A collaborative fault early warning method for power communication networks and dispatch automation systems

Abstract

This application discloses a collaborative fault early warning method for power communication networks and dispatch automation systems, relating to the field of power system fault early warning technology. This method collects multi-source data on equipment status, service operation, load, and digital twins. After time-series alignment and standardization, it uses a causal forest algorithm to extract dual-dimensional causal features related to long-term performance degradation and real-time operation, and dynamically allocates fusion weights. Combined with digital twin simulation, it dynamically calculates the fault impact level and handling priority, generates targeted handling instructions, and optimizes parameters through closed-loop feedback. It further enhances real-time performance and adaptability through an edge-cloud collaborative architecture and scenario adaptation mechanism, achieving deep fusion of multi-dimensional data. This method realizes the deep and organic integration of the full lifecycle status data of power communication network equipment, the dynamic load characteristics of dispatch automation system services, and the real-time simulated power grid operation status of the digital twin, enabling early warning several hours in advance and ensuring the safe and stable operation of the power grid.

Description

Technical Field

[0001] This invention relates to the field of power system fault early warning, and in particular to a collaborative fault early warning method between power communication networks and dispatch automation systems. Background Technology

[0002] With the increasing proportion of new energy grid integration and the growing complexity of power grid dispatching operations, the coupling between the power communication network and the dispatch automation system is becoming increasingly close. Their collaborative operation directly affects the overall safety and stability of the power grid. Currently, the industry has begun preliminary practices in cross-system collaborative fault early warning. This primarily involves collecting real-time operating parameters such as communication link latency, packet loss rate, and remote signaling frequency changes in the dispatch system to construct multivariate time series prediction models, achieving minute-level fault early warning responses. A pre-defined fault impact matrix is ​​then used to classify and assess the scope of fault impact. Digital twin technology is gradually being applied to power grid operation simulation and equipment status monitoring scenarios, and equipment remaining life prediction methods are being incorporated into the full lifecycle operation and maintenance system of power communication equipment. However, existing technologies have significant limitations. The core deficiency lies in the failure to achieve deep and organic integration among the full lifecycle status data of power communication network equipment, the dynamic load characteristics of dispatch automation system operations, and the real-time simulated power grid operation status of the digital twin. On the one hand, existing early warning models overemphasize real-time operational characteristics, neglecting the effective integration of long-term performance degradation indicators such as remaining equipment lifespan. This results in models that can only provide minute-level early warning capabilities for instantaneous faults such as sudden communication link interruptions. For example, using a unified early warning strategy for both new and old equipment fails to identify the gradual performance degradation process caused by the aging of communication equipment, such as the gradual accumulation of delays in dispatch command transmission until interruption. The lack of a mechanism for predicting faults several hours in advance causes maintenance personnel to lose the golden window for preventative handling. On the other hand, fault impact level assessment relies on a static impact matrix, which is based on fixed link importance and service priority configurations. It does not incorporate the dynamic feedback of real-time power grid operating conditions from digital twins, such as load fluctuations, unit start-up and shutdown, or line maintenance scenarios. This causes a serious disconnect between the quantitative results of fault impact and the actual operating environment. For example, the impact of a communication equipment aging fault on real-time power generation control services during peak load periods is much greater than during off-peak periods, but the static matrix cannot capture such dynamic differences. This leads to inaccurate fault priority ranking, resulting in imbalances in maintenance resource allocation and low handling efficiency. Summary of the Invention

[0003] This invention provides a collaborative fault early warning method for power communication networks and dispatch automation systems. It achieves deep and organic integration of the full life cycle status data of power communication network equipment, the dynamic load characteristics of dispatch automation system services, and the real-time simulation of power grid operation status by digital twins. This enables early warning several hours in advance and dynamically calculates the fault impact level and handling priority, significantly improving the accuracy of early warning and operation and maintenance efficiency.

[0004] To solve the above-mentioned technical problems, the present invention provides a collaborative fault early warning method for power communication networks and dispatch automation systems, comprising the following steps:

[0005] Data acquisition steps: Collect equipment status data of the power communication network, business operation data and load data of the dispatch automation system, and digital twin data reflecting the real-time operating status of the power grid; Feature processing steps: Time-series alignment of the data collected in the data acquisition step is performed to align them to the same time granularity, and normalization or standardization is applied. The normalization or standardization methods include maximum and minimum value normalization or Z-score standardization. Based on the causal inference algorithm, the causal relationship between equipment status data, business operation data, load data and digital twin data is analyzed, and long-term causal features for characterizing long-term performance degradation trends and real-time causal features for characterizing real-time operation correlations are extracted and fused. Early warning steps: Input the long-term causal features and real-time causal features into a multivariate time series prediction model to output early warning information; Assessment steps: Combining advanced early warning information with current real-time power grid operation status data, the digital twin is used to conduct simulation and dynamic calculation of the impact level and handling priority of the fault on dispatching services. Processing and optimization steps: Generate and execute disposal instructions based on the advance warning information, impact level and disposal priority, verify the disposal effect through a digital twin, and optimize the feature extraction and fusion strategy in the feature processing step, as well as the model prediction rules and evaluation rules in the warning step and evaluation step based on the verification results.

[0006] The present invention is further configured such that, in the data acquisition step: The device status data includes link latency, packet loss rate, and device remaining lifetime prediction parameters. The load data includes the real-time resource usage percentage of each service and the dynamic weight of service priority. The digital twin data is updated in association with the timestamps of the device status data and business operation data collection.

[0007] The present invention is further configured such that, in the feature processing step: The causal inference algorithm uses the causal forest algorithm; The extracted causal relationships include at least the causal relationship between the decay of the equipment remaining lifetime prediction parameters and the decline in the link performance of the power communication network, and the causal relationship between the changes in power grid operating conditions reflected by the digital twin data and the fluctuations in the dynamic weight of the service priority. Based on the strength of the extracted causal relationships, the fusion weights of the long-term causal features and real-time causal features are dynamically allocated in the subsequent model inputs.

[0008] The present invention is further configured such that: in the warning step: The multivariate time series prediction model is an LSTM-Transformer fusion model based on an improved attention mechanism; The advance warning information includes at least the probability of failure, the expected time of occurrence, the scope of affected services, and the fault causal chain identifier that traces the fault back to a specific device or power grid condition.

[0009] The present invention is further configured such that, in the evaluation step: The simulation and deduction using a digital twin specifically involves: based on the fault causal chain identifier, setting the equipment or operating condition parameters corresponding to the fault causal chain identifier to a fault state in the digital twin, performing simulation and deduction, and obtaining the evolution trajectory of the long-term causal characteristics and real-time causal characteristics. The simulated trajectory is compared with the current actual trajectory, the degree of difference is calculated based on the comparison results, and the impact level is dynamically calculated based on the degree of difference. The processing priority is calculated based on the dynamic weight of the business priority obtained from the load data and the degree of difference; The impact level is quantified into three levels: general, moderate, and severe, based on a preset threshold.

[0010] The present invention is further configured such that, in the processing and optimization steps: The handling instructions are generated based on the fault causal chain identifier and the handling priority, including equipment pre-maintenance, dynamic adjustment of link bandwidth, or high-priority service routing switching contingency plans. Verifying the effectiveness of the treatment through a digital twin is achieved by comparing whether the characteristic indicators corresponding to the fault causal chain identifier before and after the treatment have recovered to the normal threshold. The feedback optimization includes at least using the verification results to optimize the fusion weight allocation rule between the input feature set of the causal inference algorithm in the feature processing step and the long-term causal features and real-time causal features.

[0011] The present invention is further configured such that: the verification of the treatment effect through digital twins specifically includes: Based on the fault causal chain identifier and historical data of characteristic indicators before handling, the threshold curve of the characteristic indicator corresponding to the fault causal chain identifier in the fault-free state is dynamically simulated and generated in the digital twin as an adaptive recovery threshold curve. The actual characteristic index sequence after treatment is compared with the adaptive recovery threshold curve. If the actual characteristic index sequence continuously falls within the envelope of the adaptive recovery threshold curve within a preset time, the treatment is deemed effective. The generation of the adaptive recovery threshold curve incorporates the current power grid operating conditions, business load, and recovery modes of similar historical faults.

[0012] The present invention is further configured such that: the processing and optimization steps also include a dynamic adjustment sub-step: During the execution of the disposal instructions, the changing trend and deterioration rate of the impact level are continuously monitored; The deterioration rate is the magnitude of change in the impact level per unit time. If the impact level is detected to be upgraded or the deterioration rate exceeds a preset threshold, the contingency plan adjustment mechanism will be automatically triggered. The contingency plan adjustment mechanism is based on the real-time simulation of the digital twin, matching or combining contingency plan databases to generate upgraded handling instructions and push them out.

[0013] The present invention is further configured to include: In the model collaboration step, the difference degree calculated in the evaluation step is fed back as a real-time correction parameter to the fusion model in the early warning step, so as to adjust the model's attention distribution to the long-term causal features and the real-time causal features online.

[0014] The present invention is further configured such that the early warning step, the evaluation step, and the processing and optimization step are all executed using an edge-cloud collaborative architecture; The early warning step is executed on the edge computing node deployed on the plant side. The edge computing node runs a multivariate time series prediction model to generate local early warning information. The assessment and processing / optimization steps are executed on the cloud-based main station system. After receiving the local early warning information uploaded by the edge computing node, the cloud-based main station system uses full data and digital twins to perform simulation and assessment, and sends the generated handling instructions to the corresponding edge computing node and operation and maintenance terminal.

[0015] This invention, by employing the above technical solutions, achieves significant technical effects: through deep multi-dimensional data fusion, application of advanced algorithm models, deployment of edge-cloud collaborative architecture, and full-process closed-loop optimization, it comprehensively solves the core pain points in the collaborative fault early warning of existing power communication networks and dispatch automation systems. It realizes a full-chain intelligent upgrade from data acquisition to processing optimization. By accurately collecting multi-source heterogeneous data and ensuring time synchronization, it provides a high-quality foundation for subsequent processing. By leveraging the causal forest algorithm to mine deep causal relationships and dynamically allocate feature weights, it improves the targeting of feature processing. Finally, by utilizing an improved LSTM-Transformer model to output refined early warning information, it achieves… With hourly advance prediction capabilities, a scientific fault impact assessment and prioritization mechanism was constructed through dynamic simulation and difference analysis using digital twins. Combined with targeted handling instructions, dynamic adjustment strategies, and closed-loop feedback optimization, the mechanism ensured the accuracy and effectiveness of handling measures and the continuous iteration of system capabilities. The edge-cloud collaborative architecture and scenario adaptation steps further balanced the needs of real-time response and global optimization, as well as adaptation to routine operating conditions and high-risk scenarios. Overall, the system achieved a significant improvement in the accuracy of fault early warning, the timeliness of handling, and the rationality of resource allocation, effectively preventing the impact of progressive and transient faults on the power grid. This provides strong technical support for the safe and stable operation of the power grid under the background of high proportion of new energy grid connection. Attached Figure Description

[0016] Figure 1 This is a flowchart of the collaborative fault early warning method; Figure 2 This is a flowchart of the feature processing and early warning steps; Figure 3 It is a closed-loop flowchart for assessment, handling, and optimization. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0018] Example This application proposes a collaborative fault early warning method for power communication networks and dispatch automation systems, including the following steps: Data acquisition steps: Collect equipment status data of the power communication network, business operation data and load data of the dispatch automation system, and digital twin data reflecting the real-time operating status of the power grid; Feature processing steps: Time-series alignment and normalization or standardization are performed on the data collected in the data acquisition steps. Based on the causal inference algorithm, the causal relationship between equipment status data, business operation data, load data and digital twin data is analyzed. Long-term causal features for characterizing long-term performance degradation trends and real-time causal features for characterizing real-time operation correlations are extracted and fused. Early warning steps: Input the long-term causal characteristics and real-time causal characteristics into a multivariate time series prediction model to output early warning information; Assessment steps: Combining advanced early warning information with current real-time power grid operation status data, the digital twin is used to conduct simulations and dynamically calculate the impact level and handling priority of the fault on dispatching operations; Processing and optimization steps: Generate and execute disposal instructions based on the advance warning information, impact level and disposal priority, verify the disposal effect through a digital twin, and optimize the feature extraction and fusion strategy in the feature processing step, as well as the model prediction rules and evaluation rules in the warning step and evaluation step based on the verification results. The model collaboration step uses the difference calculated in the evaluation step as a real-time correction parameter to feed back to the fusion model in the early warning step, which is used to adjust the model's attention distribution to long-term causal features and real-time causal features online. In the scenario adaptation step, when the digital twin data indicates that the power grid is under peak load or planned maintenance conditions, the proportion of real-time causal features in the fusion weight is automatically increased, and enhanced collection of status data of key link equipment is triggered.

[0019] This method first performs a data acquisition step, which aims to obtain various basic data required for the operation of the power communication network and the dispatch automation system. Data acquisition can be achieved in various ways. The power communication network refers to the infrastructure network that provides communication support for the power system, carrying the transmission of key information such as power dispatch instructions and telemetry and telecontrol data. The equipment status data of the power communication network can be obtained through sensors deployed on the equipment or by periodically querying the equipment interfaces through network management protocols such as SNMP. This data covers basic physical parameters of each device, such as operating time, temperature, voltage, and current. The dispatch automation system refers to the automated system used to realize functions such as power grid operation monitoring, fault diagnosis, and dispatch control. It is the core of the safe and stable operation of the power grid. The business operation data of the dispatch automation system can be extracted from the business application by parsing the log file of the dispatch master station or through the API interface. For example, it records the start time, completion time and execution result of the business. Load data can be obtained through the monitoring tools of the operating system or virtualization platform, such as CPU utilization, memory usage, network bandwidth usage, etc. Digital twin data reflecting the real-time operation status of the power grid can be connected to the power grid simulation platform or geographic information system through data interface to periodically synchronize information such as power grid topology, power flow distribution, and equipment switching status. This data can be stored in a distributed database or data lake for subsequent processing.

[0020] After data acquisition, the feature processing step begins. This step first performs time-series alignment and standardization on the acquired heterogeneous data to eliminate time deviations and dimensional differences between different data sources. All data is unified to the same time granularity through timestamp matching and resampling techniques, and scaled to a uniform numerical range using max-min normalization or Z-score standardization. Subsequently, a deep analysis of potential causal relationships between equipment status data, business operation data, load data, and digital twin data is conducted based on causal inference algorithms. Graph-based causal inference algorithms can be used to construct causal graphs between data and identify direct causes leading to changes in other factors. From these causal relationships, two types of key features are extracted and fused: long-term causal features characterizing long-term performance degradation trends, such as the relationship between historical equipment failure rate and years of operation. Causal inference algorithms are used to identify and quantify causal relationships between different variables, as well as real-time causal features used to characterize real-time operational correlations, such as the causal relationship between network congestion and service response delay. These features can be extracted based on preset rules or statistical models, and their fusion can be achieved using a simple weighted summation method. Equipment status data refers to data reflecting the operating status of various equipment in the power communication network, such as equipment operating time, performance parameters, and health status. Service operation data refers to the operational status data of various services in the dispatch automation system, such as service execution frequency, response time, and data update status. Load data refers to the resource consumption data borne by the power communication network and dispatch automation system during operation, such as network bandwidth consumption, server CPU utilization, and memory consumption. Causal inference algorithms are statistical or machine learning methods used to identify and quantify causal relationships rather than merely correlations between different variables.

[0021] The next step is the early warning step. In this step, the long-term causal features and real-time causal features extracted and fused in the feature processing step are used as inputs and fed into a multivariate time series prediction model. This model can be a structure based on a recurrent neural network (RNN) or a long short-term memory network (LSTM). It is trained to learn the patterns of these features changing over time and predict future operating states. By applying a threshold judgment to the model output, early warning information can be generated. When a predicted key indicator deviates from the normal range to a certain extent, the warning is triggered. This warning information can be a risk alert, such as the existence of potential fault risks or a decline in the performance of a certain device.

[0022] After the early warning information is generated, the evaluation step begins. This step combines the early warning information with the current real-time operating status data of the power grid and uses a digital twin for simulation. A digital twin is a virtual mapping of the physical power grid and its communication system in digital space, which can reflect the operating status, behavior, and performance of the physical entities in real time and perform simulation. According to the potential fault type indicated by the early warning information, the corresponding fault conditions can be set manually or semi-automatically in the digital twin. Subsequently, the simulation model is run to observe the possible impact of the fault on power grid operation and dispatching services. Through the analysis of the simulation results, the impact level of the fault on dispatching services can be dynamically calculated. The impact level can be divided into different levels according to indicators such as the number of service interruptions and recovery time. Combining factors such as the importance of the service and the urgency of the fault, the corresponding handling priority can be determined, and faults that affect core dispatching services can be given higher priority.

[0023] Next comes the processing and optimization steps. Based on the early warning information, the assessed impact level, and the handling priority, corresponding handling instructions can be generated by the system. These instructions can be pre-set standardized operating procedures, such as suggesting to check a certain device, adjust network configuration parameters, or notify maintenance personnel to conduct on-site troubleshooting. These instructions are then executed. After the handling instructions are executed, the handling effect is verified through a digital twin. The handling instructions refer to the specific operational plans generated by the system and issued to maintenance personnel or automated equipment for potential or actual faults that have been warned, used to guide fault handling and system recovery. For example, the handling instructions can be simulated and executed in the digital twin, and their impact on the power grid operation status and business indicators can be observed and compared with the state before the handling to determine whether the handling is effective. Based on the verification results, the relevant parameters in the feature processing steps and the early warning and assessment steps can be fed back for optimization. For example, if a certain early warning model does not perform well after actual handling, the model's training parameters or feature selection strategy can be manually adjusted to improve its accuracy.

[0024] This method, by deeply integrating the full lifecycle status of power communication network equipment, the dynamic load characteristics of dispatch automation system services, and the real-time simulated power grid operation status of digital twins, effectively overcomes the shortcomings of existing technologies in progressive collaborative fault early warning. By dynamically calculating the impact level and handling priority of faults on dispatch services, this method achieves precise quantification of fault impact, thus providing maintenance personnel with a window for early prediction and preventative handling. This improves the rationality of maintenance resource allocation and handling efficiency, ensuring the safe and stable operation of the power grid under complex operating conditions. In some of the embodiments described above in this application, a collaborative fault early warning method for power communication networks and dispatch automation systems is proposed. Its data acquisition steps aim to obtain multi-source data. However, if the types of data collected are not detailed and comprehensive enough, or if the time granularity between different data sources is inconsistent, it may lead to the subsequent causal relationship analysis and prediction model failing to accurately capture the complex correlations within the system, thereby affecting the timeliness and accuracy of fault early warning.

[0025] This application further proposes refining and standardizing the data type and synchronization of the collected data during the data acquisition step to ensure the accuracy of subsequent analysis. In the data acquisition step, equipment status data includes equipment operating years, performance degradation rate, link latency, packet loss rate, and predicted parameters of remaining equipment lifespan; service operation data includes remote signaling change frequency and telemetry data refresh cycle; load data includes the real-time resource occupancy ratio of each service and the dynamic weight of service priority; and the update granularity of the digital twin data is synchronized with the acquisition time granularity of the equipment status data and service operation data. The equipment operating years in the equipment status data directly reflect the service time of the equipment and are a basic indicator for assessing the degree of equipment aging and potential failure risks. As the operating years increase, the equipment... The probability of failure usually increases. Performance degradation rate quantifies the rate at which equipment performance declines over time, such as signal strength attenuation and processing capacity reduction in communication modules. It is a key parameter for predicting equipment performance degradation. Link latency refers to the time required for data packets to travel from the source to the destination on a communication link. Excessive latency may indicate network congestion or link failure. Packet loss rate represents the proportion of data packets lost during data transmission out of the total number of data packets sent. It is an important indicator for measuring the reliability of communication links. A high packet loss rate usually means a decline in communication quality or link instability. Equipment remaining life prediction parameters combine historical data and prediction models to estimate how long the equipment can still operate normally, providing a basis for advance planning of maintenance and replacement. It is one of the core data for achieving early warning.

[0026] The remote signaling change frequency in the operational data indicates the frequency of changes in the status of switch signals such as circuit breaker status and protection action signals in the dispatch automation system. An abnormally high change frequency may indicate frequent equipment operation, system instability, or false alarms. The telemetry data refresh cycle reflects the frequency at which analog data such as voltage, current, and power are collected from the field and updated to the master station system. An abnormal extension or shortening of the refresh cycle may indicate communication link failure, abnormal acquisition equipment, or excessive system load. The real-time resource consumption ratio of each service in the load data provides the real-time consumption of CPU, memory, bandwidth, and other resources by various services in the current power communication network and dispatch automation system, which helps to identify resource bottlenecks and potential overload risks. The dynamic weight of service priorities reflects the importance of different dispatch services under the current power grid operating status. For example, in an emergency, the priority of certain key services will be dynamically increased, which is crucial for assessing the impact of faults and formulating response strategies.

[0027] Furthermore, the update granularity of digital twin data is synchronized with the collection time granularity of equipment status data and business operation data, ensuring that the digital twin model can reflect the operating status of the physical power grid in real time and accurately. This synchronization is the basis for conducting high-precision simulation and deduction, causal relationship analysis and fault impact assessment, avoiding data misalignment or information lag caused by inconsistent data timestamps, thereby ensuring the effectiveness of early warning and assessment.

[0028] Through the above technical solutions, more detailed and comprehensive equipment status, business operation, and load data are acquired during the data acquisition phase. This ensures that the digital twin data is synchronized with the acquisition time granularity of these key operational data. This specific and synchronized data provides high-quality input for subsequent feature processing steps, enabling causal inference algorithms to more accurately identify deep causal relationships between equipment degradation trends, abnormal business patterns, and changes in power grid operating conditions. For example, by using equipment remaining life prediction parameters, long-term performance degradation trends can be detected earlier; by using link latency and packet loss rate, the health status of the communication network can be monitored in real time; and the remote signaling change frequency and telemetry data refresh cycle directly reflect the real-time operational quality of scheduling services. The synchronized updating of digital twin data ensures the accuracy and real-time nature of simulations, making the calculation of fault impact levels and handling priorities more realistic. This significantly improves the accuracy, timeliness, and effectiveness of collaborative fault early warning and decision support.

[0029] In some of the embodiments described above in this application, it is proposed to collect multi-source data and perform time-series alignment and standardization processing, and then extract and fuse long-term causal features and real-time causal features for fault early warning. However, in practical applications, the data of power communication networks and dispatch automation systems are complex and interconnected. Simple data processing may not be able to accurately reveal the deep causal mechanisms behind the data, resulting in the extracted features not being able to fully reflect the real causes and evolution trends of potential faults, thereby affecting the accuracy and reliability of early warning.

[0030] In the feature processing step: the causal inference algorithm adopts the causal forest algorithm; the extracted causal relationships include at least the causal relationship between the decay of the remaining lifetime prediction parameters and the decline in the link performance of the power communication network, and the causal relationship between the changes in power grid operating conditions reflected by the digital twin data and the fluctuation of the dynamic weight of business priority; based on the strength of the extracted causal relationships, the fusion weights of long-term causal features and real-time causal features in the subsequent model input are dynamically allocated.

[0031] Causal forest algorithm is a nonparametric machine learning method based on decision trees, specifically designed to estimate heterogeneous processing effects and identify complex causal relationships. In the collaborative fault early warning scenario of power communication networks and dispatch automation systems, this algorithm can effectively handle high-dimensional, nonlinear data with confounding variables. By constructing multiple causal trees and performing ensemble learning, it accurately identifies potential, nonlinear causal dependencies between equipment status data, business operation data, load data, and digital twin data, rather than just statistical correlations. Its core lies in using the idea of ​​random forests, where, when splitting at each node of the decision tree, it not only considers predictive performance but also focuses on maximizing the heterogeneity of causal effects, thereby more accurately quantifying the causal strength and direction between different variables.

[0032] Among the extracted causal relationships, the causal relationship between the decay of remaining lifetime prediction parameters and the decline in link performance of power communication networks reveals the intrinsic link between the performance degradation caused by long-term operation of equipment in power communication networks and the quality of service of network links. The decay of remaining lifetime prediction parameters directly reflects the aging degree or deterioration of the health status of equipment hardware, while the increase in link latency, the rise in packet loss rate, and the decline in link performance directly affect the reliability and efficiency of communication networks. Identifying and quantifying this causal relationship helps to predict and take intervention measures in advance before the equipment reaches the end of its physical lifespan, avoid communication interruptions or performance bottlenecks caused by equipment failure, and thus achieve proactive equipment maintenance and network optimization.

[0033] Digital twin data reflects the causal relationship between changes in power grid operating conditions and dynamic weight fluctuations in business priorities. It focuses on how real-time changes in the power grid's operating status affect the importance ranking of various services in the dispatch automation system. Digital twin data can reflect key operating condition information of the power grid, such as voltage, current, frequency, and load distribution, in real time and with high fidelity. When power grid operating conditions change, such as peak load, line tripping, or fluctuations in new energy grid connection, the priority of certain services in the dispatch automation system, such as relay protection and safety and stability control, may be dynamically increased to ensure the safe and stable operation of the power grid. Identifying this causal relationship enables the early warning system to better understand the dynamic coupling between the real-time operating needs of the power grid and the priority of dispatch services. Thus, in early warning and response decisions, priority is given to ensuring the normal operation of key services and avoiding the impact of secondary service failures on the overall safety of the power grid.

[0034] Based on this, and according to the strength of the extracted causal relationships, the fusion weights of long-term causal features and real-time causal features are dynamically allocated in subsequent model inputs. This dynamic allocation of fusion weights refers to adjusting the relative importance of long-term and real-time causal features in real time when input into a multivariate time series prediction model based on the strength of different causal relationships quantified by the causal forest algorithm. For example, when the decay intensity of the remaining lifetime prediction parameter of a certain device is high, it indicates a significant long-term deterioration trend, and the weight of long-term causal features can be appropriately increased. Conversely, when the digital twin data shows drastic changes in the power grid operating conditions and a strong causal relationship with the dynamic weight fluctuations of business priorities, the weight of real-time causal features should be increased. This dynamic adjustment mechanism ensures that the early warning model can adaptively focus on the most relevant causal features according to the current system state and the nature of potential faults, thereby improving the accuracy and pertinence of the early warning.

[0035] By employing the causal forest algorithm, this application delves into the complex causal relationships between multi-source heterogeneous data from power communication networks and dispatch automation systems, rather than merely focusing on superficial statistical correlations. It identifies the causal relationship between the decay of remaining lifetime prediction parameters and the performance degradation of power communication network links, enabling the system to predict long-term equipment degradation trends earlier and more accurately, providing a scientific basis for preventative maintenance. By revealing the causal relationship between changes in power grid operating conditions reflected in digital twin data and the dynamic weight fluctuations of business priorities, the early warning system can perceive the impact of the power grid operating environment on the criticality of dispatch operations in real time. Therefore, in early warning and subsequent handling, priority is given to services crucial to the safety and stability of the power grid. Furthermore, based on the strength of the extracted causal relationships, the fusion weights of long-term and real-time causal features are dynamically allocated. This allows the early warning model to adaptively adjust its focus on different types of features according to the current system state and the nature of potential faults. This mechanism ensures that the features input into the multivariate time series prediction model are more targeted and effective, significantly improving the accuracy, timeliness, and interpretability of early warning information, thus providing a more solid and intelligent foundation for collaborative fault early warning of power communication networks and dispatch automation systems.

[0036] In some of the embodiments described above in this application, after data acquisition and feature processing, long-term causal features and real-time causal features are input into a multivariate time series prediction model to output early warning information. However, in practical applications, the fault modes of power communication networks and dispatch automation systems are complex and varied, involving multi-source heterogeneous data and long-term and short-term dependencies. If a general or single prediction model is used, it may be difficult to accurately capture these complex time-series dynamics and causal relationships, resulting in insufficient accuracy and operability of the early warning information, and making it difficult to effectively guide subsequent fault assessment and handling.

[0037] This application further proposes that in the above-mentioned early warning steps, the multivariate time series prediction model is an LSTM-Transformer fusion model based on an improved attention mechanism; the early warning information includes at least the probability of fault occurrence, the expected time of occurrence, the scope of affected services, and the fault causal chain identifier traced back to specific equipment or power grid conditions.

[0038] Multivariate time series forecasting models aim to process multiple interrelated time series data and predict future trends or events based on historical data. In the collaborative fault early warning scenario of power communication networks and dispatch automation systems, it is necessary to consider multiple data such as equipment status data, business operation data, load data, and digital twin data. These data have complex temporal dependencies and mutual influences. Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network that excels at processing and predicting long- and short-term dependencies in time series, effectively learning and remembering information over longer time steps. Transformer models, through their self-attention mechanism, can process all elements in the sequence in parallel and capture the dependencies between any two positions in the sequence, especially excelling in handling long-distance dependencies and possessing strong parallel computing power. The computing power is enhanced by an improved LSTM-Transformer fusion model based on an attention mechanism. This model combines the advantages of LSTM in capturing local temporal features with the ability of Transformer in handling global long-distance dependencies. The attention mechanism allows the model to dynamically focus on the most important parts of the input sequence during prediction, thereby improving the ability to identify key features and the accuracy of prediction. For example, in the fusion model, the LSTM part can be responsible for extracting local temporal dependency patterns, while the Transformer part can use the attention mechanism to capture the global correlation between different time steps and different variables, such as the potential correlation between the historical performance degradation trend of a device and the current peak load of the power grid. This fusion approach can more comprehensively and accurately understand complex causal features, thereby improving the performance of the prediction model.

[0039] The advanced early warning information is further refined, including at least the probability of failure, the expected time of occurrence, the scope of affected services, and a fault causal chain identifier tracing the fault back to specific equipment or power grid conditions. The probability of failure refers to the likelihood of a failure occurring within a future time window, as output by the prediction model, usually expressed as a percentage or a value between 0 and 1, providing maintenance personnel with a quantitative basis for risk assessment. The expected time of occurrence refers to the most likely time point or time range for the failure, given by the prediction model, enabling maintenance personnel to plan maintenance, resource allocation, or service switching measures in advance, achieving preventative maintenance. The scope of affected services refers to the types, volumes, or specific service instances that may be affected once a failure occurs, such as the transmission of remote control or telemetry data, or power distribution in a specific area. Automated operations facilitate rapid assessment of the breadth and depth of a fault's impact and prioritize the continuity of critical business operations. A fault causal chain identifier, tracing the source to specific equipment or power grid conditions, is crucial diagnostic information. It not only indicates the possibility of a fault but also reveals the potential causes and evolution paths leading to it. For example, the identifier might indicate that wear on the first gear 100 leads to abnormal load on the first motor 200, potentially causing a communication link interruption. Alternatively, persistently high power grid load in a certain area might cause a communication module to overheat, ultimately affecting data transmission stability. This causal chain identifier is generated based on causal relationships extracted during feature processing. It concretizes abstract warning information to physical equipment or specific power grid operating conditions, providing a direct basis for accurate fault location and targeted handling.

[0040] By employing an improved LSTM-Transformer fusion model based on an attention mechanism as a multivariate time series prediction model, this application can more effectively process multi-source heterogeneous data in power communication networks and dispatch automation systems, and capture complex long-term and short-term time series dependencies and causal relationships. This fusion model combines the advantages of LSTM in sequence modeling with the ability of Transformer in capturing global dependencies, and dynamically focuses key information through an attention mechanism, significantly improving the depth of understanding and prediction accuracy of long-term and real-time causal features. Furthermore, the advanced warning information output by this application not only includes the probability of fault occurrence and the expected time of occurrence, but also provides information on the affected parties. The detailed and diagnostic early warning information, including the scope of the affected business and the fault causal chain identification tracing back to specific equipment or power grid operating conditions, enables operation and maintenance personnel to clearly understand the risk level, timing, potential impact, and root cause of the fault. For example, through the fault causal chain identification, the specific equipment or abnormal operating condition causing the fault can be quickly located, thereby transforming the traditional passive response into proactive prevention. This greatly improves the accuracy, timeliness, and operability of fault early warning, providing a solid foundation for subsequent fault assessment, priority determination of handling, and generation of targeted handling instructions. It effectively avoids handling delays or resource waste caused by ambiguous early warning information, thereby ensuring the safe and stable operation of the power system.

[0041] In some of the above implementations, although it is possible to output early warning information through multivariate time series prediction models and identify the causal chain identifiers of potential faults, it is difficult to intuitively and quantitatively assess the specific impact of faults on scheduling services based solely on this information, nor can it effectively determine the priority of handling the fault, which may lead to delayed response or improper resource allocation.

[0042] This application further proposes to use a digital twin for simulation and deduction in the evaluation step. Specifically, based on the fault causal chain identifier, the equipment or operating condition parameters corresponding to the fault causal chain identifier are set to a fault state in the digital twin, and simulation and deduction are performed to obtain the evolution trajectory of the long-term causal characteristics and real-time causal characteristics; the simulated trajectory is compared with the current actual trajectory, and the difference is used to dynamically calculate the impact level; the calculation of the handling priority is further combined with the dynamic weight of the business priority obtained from the load data and the difference; the impact level is quantified into three levels: general, relatively severe, and serious according to a preset threshold.

[0043] Upon receiving the advanced warning information output from the early warning step, especially the fault causal chain identifier contained therein, this application utilizes a digital twin for simulation and deduction. The fault causal chain identifier clearly indicates the root cause of the potential fault and its propagation path in the power communication network and dispatch automation system. The digital twin is a virtual mapping of the physical power grid and communication network, capable of reflecting its operating status and topology in real time. To simulate the system behavior after a fault occurs, this application sets the corresponding fault dependent variable in the digital twin to a fault state or expected deterioration state based on the fault point or affected equipment indicated by the fault causal chain identifier. Subsequently, the digital twin, based on its built-in physical model, behavioral model, and business logic, simulates the impact of these fault dependent variables on the operation of the entire system over a period of time. This allows it to deduce the long-term causal characteristics such as the trend of equipment performance degradation and the evolution trajectory of real-time causal characteristics such as link latency, packet loss rate, and service resource consumption. For example, if the fault causal chain identifier indicates that a critical communication device is about to fail, the communication capability of that device is simulated as a decrease or interruption in the digital twin, and its impact on the real-time causal characteristics of the services it carries, such as the telemetry data refresh cycle, and the long-term causal characteristics such as the performance degradation rate of related links, is observed.

[0044] After simulating the evolution trajectory of long-term and real-time causal characteristics after a fault occurs in a digital twin, this application compares these simulated trajectories with the actual trajectories reflected by the actual operating data of the current power communication network and dispatch automation system. The actual trajectory represents the characteristic evolution trend of the system under fault-free or normal operating conditions. By comparing the simulated trajectory with the actual trajectory, the degree of difference between the two can be quantified, i.e., the degree of difference. This degree of difference reflects the extent to which the system performance will deviate from the normal state once a potential fault occurs. The greater the degree of difference, the more significant the impact of the fault on the system operation. For example, the degree of difference can be obtained by calculating the mean square error, dynamic time warping distance, or percentage deviation at a specific time point between the simulated trajectory and the actual trajectory on key performance indicators. This degree of difference is dynamically used to calculate the impact level of the fault on dispatch services, ensuring that the evaluation results are closely related to the current system state.

[0045] To more comprehensively assess the urgency and importance of a fault, this application, when calculating the handling priority, not only considers the degree of difference obtained through simulation, but also further incorporates the dynamic weights of service priorities obtained from load data. For example, the load data on the real-time resource occupancy percentage of each service and the dynamic weights of service priorities reflect the importance of current power grid dispatching services and their real-time resource demands. For instance, during power grid emergency operations or peak load periods, the priority of certain dispatching services will significantly increase. Therefore, the calculation of handling priority comprehensively considers the impact of the potential degree of difference in the fault and the dynamic weights of the affected service priorities. When the degree of difference is high and affects high-priority services, the handling priority will be significantly increased; conversely, if the degree of difference is low or only affects low-priority services, the handling priority will be relatively low. This combined approach ensures that the handling decision can take into account both the severity of the fault and the criticality of the service, achieving optimal resource allocation.

[0046] To make the assessment of fault impact more operational and intuitive, this application quantifies the degree of continuous impact obtained through difference calculation into discrete, easily understandable levels based on preset thresholds. The impact levels are quantified into three categories: general, moderate, and severe. These thresholds are pre-set based on historical fault data, expert experience, and the system's capacity to withstand different levels of faults. For example, when the difference is below a certain threshold T1, it is considered a general impact; when the difference is between T1 and T2, it is considered a moderate impact; and when the difference is above T2, it is considered a severe impact. Specifically, a difference ≤ 0.3 is quantified as a general level, 0.3 < difference ≤ 0.7 as a moderate level, and a difference > 0.7 as a severe level. This quantification process allows maintenance personnel to quickly determine the severity of the fault and take corresponding measures based on different levels, avoiding direct interpretation of complex continuous indicators and improving decision-making efficiency.

[0047] Through the above technical solution, this application solves the problem that it is difficult to quantify the impact of faults and determine the priority of handling based solely on early warning information. By using digital twins to simulate and extrapolate the fault causal chain, the evolution trajectory of system performance after a fault occurs can be intuitively predicted and compared with the actual trajectory, thereby dynamically calculating the degree of difference in the impact of the fault on system operation. On this basis, dynamic weights of business priorities are further combined, making the calculation of handling priorities more comprehensive and accurate. It not only considers the severity of the fault itself, but also takes into account the criticality of the affected business. Finally, the impact level is quantified into three levels: general, relatively severe, and severe, providing clear and operable decision-making basis for operation and maintenance personnel. This assessment method makes the early warning information more instructive, effectively avoids response delays or resource misallocation, and ensures that efficient handling strategies can be formulated and implemented in a targeted manner before potential faults occur, thereby significantly improving the operational reliability and security of power communication networks and dispatch automation systems.

[0048] In some of the embodiments described above in this application, although steps are proposed to generate and execute disposal instructions based on advance warning information, impact level, and disposal priority, verify the disposal effect through a digital twin, and optimize relevant parameters based on the verification results, in actual operation, how to ensure that the generated disposal instructions are highly targeted and operable, and how to efficiently and accurately feed back the verification results of the disposal effect to the preceding feature processing and early warning assessment stages to achieve continuous adaptive optimization of the entire early warning system, are still key aspects that need to be further clarified and improved.

[0049] This application further proposes that, in the processing and optimization steps, the handling instructions are generated based on the fault causal chain identifier and handling priority, including targeted equipment pre-maintenance, dynamic adjustment of link bandwidth, or high-priority service routing switching plans; the handling effect is verified through a digital twin by comparing whether the feature indicators corresponding to the fault causal chain identifier before and after handling have recovered to the normal threshold; feedback optimization includes at least feeding the verification results back to the feature processing step to optimize the allocation rules of the input feature set and fusion weights of the causal forest algorithm.

[0050] Disposal instructions are specific action plans taken to address predicted faults. Their generation process is highly dependent on the fault causal chain identifier and disposal priority. The fault causal chain identifier provides the root cause and propagation path of the fault, enabling disposal instructions to directly address the core problem. Disposal priority determines the urgency of the action and the degree of resource allocation. For example, when the fault causal chain identifier points to the performance degradation of a specific device, the generated disposal instruction may be a pre-maintenance plan for that device, such as replacing aging parts or performing preventative maintenance. If the fault causal chain identifier reveals the potential congestion risk of a communication link, the disposal instruction may be a dynamic adjustment of the link bandwidth to ensure the communication quality of critical services. When high-priority services face the risk of interruption, the system will generate a high-priority service routing switchover plan to quickly switch service traffic to an alternative path, thereby ensuring service continuity. These targeted disposal instructions ensure the accuracy and effectiveness of intervention measures.

[0051] To objectively evaluate the effectiveness of the handling instructions after their execution, this application utilizes a digital twin for verification. The verification process involves comparing whether the characteristic indicators corresponding to the fault causal chain identifier before and after the handling have returned to normal thresholds. For example, if the fault causal chain identifier indicates an abnormally high packet loss rate on a certain communication link, and the handling instruction is to adjust the link bandwidth, then the verification will focus on monitoring the packet loss rate of that link. If, after handling, the packet loss rate of that link falls back to the preset normal threshold range in the digital twin simulation environment or in actual monitoring, then the handling is deemed effective. This quantitative comparison based on key characteristic indicators provides a clear basis for evaluating the handling effect and avoids the bias of subjective judgment.

[0052] Furthermore, this application proposes a feedback optimization mechanism, which includes at least feeding back the verification results to the feature processing step to optimize the allocation rules of the input feature set and fusion weights of the causal forest algorithm. When the treatment effect is verified to be effective, it indicates that the current feature processing and early warning model is accurate and the relevant parameters can be strengthened. When the treatment effect is poor, it prompts the system to make adjustments. For example, if a treatment fails to effectively solve the problem, the system will feed back this verification result to the feature processing step. This feedback information can be used to adjust the input feature set of the causal forest algorithm, for example, by adding new features that may be ignored, or by adjusting the weights of existing features to improve the accuracy of causal relationship analysis. The feedback results can also be used to optimize the fusion weight allocation rules of long-term causal features and real-time causal features in subsequent model inputs, so that the model can better adapt to different power grid operating conditions and fault modes, thereby continuously improving the prediction accuracy and the intelligence level of the treatment strategy of the entire early warning system.

[0053] Through the above technical solutions, the generation of handling instructions is no longer general but can accurately generate targeted plans for equipment pre-maintenance, dynamic adjustment of link bandwidth, or high-priority service routing switching based on the fault causal chain identifier and handling priority. This greatly improves the efficiency and effectiveness of fault handling. Digital twins are used to quantitatively verify the handling effect. By comparing whether the characteristic indicators corresponding to the fault causal chain identifier before and after handling have recovered to the normal threshold, objective and measurable data support is provided for the handling effect. More importantly, the verification results are fed back to the feature processing step to optimize the allocation rules of the input feature set and fusion weight of the causal forest algorithm. A closed-loop adaptive optimization mechanism is constructed, which enables the entire collaborative fault early warning method to continuously learn and improve, and continuously improve its prediction accuracy and the intelligence level of the handling strategy. This effectively solves the technical problems of insufficient targeting of handling instructions and difficulty in continuous system optimization.

[0054] In some of the embodiments described above in this application, a digital twin is proposed to verify the handling effect. This is achieved by comparing whether the characteristic indicators corresponding to the fault causal chain identifier before and after the handling have recovered to the normal threshold. However, in the actual operation of power communication networks and dispatch automation systems, the power grid conditions and business loads are dynamic. Simple fixed thresholds may not accurately reflect the handling effect, leading to misjudgment or failure to fully assess the true state of system recovery. Especially in complex and ever-changing operating environments, a more refined verification mechanism is needed.

[0055] This application further proposes to verify the handling effect through a digital twin, specifically including: dynamically simulating and generating an adaptive recovery threshold curve of the characteristic indicators corresponding to the fault causal chain identifier in the digital twin under a fault-free state based on the fault causal chain identifier and historical data of characteristic indicators before handling; comparing the actual characteristic indicator sequence after handling with the adaptive recovery threshold curve, and if the actual characteristic indicator sequence continuously falls within the envelope of the adaptive recovery threshold curve within a preset time, the handling is deemed effective; the generation of the adaptive recovery threshold curve integrates the current power grid operating conditions, business load, and recovery modes of similar historical faults.

[0056] This verification mechanism first uses a digital twin to perform refined modeling of the characteristic indicators associated with specific fault causal chain identifiers, aiming to overcome the limitations of traditional fixed threshold verification methods. Based on the fault causal chain identifiers identified in the early warning information, the system determines the core characteristic indicators that need attention. Subsequently, using historical data before the intervention, combined with the digital twin's accurate simulation capability of the power grid's operating state, it dynamically extrapolates the recovery trajectory that these characteristic indicators should have under fault-free conditions. This fault-free state does not refer to an ideal static value, but rather considers the natural recovery trend of the system under normal operating fluctuations. For example, for the link delay indicator, its recovery under fault-free conditions may be a gradual convergence process, rather than instantaneously reaching a certain fixed value. The digital twin can simulate the system behavior under various operating conditions, thereby generating a curve that reflects this dynamic recovery process.

[0057] This step is the core of verifying the effectiveness of the treatment. After the treatment command is executed, the system will continuously collect actual characteristic index data related to the fault causal chain identifier to form a time series. This actual characteristic index series is then input into the comparison module and compared with the aforementioned dynamically generated adaptive recovery threshold curve in real time or near real time. The comparison process can use various techniques, such as calculating the mean square error and correlation coefficient between the two curves, or using a sliding window to evaluate the degree of deviation between the actual sequence and the threshold curve. This comparison not only focuses on the final value, but also on the dynamic matching of the recovery process.

[0058] To ensure the accuracy and robustness of the judgment, this application introduces the concepts of preset time and envelope. The preset time refers to the time window required for the expected characteristic indicators of the system to recover to the normal state. For example, it can be set to several minutes or several hours based on historical experience or business requirements. The envelope refers to an acceptable range of fluctuation around the adaptive recovery threshold curve. It takes into account the unavoidable random fluctuations in the actual system operation. For example, it can be set to ±X% or ±Y standard deviations of the threshold curve. Only when the actual characteristic indicator sequence is stably within this dynamic envelope throughout the preset time is it considered to be a truly effective treatment, thereby avoiding misjudgments caused by instantaneous fluctuations or short-term improvements.

[0059] This technical solution further clarifies the source of the intelligence and accuracy of the adaptive recovery threshold curve. Currently, factors such as load level, power grid operating conditions with topology changes, and business loads such as data traffic and critical business priorities are real-time factors affecting the system's recovery speed and stability. By incorporating these factors into the curve generation process, the threshold curve can be made closer to the current actual operating environment. By integrating recovery patterns from similar historical faults, it is possible to learn from past experience. For example, how long does it typically take to recover from a certain type of fault under specific operating conditions, and what are the typical fluctuation ranges of indicators during the recovery process? This fusion of multi-dimensional information can be achieved through machine learning models or rule-based expert systems, ensuring that the generated threshold curve is both real-time adaptive and supported by historical experience.

[0060] Through the above technical solution, this application overcomes the limitations of traditional fixed threshold verification methods, achieving dynamic and adaptive evaluation of the handling effect. By dynamically simulating and generating an adaptive recovery threshold curve in a digital twin that considers the current power grid operating conditions, business load, and historical recovery modes of similar faults, and comparing it with the actual characteristic index sequence after handling, the true dynamic process of system recovery can be captured more accurately. Only when the actual characteristic index sequence continuously falls within the envelope of the adaptive recovery threshold curve within a preset time can the handling be deemed effective. This significantly improves the accuracy and robustness of handling effect verification, avoiding misjudgments caused by dynamic environmental changes or instantaneous fluctuations. This refined verification mechanism provides a more reliable basis for subsequent feedback optimization, thereby effectively improving the overall performance and reliability of the collaborative fault early warning method between the power communication network and the dispatch automation system. Although a scheme has been proposed to generate and execute disposal instructions based on advance warning information, impact level, and disposal priority, and to verify the disposal effect through a digital twin, in the actual fault disposal process, the power grid operating status may change dynamically. The initial disposal instructions may not be able to completely curb the deterioration trend of the fault, and there may even be a situation where the impact level is upgraded or the deterioration rate is accelerated. If the disposal strategy is not adjusted in time, the fault may be further expanded, affecting the stable operation of dispatch services.

[0061] This application further proposes that the processing and optimization steps also include a dynamic adjustment sub-step. This sub-step aims to continuously monitor the changing trend and deterioration rate of the impact level during the execution of the disposal instructions. If the impact level is detected to be upgraded or the deterioration rate exceeds a preset threshold, the contingency plan adjustment mechanism is automatically triggered. The contingency plan adjustment mechanism is based on real-time simulation of the digital twin, matching or combining the contingency plan library to generate an upgraded disposal instruction and push it out. If the impact level is not detected to be upgraded and the deterioration rate is within the normal range, the system determines that the current disposal instruction is valid, continues to execute the original instruction, and maintains the monitoring status.

[0062] The dynamic adjustment sub-step is a key step in tracking the effects of executed handling instructions and making dynamic interventions. Its core lies in the continuous assessment of the impact level of the fault and the real-time perception of the fault's deterioration trend.

[0063] Continuous monitoring of the changing trend and deterioration rate of the impact level refers to the system continuously acquiring the latest power grid operation data after the initial handling command is issued and executed, and combining it with digital twin simulations to update the assessment of the fault impact level in real time. The changing trend of the impact level can be judged by comparing the current assessed impact level with the historical assessed impact level, for example, determining whether it remains unchanged, decreases, or increases. The deterioration rate quantifies the speed at which the impact level escalates, for example, whether the impact level rapidly escalates from general to severe or serious within a unit of time. This monitoring can be achieved by setting periodic assessment tasks or by using streaming data processing technology to calculate and analyze the impact level and its rate of change in real time.

[0064] If the rate of escalation or deterioration of the impact level exceeds a preset threshold, the contingency plan adjustment mechanism will be automatically triggered. The escalation of the impact level refers to the change in the degree of impact of the fault on the scheduling business from a lower level to a higher level, such as from general to severe, or from severe to critical. The preset threshold is an upper limit set for the rate of deterioration. For example, if the magnitude or speed of the escalation of the impact level within a specific time window exceeds the threshold, it is considered that the fault is deteriorating rapidly. The automatic triggering mechanism ensures that when the system detects signs of fault deterioration, it can quickly start the subsequent adjustment process without manual intervention, thereby gaining valuable time for handling.

[0065] The contingency plan adjustment mechanism, based on real-time simulations using a digital twin, matches or combines data from the contingency plan library to generate upgraded handling instructions and push them out. This mechanism is an intelligent decision-making module for responding to dynamic changes in faults, where the real-time simulation capability of the digital twin plays a crucial role. It can simulate the potential effects of different handling schemes based on the current real-time state of the power grid and fault characteristics, predicting the evolution trajectory of the fault under different interventions, thus providing a scientific basis for decision-making. The contingency plan library stores multiple preset handling schemes for different fault types, impact levels, and deterioration rates. These schemes can be tailored to specific conditions. For operations such as pre-maintenance, switching, and load reduction of backup, links, or services, the contingency plan adjustment mechanism can intelligently match the single contingency plan that best suits the current situation from the contingency plan library based on the inference results of the digital twin, or intelligently combine multiple contingency plans to form a more comprehensive and powerful upgraded handling instruction. For example, when it is found that a single link bandwidth adjustment is insufficient to contain the fault, the system may combine and generate a comprehensive instruction that includes equipment pre-maintenance and high-priority service routing switching. Finally, the generated upgraded handling instruction will be pushed to the operation and maintenance personnel or the automated execution system in a timely manner to guide or execute new handling actions.

[0066] By introducing a dynamic adjustment sub-step for handling, this application can continuously monitor the changing trend and deterioration rate of the fault impact level during the execution of handling instructions. Once the impact level is detected to be escalating or the deterioration rate exceeds a preset threshold, the system can automatically trigger the contingency plan adjustment mechanism. Based on the real-time extrapolation capability of the digital twin, this mechanism can intelligently match or combine from the contingency plan library to generate upgraded handling instructions and push them in a timely manner. This effectively solves the problem that the initial handling instructions may not be able to completely curb the deterioration of the fault. This significantly improves the timeliness, pertinence, and effectiveness of fault handling, avoids further expansion of the fault, ensures the stable operation of the power communication network and dispatch automation system, and improves the overall resilience and reliability of the power grid.

[0067] In some of the above implementation methods, although multivariate time series prediction models can output early warning information and assess the impact of faults through digital twins, the prediction models may face dynamically changing power grid conditions and business loads in actual operation. This may cause the model's focus on long-term causal characteristics and real-time causal characteristics to not always maintain the best match with the actual situation. This may affect the accuracy and timeliness of the model's predictions in some complex or sudden scenarios, thereby reducing the effectiveness of the early warning.

[0068] The model coordination step aims to establish a feedback loop between early warning and assessment, enabling the early warning model to self-optimize and adjust based on assessment results during actual operation. By introducing the model coordination step, the system can shift from passive early warning to proactive adaptation, improving the overall intelligence and robustness of the early warning system. The difference degree calculated in the assessment step is the degree of deviation quantified by comparing the fault evolution trajectory simulated by the digital twin with the current actual power grid operation trajectory. This difference degree intuitively reflects the degree of consistency between the early warning information and the actual situation; the larger the value, the greater the deviation between the prediction and reality. In the model coordination step, this difference degree is given a new role, no longer merely as an output of the assessment result, but as a quantitative indicator used to guide the adjustment of the early warning model. As a real-time correction parameter, the difference degree can be continuously and dynamically input into the early warning model to reflect the latest changes in the power grid operating status. This real-time nature ensures the timeliness of model adjustment, enabling it to quickly respond to fluctuations or sudden events in the power grid operating conditions and avoid model-related errors. The feedback mechanism addresses the failure of early warnings due to model lag. It directly transmits the insights from the assessment phase to the core prediction model in the early warning phase. The aforementioned fusion model, namely the LSTM-Transformer fusion model improved based on the attention mechanism, is key to generating advanced early warning information. By feeding back the difference to the model, its internal mechanism can be adjusted according to actual performance, thereby improving its predictive ability. The online adjustment of the model's attention distribution to the aforementioned long-term causal features and real-time causal features refers to dynamically changing the model's weight allocation to long-term causal features and real-time causal features based on the real-time correction difference parameter during continuous model operation. When the difference is large, it may mean that the model is not sensitive enough to real-time changes. In this case, the attention weight for real-time causal features can be increased. Conversely, if the difference indicates that the model does not grasp the long-term trend well, the attention weight for long-term causal features can be appropriately increased. This adjustment does not require retraining the entire model, but rather achieves rapid adaptation and optimization of the model by fine-tuning the attention weights.

[0069] By introducing a model collaboration step, the difference calculated in the evaluation step is fed back to the fusion model in the early warning step as a real-time correction parameter. This parameter is used to adjust the model's attention distribution to long-term causal features and real-time causal features online. This application effectively solves the prediction bias problem that may occur in the early warning model in a dynamic operating environment. When the evaluation finds a large difference between the simulated trajectory of the digital twin and the actual trajectory, the difference can immediately serve as a signal to prompt the fusion model to dynamically adjust its internal attention mechanism. For example, when facing sudden and rapidly evolving faults, the model will increase its attention to real-time causal features, thereby capturing early signs of faults more quickly and accurately. When facing faults caused by long-term performance degradation, the model will focus more on the analysis of long-term causal features. This closed-loop feedback and adaptive adjustment mechanism enables the early warning model to continuously optimize its sensitivity to different types of causal features, significantly improving the accuracy, timeliness, and robustness of early warning information. This ensures that the early warning system can better adapt to the complex power grid operating environment, thus providing a more reliable foundation for subsequent fault assessment and handling.

[0070] In some of the embodiments described above in this application, the collaborative fault early warning method of power communication network and dispatch automation system involves multiple steps such as data acquisition, feature processing, early warning, evaluation, and processing and optimization. However, in actual power systems, data sources are widely distributed, the real-time requirements for early warning are high, and the evaluation and optimization process may require a large amount of computing resources and global data. If all steps are concentrated on a single computing entity, it may lead to data transmission delay and excessive computing load. Especially in local fault early warning scenarios that require rapid response, it is difficult to balance real-time performance and global optimization.

[0071] This application further proposes that the early warning step, assessment step, and processing and optimization step are all executed using an edge-cloud collaborative architecture. This edge-cloud collaborative architecture is a distributed computing paradigm that offloads some computing tasks from the centralized cloud to devices or nodes at the network edge. This architecture aims to balance real-time performance, bandwidth efficiency, and data security. By performing preliminary processing near the data source, it reduces data transmission volume and latency, while leveraging the powerful computing capabilities of the cloud for complex analysis and global optimization. The early warning step is executed on edge computing nodes deployed at the power plant / substation side. Edge computing nodes are computing devices close to the data source, such as servers or dedicated hardware deployed inside power plants or substations. These nodes typically have certain... With its computing and storage capabilities, the edge computing node can independently run some applications. In the power system, the edge computing node on the power plant side can directly obtain local equipment status data and business operation data, thereby achieving rapid response. The edge computing node runs a lightweight multivariate time series forecasting model to generate local early warning information. The lightweight model refers to a model with low computing resource requirements and a relatively simple model structure, but which can still effectively perform forecasting tasks. For example, a deep learning model or statistical model with fewer parameters and fast inference speed can be used. Its purpose is to quickly analyze local data under the limited computing resources of the edge node and generate preliminary early warning information for local areas or specific equipment to meet real-time requirements.

[0072] The evaluation, processing, and optimization steps are executed on the cloud-based main station system. This system is a centralized platform with powerful computing and storage capabilities, typically deployed in a data center. It aggregates data from various edge nodes and performs large-scale data processing, complex model training, and global decision-making. After receiving local early warning information uploaded by edge computing nodes, the cloud-based main station system uses full data and a high-fidelity digital twin for simulation and comprehensive evaluation. Full data refers to a complete dataset including historical data, cross-regional data, and various auxiliary information. This data is typically not fully accessible at edge nodes. High-fidelity data... A digital twin is a virtual model of a power system that accurately reflects the real-time status, topology, equipment parameters, and operational patterns of the physical power grid. It supports complex simulations and predictions. By utilizing these resources, the cloud-based master station system can conduct deeper and more comprehensive analysis of local early warning information, assess the global impact of potential faults, and formulate optimal handling strategies. Subsequently, the cloud-based master station system distributes the generated handling instructions to the corresponding edge computing nodes and maintenance terminals. These instructions are specific responses to the early warning faults, such as equipment pre-maintenance plans, link bandwidth adjustment schemes, or service routing switching contingency plans. These instructions are then distributed to the edge computing nodes. The computing nodes can coordinate the execution of local devices via edge nodes; when distributed to the operation and maintenance terminals, they can directly notify on-site operation and maintenance personnel to perform operations. By deploying early warning, assessment, processing, and optimization steps collaboratively via the edge and cloud, this effectively solves the problem of balancing real-time performance and global optimization in traditional centralized processing for power system fault early warning. Time-sensitive early warning steps are devolved to edge computing nodes at the plant / substation side, and lightweight models are used to quickly generate local advance warning information, significantly shortening the early warning response time. This allows for timely detection and initial response to local faults, improving the system's real-time perception and processing capabilities for emergencies. The evaluation, processing, and optimization steps, supported by massive computing resources and global data, are executed on the cloud-based main station system. The cloud can utilize full data and high-fidelity digital twins to perform more accurate and comprehensive simulations and evaluations, thereby generating more optimized and globally-oriented handling instructions. This edge-cloud collaborative working mode not only reduces the real-time computing pressure on the cloud-based main station system and optimizes network bandwidth utilization, but more importantly, it achieves an organic combination of rapid local response and in-depth global analysis. This ensures the efficiency, accuracy, and reliability of fault early warning in the power communication network and dispatch automation system, thereby improving the operational resilience of the entire power system.

[0073] In collaborative fault early warning methods for power communication networks and dispatch automation systems, although advanced fault warning and assessment can be achieved and disposal instructions can be generated through steps such as data acquisition, feature processing, early warning, evaluation, and processing and optimization, in actual operation, the power grid conditions are complex and changeable. Especially in special scenarios such as peak load or planned maintenance, the system's ability to perceive real-time risks and the demand for data support will increase significantly. If the early warning strategy and data acquisition focus are not adjusted in time, the early warning of sudden or rapidly evolving faults may be untimely or inaccurate, thereby affecting the disposal effect.

[0074] The scenario adaptation step aims to enable the collaborative fault early warning method to dynamically adjust its early warning strategy and data acquisition behavior according to the real-time operating scenario of the power grid. Its purpose is to improve the early warning accuracy and response speed of the system under specific high-risk operating conditions. This step continuously monitors the operating status of the power grid, identifies the specific operating scenario, and triggers corresponding strategy adjustments based on this scenario. Digital twin data provides real-time and comprehensive operating status information of the power grid, including but not limited to power grid topology, load distribution, equipment operating parameters, and environmental factors. Analysis of this data can accurately determine the current operating condition of the power grid. Peak load conditions refer to the power grid's operating conditions. When the load level reaches or approaches its design capacity limit, the pressure on the power grid equipment increases, and the probability and scope of failure may increase significantly. The system can determine this by monitoring whether key indicators such as total load, line power flow, and transformer load rate reflected in the digital twin data exceed preset thresholds. Planned maintenance conditions refer to the period when some equipment or lines in the power grid are temporarily out of service due to maintenance, upgrades, or other reasons. Under these conditions, the operating topology and power flow distribution of the power grid will change, the importance of the remaining operating equipment will become more prominent, and the tolerance for failure will decrease. The system can identify this by querying the dispatching plan system or analyzing the equipment status in the digital twin data.

[0075] When digital twin data indicates that the power grid is under peak load or planned maintenance conditions, the system will automatically increase the proportion of real-time causal features in the fusion weight. Real-time causal features refer to the causal features extracted in the feature processing step that characterize real-time operational correlations. They reflect the immediate changes and rapid evolution trends of the power grid's operating status. The fusion weight determines the relative importance of long-term causal features and real-time causal features in the input of the multivariate time series prediction model. When the system identifies high-risk conditions such as peak load or planned maintenance, the weight allocation module inside the early warning system will automatically adjust, increasing the weight value allocated to real-time causal features while correspondingly decreasing the weight value of long-term causal features. For example, a new weight ratio can be directly set according to the current operating conditions through a preset weight adjustment function or lookup table to ensure that the prediction model can more sensitively capture real-time and rapidly changing risk signals.

[0076] The system will also trigger enhanced acquisition of status data for critical link equipment. Critical link equipment refers to equipment whose operating status has a decisive impact on the overall stability of the power grid and the reliability of communication under specific operating conditions. Examples include backbone communication optical cables, core routers, communication equipment in important substations, and key transmission lines. Enhanced acquisition means that, based on the original data acquisition frequency, granularity, or scope, specific status data of these critical equipment will be collected more intensively, more precisely, or more comprehensively. When a high-risk operating condition is identified, the system will send an instruction to the data acquisition module, requiring it to increase the acquisition frequency of equipment status data for predefined or dynamically identified critical link equipment based on the current operating condition. Equipment status data generally includes link latency, packet loss rate, equipment temperature, current, voltage fluctuations, etc., or increases the types of parameters to be collected to obtain more timely and detailed operating data.

[0077] By introducing a scenario adaptation step, the collaborative fault early warning method of this application can dynamically adjust according to the actual operating conditions of the power grid. This effectively solves the problem that the early warning system may be untimely or inaccurate due to rigid strategies in high-risk scenarios such as peak load or planned maintenance. When digital twin data indicates that the power grid is in a high-risk condition, the system automatically increases the proportion of real-time causal features in the fusion weight, enabling the multivariate time series prediction model to focus more on the real-time operation correlation and rapid change trend of the power grid. This improves the ability to perceive sudden or rapidly evolving faults, triggers enhanced collection of status data of key link equipment, and provides more intensive and refined real-time data support for the early warning model, further improving the accuracy and timeliness of the early warning. This adaptive strategy adjustment ensures that the early warning system can maintain efficient and reliable operation under various complex operating conditions, significantly enhancing the robustness and practicality of power grid fault early warning.

[0078] In the coordinated operation of the monitoring power communication network and dispatch automation system, the communication link connecting an important substation and the master station is showing a trend of gradual performance degradation due to the fact that its core transmission equipment has been in operation for many years. This link carries the real-time remote control command transmission for key generator units in the region, and its stability directly affects the operational safety of the power grid. Traditional early warning systems can only provide minute-level warnings for instantaneous faults such as sudden link interruptions. For the gradual performance degradation caused by equipment aging, it is difficult to predict several hours in advance, thus missing the best opportunity for preventive maintenance. Existing fault impact assessment mechanisms are mostly statically set and cannot accurately assess the actual impact and handling priority of the same communication fault on dispatch operations based on dynamic operating conditions such as real-time load fluctuations and generator start-up and shutdown.

[0079] To address the aforementioned issues, the system first initiates a data acquisition step, which continuously collects heterogeneous data from multiple sources: Equipment status data of the power communication network: including the operating years, performance degradation rate, link delay, packet loss rate of the communication equipment of the important substation, and the predicted parameters of the remaining life of the equipment obtained by the prediction model.

[0080] Operational data of the dispatch automation system includes the remote signaling frequency and telemetry data refresh cycle for key generator units.

[0081] Load data: This includes the proportion of communication resources occupied by each service in real time and the dynamic weight of service priority that is dynamically adjusted according to the current power grid operation strategy.

[0082] Digital twin data reflects the real-time operating status of the power grid, such as current load levels, unit start-up and shutdown status, and line maintenance plans. The update granularity of this digital twin data is synchronized with the collection time granularity of equipment status data and business operation data, ensuring data consistency.

[0083] When digital twin data indicates that the power grid is about to enter a peak load period or that a planned maintenance operation is about to occur, the system will automatically increase the proportion of real-time causal features in the subsequent fusion weights and trigger enhanced collection of status data of key link equipment, such as increasing the collection frequency of link latency and packet loss rate from minutes to seconds to obtain more refined real-time data.

[0084] After data collection is completed, the feature processing step begins. The system first performs strict time-series alignment and standardization on the collected data to eliminate time deviations and dimensional differences between different data sources. Then, based on causal inference algorithms, specifically the causal forest algorithm, it deeply analyzes the complex causal relationships between equipment status data, business operation data, load data, and digital twin data.

[0085] For example, the system identified a significant causal relationship between the decay of the remaining lifetime prediction parameters of the communication equipment and the decline in the link performance of the power communication network, which characterizes the trend of long-term performance degradation of the equipment. The system also found that the digital twin data reflects a causal relationship between changes in grid operating conditions such as the arrival of peak loads and the dynamic weight fluctuations of dispatching service priorities, which reflects real-time operational correlations.

[0086] Based on the strength of the extracted causal relationships, the system dynamically allocates the fusion weights of long-term causal features such as the degradation trend of equipment remaining lifespan and real-time causal features such as sudden changes in link delay and changes in grid load in subsequent model inputs. For example, when the long-term degradation trend is obvious but the real-time fluctuations are not drastic, the weight of long-term causal features will be higher; while when the grid operating conditions change drastically, the weight of real-time causal features will be increased accordingly.

[0087] The next step is the early warning process. The system inputs the fused long-term causal features and real-time causal features into a lightweight multivariate time series prediction model deployed on the edge computing nodes of the plant. This model is an LSTM-Transformer fusion model based on an improved attention mechanism, which can effectively capture long-term dependencies and short-term correlations in time series.

[0088] The edge computing node outputs local early warning information, such as predicting that the probability of the communication link failing within the next 4 hours is 70%, and that the link latency will exceed the preset threshold after 3 hours and 15 minutes. The affected business scope includes the real-time power generation control business of PowerPlantB, and traces the failure to specific aging components of the communication equipment, generating a fault causal chain identifier.

[0089] The system also includes a model collaboration step, in which the difference calculated in the evaluation step is fed back to the fusion model in the early warning step as a real-time correction parameter. This is used to adjust the model’s attention distribution to long-term causal features and real-time causal features online, enabling the model to dynamically optimize its prediction focus based on actual operating conditions.

[0090] Edge computing nodes upload local early warning information to the cloud main station system. In the cloud, the evaluation process combines the early warning information with the current real-time power grid operation status data and uses a high-fidelity digital twin for simulation and deduction.

[0091] Based on the fault causal chain identifier, the corresponding fault dependent variable will be masked in the digital twin, and the evolution trajectory of long-term causal characteristics and real-time causal characteristics will be simulated over a period of time. Subsequently, the simulated trajectory will be compared with the current actual trajectory, and the degree of difference will be used to dynamically calculate the impact level of the fault on the scheduling business.

[0092] The calculation of handling priorities further combines the dynamic weight of business priorities obtained from load data with the aforementioned difference degree. The impact level is quantified into three levels: general, moderate, and severe based on preset thresholds. For example, during peak load periods, even a moderate increase in link latency may be assessed as a severe impact level because its potential impact on power generation control business is huge.

[0093] Based on the early warning information, impact level, and handling priority generated by the cloud master station system, the system generates handling instructions and issues them to the corresponding edge computing nodes and operation and maintenance terminals. These handling instructions are generated according to the fault causal chain identifier and handling priority. For example, targeted pre-maintenance such as replacing aging parts of equipment, dynamic adjustment of link bandwidth to reserve more bandwidth for high-priority services, or high-priority service routing switching plans such as switching power generation control services to backup links.

[0094] During the execution of the disposal instructions, the system continuously monitors the changing trend and deterioration rate of the impact level. If the impact level is detected to be upgraded or the deterioration rate exceeds the preset threshold, the contingency plan adjustment mechanism is automatically triggered. This mechanism is based on real-time simulation of the digital twin, matching or combining the contingency plan library to generate upgraded disposal instructions and push them out. For example, upgrading from replacing aging components to immediately switching the primary and backup links and urgently replacing components.

[0095] Verifying the effectiveness of handling through digital twins specifically includes: dynamically simulating and generating adaptive recovery threshold curves for the characteristic indicators corresponding to the fault causal chain identifiers in a fault-free state within the digital twin based on the fault causal chain identifiers and historical data of characteristic indicators before handling. The generation of this curve integrates the current power grid operating conditions, business load, and recovery modes of similar historical faults. The actual characteristic indicator sequence after handling is compared with the adaptive recovery threshold curve. If the actual characteristic indicator sequence continuously falls within the envelope of the adaptive recovery threshold curve within a preset time, the handling is deemed effective, and a feedback optimization process is triggered. If the actual characteristic indicator sequence fails to fall within the envelope within the preset time, or deteriorates again after a brief period of falling within it, the handling is deemed ineffective or partially effective. The system marks this result as an abnormal handling case, immediately triggering a dynamic adjustment sub-step for handling, re-evaluating, and generating new handling instructions. This case is also stored in a historical database for subsequent optimization of the early warning model and evaluation rules for special scenarios.

[0096] Based on the verification results, the system feeds back relevant parameters in the feature processing step and the early warning and evaluation steps. For example, the verification results are fed back to the feature processing step to optimize the allocation rules of the input feature set and fusion weights of the causal forest algorithm, thereby improving the accuracy of subsequent early warnings and the precision of evaluation.

[0097] Compared with existing technologies, this solution achieves advanced early warning of progressive collaborative faults caused by the aging of communication equipment by deeply integrating the full life cycle status data of equipment, the dynamic load characteristics of scheduling services, and the real-time power grid operation status of digital twins. Instead of being limited to minute-level warnings for instantaneous faults, it provides maintenance personnel with a preventive handling window of several hours. In addition, this solution uses digital twins for dynamic simulation and deduction, combined with real-time power grid operating conditions and dynamic weights of service priorities, to achieve dynamic and accurate quantification of fault impact levels and intelligent sorting of handling priorities. This overcomes the limitations of existing static impact matrices that cannot adapt to complex dynamic power grid environments, and significantly improves the rationality of maintenance resource allocation and handling efficiency.

Claims

1. A collaborative fault early warning method for power communication networks and dispatch automation systems, characterized in that, Includes the following steps: Data acquisition steps: Collect equipment status data of the power communication network, business operation data and load data of the dispatch automation system, and digital twin data reflecting the real-time operating status of the power grid; Feature processing steps: Time-series alignment of the data collected in the data acquisition step is performed to align them to the same time granularity, and normalization or standardization is applied. The normalization or standardization methods include maximum and minimum value normalization or Z-score standardization. Based on the causal inference algorithm, the causal relationship between equipment status data, business operation data, load data and digital twin data is analyzed, and long-term causal features for characterizing long-term performance degradation trends and real-time causal features for characterizing real-time operation correlations are extracted and fused. Early warning steps: Input the long-term causal features and real-time causal features into a multivariate time series prediction model to output early warning information; Assessment steps: Combining advanced early warning information with current real-time power grid operation status data, the digital twin is used to conduct simulation and dynamic calculation of the impact level and handling priority of the fault on dispatching services. Processing and optimization steps: Generate and execute disposal instructions based on the advance warning information, impact level and disposal priority, verify the disposal effect through a digital twin, and optimize the feature extraction and fusion strategy in the feature processing step, as well as the model prediction rules and evaluation rules in the warning step and evaluation step based on the verification results.

2. The collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 1, characterized in that, In the data acquisition step: The device status data includes link latency, packet loss rate, and device remaining lifetime prediction parameters. The load data includes the real-time resource usage percentage of each service and the dynamic weight of service priority. The digital twin data is updated in association with the timestamps of the device status data and business operation data collection.

3. The collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 2, characterized in that, In the feature processing step: The causal inference algorithm uses the causal forest algorithm; The extracted causal relationships include at least the causal relationship between the decay of the equipment remaining lifetime prediction parameters and the decline in the link performance of the power communication network, and the causal relationship between the changes in power grid operating conditions reflected by the digital twin data and the fluctuations in the dynamic weight of the service priority. Based on the strength of the extracted causal relationships, the fusion weights of the long-term causal features and real-time causal features are dynamically allocated in the subsequent model inputs.

4. The collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 3, characterized in that, In the aforementioned warning step: The multivariate time series prediction model is an LSTM-Transformer fusion model based on an improved attention mechanism; The advance warning information includes at least the probability of failure, the expected time of occurrence, the scope of affected services, and the fault causal chain identifier that traces the fault back to a specific device or power grid condition.

5. The collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 4, characterized in that, In the evaluation step: The simulation and deduction using a digital twin specifically involves: based on the fault causal chain identifier, setting the equipment or operating condition parameters corresponding to the fault causal chain identifier to a fault state in the digital twin, performing simulation and deduction, and obtaining the evolution trajectory of the long-term causal characteristics and real-time causal characteristics. The simulated trajectory is compared with the current actual trajectory, the degree of difference is calculated based on the comparison results, and the impact level is dynamically calculated based on the degree of difference. The processing priority is calculated based on the dynamic weight of the business priority obtained from the load data and the degree of difference; The impact level is quantified into three levels: general, moderate, and severe, based on a preset threshold.

6. A collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 5, characterized in that, In the processing and optimization steps: The handling instructions are generated based on the fault causal chain identifier and the handling priority, including equipment pre-maintenance, dynamic adjustment of link bandwidth, or high-priority service routing switching contingency plans. Verifying the effectiveness of the treatment through a digital twin is achieved by comparing whether the characteristic indicators corresponding to the fault causal chain identifier before and after the treatment have recovered to the normal threshold. The feedback optimization includes at least using the verification results to optimize the fusion weight allocation rule between the input feature set of the causal inference algorithm in the feature processing step and the long-term causal features and real-time causal features.

7. A collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 6, characterized in that, The verification of the treatment effect through digital twins specifically includes: Based on the fault causal chain identifier and historical data of characteristic indicators before handling, the threshold curve of the characteristic indicator corresponding to the fault causal chain identifier in the fault-free state is dynamically simulated and generated in the digital twin as an adaptive recovery threshold curve. The actual characteristic index sequence after treatment is compared with the adaptive recovery threshold curve. If the actual characteristic index sequence continuously falls within the envelope of the adaptive recovery threshold curve within a preset time, the treatment is deemed effective. The generation of the adaptive recovery threshold curve incorporates the current power grid operating conditions, business load, and recovery modes of similar historical faults.

8. A collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 6, characterized in that, The processing and optimization steps also include a dynamic adjustment sub-step: During the execution of the disposal instructions, the changing trend and deterioration rate of the impact level are continuously monitored; The deterioration rate is the magnitude of change in the impact level per unit time. If the impact level is detected to be upgraded or the deterioration rate exceeds a preset threshold, the contingency plan adjustment mechanism will be automatically triggered. The contingency plan adjustment mechanism is based on the real-time simulation of the digital twin, matching or combining contingency plan databases to generate upgraded handling instructions and push them out.

9. A collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 5, characterized in that, Also includes: In the model collaboration step, the difference degree calculated in the evaluation step is fed back as a real-time correction parameter to the fusion model in the early warning step, so as to adjust the model's attention distribution to the long-term causal features and the real-time causal features online.

10. A collaborative fault early warning method for power communication networks and dispatch automation systems according to claim 1, characterized in that, The warning, assessment, and processing and optimization steps are all executed using an edge-cloud collaborative architecture. The early warning step is executed on the edge computing node deployed on the plant side. The edge computing node runs a multivariate time series prediction model to generate local early warning information. The assessment and processing / optimization steps are executed on the cloud-based main station system. After receiving the local early warning information uploaded by the edge computing node, the cloud-based main station system uses full data and digital twins to perform simulation and assessment, and sends the generated handling instructions to the corresponding edge computing node and operation and maintenance terminal.

Images