criteria are shown in Table 3.
Table 3 Experimental Comparison Scheme Table
| Comparison Method | Detection Mechanism | Interaction Mode | Explainability | Evaluation Metrics |
| Traditional Impedance Method | Impedance Calculation | None | None | Accuracy, Latency |
| Deep Learning + Non-Adaptive Interface | CNN-LSTM-Transformer Model | Fixed Interface | None | Accuracy, Recall, F1, Latency |
| Proposed Method (Deep Learning + Adaptive HMI) | CNN-LSTM-Transformer Model | Automatic / Semi-automatic / Assisted Interaction Modes | XAI-based Explainability | Accuracy, Recall, F1, Latency, NASA-TLX, Satisfaction |
4.3 Experimental results
In terms of technical performance, in the comparison of ROC curves among the three methods, the traditional impedance method showed a significant decrease in noise environment, deviating from the upper left corner; The detection performance of the deep learning+non adaptive interface system has been improved, with an AUC of 0.963; And the method in this article is closest to the top left corner, with an AUC of 0.992, which is the best.
Table 4 presents further quantified results. The traditional impedance method has an accuracy of 82.5% and a recall of 80.4%, which is not effective in complex working conditions. The accuracy of non adaptive deep learning systems has been improved to 95.7%, but there are still shortcomings in terms of recall and latency. This method achieved 98.9%, 98.3%, and 0.98 on accuracy, recall, and f1 score, respectively. At the same time, the detection delay was reduced to 86ms, which is 179ms shorter than traditional methods and 62ms shorter than non adaptive methods.
Table 4 Comparison of Experimental Performance of Three Methods
| Method | Accuracy | Recall | F1-score | Latency (ms) |
| Traditional Impedance Method | 0.825 | 0.804 | 0.81 | 265 |
| Deep Learning + Non-Adaptive HMI | 0.957 | 0.948 | 0.95 | 148 |
| Proposed Method (Adaptive HMI) | 0.989 | 0.983 | 0.98 | 86 |
From the perspective of detection delay, as shown in Figure 4, there is a gap between the three methods. The traditional impedance method relies on the impedance ratio between voltage and current signals to calculate, so the overall response speed is relatively slow, with an average detection delay of 265ms. Deep learning algorithms shorten the detection delay to 148ms through advantages such as feature extraction and pattern recognition. However, their fixed interface design still leaves some redundant information in alarm processing. This prevents the operator’s reaction time from being further reduced. The method proposed in this article improves the detection speed from a model perspective and utilizes adaptive human-computer interaction to reduce the delay caused by redundant alarms. The average detection time is 86 milliseconds, which meets real-time requirements. This reflects the crucial role of adaptive interaction mechanisms in improving overall response performance.
Figure 4 Line graph comparing delay detection using different methods
The subjective psychological burden and satisfaction obtained by 30 participants after completing the fault diagnosis task in the simulation environment in the human-computer interaction experiment are shown in Table 5. The adaptive HMI system has significantly reduced load scores in six aspects of NASA-TLX, with psychological load, time pressure, and operational complexity reduced by approximately 22%, 23%, and 26%, respectively, and overall satisfaction rate increased by 18%. The radar chart in Figure 5 shows that there are differences in user experience between two categories of systems, indicating that the method used in this article can reduce the cognitive pressure on operators to a certain extent and improve their efficiency in completing tasks.
Table 5 Subjective Experimental Results of Operators
| Evaluation Dimension | Non-Adaptive System | Proposed Method (Adaptive HMI) | Improvement |
| NASA-TLX Mental Demand (0–100) | 68 | 53 | ↓22% |
| NASA-TLX Temporal Demand | 71 | 55 | ↓23% |
| NASA-TLX Effort (Task Complexity) | 66 | 49 | ↓26% |
| Overall Satisfaction (0–10) | 7.1 | 8.4 | ↑18% |
Figure 5 User Experience Radar Chart
4.4 Result Analysis and Discussion
From the experimental results, it can be seen that the algorithm proposed in this paper has greatly improved detection performance and interactive experience compared to other comparative methods. From an algorithmic perspective, although the traditional impedance method is relatively simple and intuitive in theory, it relies on the linear relationship between voltage and current, and is easily disturbed by noise and load changes in complex power grid environments, resulting in unsatisfactory positioning accuracy. On the other hand, deep learning models can capture the nonlinear relationships of power signals in a multidimensional feature space, and use CNN to extract transient features, LSTM to capture temporal dependencies, and Transformer’s multi head attention mechanism for global pattern recognition. Despite multiple faults and disturbances, they still maintain a high detection accuracy of 98.9%. This also indicates that deep learning methods have natural advantages for complex operating environments.
From the perspective of human-computer interaction, experimental results show that adaptive HMI is effective. In non adaptive systems, although the algorithm detection performance has been improved compared to traditional methods, the lack of alarm priority management and user profile modeling require operators to handle a large amount of redundant information, resulting in an average psychological load of 68 points. The method proposed in this article uses a dynamic approach to filter out unimportant alarm signals, and changes the way information is presented to the staff based on their experience and workload level. This reduces the psychological burden to 53 points, and reduces time pressure and operational difficulty by 23% and 26%, respectively. This means that the system can provide faster inspection results and assist operators in making more effective decisions in complex environments, thus forming an efficient closed loop between algorithms, interactions, and judgments.
From the perspective of engineering application, the average detection delay of the method proposed in this article is only 86ms, which can meet the requirement of the power dispatch center for real-time alarms not exceeding 100ms. This not only ensures that faults can be detected in a timely manner, but also leaves sufficient response space for relay protection and accident handling. At the same time, operator satisfaction has increased by 18%, indicating that the system has been recognized for its practicality, thereby reducing manual errors and delays, and improving the overall reliability of power grid operation and maintenance.
However, this study still has limitations. On the one hand, the experiment was mainly conducted based on the IEEE testing system and operational data of a certain regional power grid, and did not involve the verification of models in complex scenarios of cross regional interconnected large power grids. On the other hand, real-time deployment of deep learning models requires a significant amount of computing resources, and as PMU and SCADA have become popular, further research is needed on how to reduce computing power consumption while maintaining low latency. The optimization of adaptive interaction mechanisms can be combined with cognitive science and behavioral models to more finely correspond to individual differences among different operators.
5 Conclusions
As the scale of the power system increases and operating conditions become increasingly complex, timely fault identification and execution of operation and maintenance decision support have become key challenges to ensure the safety and stability of the power grid. This article proposes a deep learning detection model and an adaptive human-machine interaction combination method to address the problems of insufficient real-time fault detection in power transmission networks and excessive cognitive load on operators. After creating a four layer structure of data collection and preprocessing, deep learning detection, adaptive interaction, and operator feedback, the system can achieve high accuracy and low latency in identifying various types of faults, and change the displayed message format according to the operator profile and real-time work pressure, thus forming a cycle system of detection, communication, and feedback. The experimental results show that the method proposed in this paper can achieve an average detection accuracy of 98.9% for IEEE standard power systems and actual power grid datasets, with a detection delay shortened to 86ms, and better accuracy than traditional impedance methods and non adaptive deep learning systems. According to NASA-TLX evaluation and satisfaction studies, adaptive interaction methods can reduce the psychological pressure on operators and improve decision-making speed. The intelligent development of the power system places greater emphasis on algorithm, interaction, and overall optimization of decision-making, rather than relying on a single detection technology. After the continuous popularization of PMU and SCADA, real-time monitoring and automatic interaction of large-scale power grids will become a trend. Meanwhile, through multimodal data fusion, intelligent visualization and other technological means, the intelligence level of fault detection and decision support can be further improved.
References:
- (2025). Smart fault detection system power transmission market report. Market Research Future.
- Khandelwal, D., et al. (2025). Fault detection in electrical power systems using attention-based model. Scientific Reports.
- Mahmoud, R. A., et al. (2025). Transmission line faults detection and classification using coherence-based scheme. PMC.
- (n.d.). Human–machine interaction. ScienceDirect Topics. Retrieved September 28, 2025, from.
- Dash, P. K., Mishra, S., & Panda, G. (2007). Fault classification and section identification of transmission lines using support vector machines. IEEE Transactions on Power Delivery, 22(1), 67–73.
- Santoso, S., Powers, E. J., Grady, W. M., & Bhatt, S. C. (2000). Power quality disturbance waveform recognition using wavelet-based neural classifier: I. Theoretical foundation. IEEE Transactions on Power Delivery, 15(1), 222–228.
- Özdemir, A., Yıldız, B., & Uyaroglu, Y. (2025). A transformer-based deep learning framework for fault classification and location in power transmission networks. IEEE Access, 13, 15234–15245.
- Kundur, P., Balu, N. J., & Lauby, M. G. (1994). Power system stability and control. New York: McGraw-Hill.
- Reddy, K. P., Kumar, A., & Karthikeyan, M. (2021). Real-time visualization and augmented reality applications in power system monitoring. International Journal of Electrical Power & Energy Systems, 127, 106652.
- Gong Changxing (2024). Fault detection of highway electromechanical equipment based on human-computer interaction Transportation World, (24), 100-102.
- Wang, H., Zhao, J., Liu, Y., et al. (2022). Human–machine interaction for smart grid operation: A survey. CSEE Journal of Power and Energy Systems, 8(3), 626–639.