A power system transient stability intelligent evaluation and early warning method based on multi-source data fusion

Abstract

This invention discloses an intelligent assessment and early warning method for power system transient stability based on multi-source data fusion. The method includes acquiring multimodal data during power system operation; performing fusion analysis on the multimodal data using a pre-defined transient stability assessment model to output transient stability discrimination results, stability margin indices, and confidence scores; the assessment model uses historical multimodal data of the power system as training samples and is trained using a neural network model trained by embedding a data-driven loss function through physical regularization constraints; the assessment model generates a causal evidence chain starting from the fault point through feature attribution analysis, and generates assessment results and graded early warning levels along with the causal evidence chain and confidence scores; sensitivity decoupling is performed on the causal evidence chain to locate core instability influencing factors and generate prevention and control strategies; and the credibility and accuracy of power system assessment and early warning are improved through physical regularization embedding and causal tracing assessment mechanisms.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, specifically to a method for intelligent assessment and early warning of transient stability of power systems based on multi-source data fusion. Background Technology

[0002] With the rapid development of large-scale grid connection of new energy sources and ultra-high voltage AC / DC hybrid power grids, intelligent assessment and early warning technology for power system transient stability based on multi-source data fusion has become a research hotspot. Various dispatching agencies are exploring more efficient online safety analysis methods in an attempt to improve the capacity for new energy absorption and reduce the risk of major power outages while ensuring the safe operation of the power grid. However, the following prominent problems still exist in the existing technology: the evaluation process lacks effective embedding of transient stability physical mechanisms such as the generator rotor motion equation and the equal area rule, which may lead to purely data-driven judgment results that violate physical common sense and the decision-making logic is untraceable; the early warning conclusion only outputs an instability label and cannot generate a causal transmission path from the fault event to the instability result, making it difficult for dispatchers to accept early warning information that lacks evidence; the mapping of evaluation results to early warning levels depends on static thresholds and lacks a dynamic grading mechanism based on confidence calibration, resulting in a coexistence of high false alarm and false alarm risks. Summary of the Invention

[0003] The purpose of this invention is to provide a method for intelligent assessment and early warning of transient stability in power systems based on multi-source data fusion. By fusing multimodal data and using a neural network model with embedded physical regularization constraints, transient stability discrimination and margin prediction are achieved. Feature attribution analysis is combined to generate a fault propagation causal evidence chain, sensitivity decoupling is used to locate the core instability influencing factors, and adaptive grading rules are used to generate graded early warning levels and match corresponding transient stability prevention and control strategies.

[0004] The objective of this invention can be achieved through the following technical solution: This application provides a method for intelligent assessment and early warning of transient stability of power systems based on multi-source data fusion, including; S1. Acquire multimodal data during the operation of the power system in real time; perform feature fusion and inference by combining the multimodal data with a preset transient stability model, and output transient stability discrimination results, stability margin index and confidence score; The preset transient stability model is a neural network model trained by embedding a data-driven loss function with physical regularization constraints, using historical multimodal data of the power system as training samples. S2. Based on the preset transient stability model, a fault transmission causal evidence chain starting from the fault point is generated through feature attribution analysis, and a transient stability assessment result is generated including the causal evidence chain, transient stability discrimination result, stability margin index and confidence score. Based on the transient stability assessment result, a graded early warning level is generated through preset adaptive grading rules. S3. Calculate the sensitivity of each power equipment node in the causal evidence chain to the stability margin index through sensitivity decoupling analysis, and locate the core instability influencing factors after sorting the sensitivity in descending order. S4. Combining the core instability influencing factors and graded early warning levels, a transient stability prevention and control strategy for the current power system operating condition is generated by matching the preset prevention and control strategy mapping table.

[0005] The beneficial effects of this invention are as follows: This invention embeds the generator rotor motion equation and transient energy function into the neural network loss function by using motion trajectory regularization term and energy consistency regularization term. This solves the technical problem that the evaluation results of pure data-driven models are prone to violate the physical common sense of power system and the decision-making lacks physical constraints. It ensures that the transient stability discrimination results and stability margin index strictly follow physical laws, thereby improving the evaluation credibility and generalization ability. By calculating the feature contribution rate using the integral gradient method and combining it with a step-by-step search of the power system topology, a fault propagation causal evidence chain starting from the fault point is automatically generated. This solves the technical problems of existing technology warning conclusions that only output instability labels, cannot provide fault propagation paths and decision-making basis, and are difficult for dispatchers to accept, thus achieving interpretability of the evaluation process. The confidence threshold and margin threshold are dynamically determined by the quantile method based on historical statistical distribution. Combined with transient stability discrimination results, confidence scores and stability margin indicators, a four-level adaptive early warning level is generated to reduce the risk of false alarms and missed alarms. By using sensitivity decoupling analysis to locate the core instability influencing factors, and matching the prevention and control strategy mapping table to generate precise prevention and control strategies, we can achieve physical reliability, causal traceability, hierarchical self-adaptation, and precise prevention and control for power system transient stability assessment and early warning. Attached Figure Description

[0006] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0007] Figure 1 A flowchart illustrating a method for intelligent assessment and early warning of transient stability in power systems based on multi-source data fusion, provided for this application; Figure 2 This application provides a schematic diagram of the transient stability model feature fusion and inference process for a power system transient stability intelligent assessment and early warning method based on multi-source data fusion. Detailed Implementation

[0008] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0009] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0010] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0011] Example 1, please refer to Figures 1 to 2 This embodiment provides a method for intelligent assessment and early warning of power system transient stability based on multi-source data fusion. By acquiring multimodal data during the operation of the power system in real time, it uses a neural network model with embedded physical regularization constraints to perform feature fusion and inference, and outputs transient stability discrimination results, stability margin index and confidence score. Based on feature attribution analysis, it generates a fault propagation causal evidence chain starting from the fault point, and generates a graded early warning level by combining adaptive grading rules. Through sensitivity decoupling analysis, it locates the core instability influencing factors, matches the prevention and control strategy mapping table to generate transient stability prevention and control strategies, and realizes the physical credibility, causal traceability, graded adaptation and precise prevention and control of power system transient stability assessment and early warning.

[0012] The specific steps of the intelligent assessment and early warning method for transient stability of power systems based on multi-source data fusion described in this embodiment are as follows: S1. Acquire multimodal data during the operation of the power system in real time; perform feature fusion and inference by combining the multimodal data with a preset transient stability model, and output transient stability discrimination results, stability margin index and confidence score; Specifically, multimodal data during the real-time operation of the power system is acquired, including: phasor measurement data, SCADA operation data, and fault recording data; the phasor measurement data includes voltage amplitude, voltage phase angle, active power, reactive power, and frequency at each generator terminal and important bus node; the SCADA operation data includes circuit breaker status, disconnector switch position, transformer tap position, line active and reactive power flow, generator active and reactive power output, and generator terminal voltage setpoint; the fault recording data includes instantaneous voltage and current waveforms before and after the fault. Furthermore, the training process of the preset transient stability model is as follows: Training sample construction: Multiple sets of multimodal data are extracted from historical power system operation data as training samples; each training sample includes: pre-fault operating conditions, fault type, fault location, fault clearing time, and corresponding actual transient stability results; Calculation of theoretical physical benchmark values: Based on the pre-fault operating conditions, fault type, fault location, and fault clearing time of each sample group, the theoretical power angle and theoretical speed of each generator rotor after the fault are calculated by numerical integration of the generator rotor motion equation. Simultaneously, the difference between the system transient energy and the critical transient energy at the fault clearing time is calculated using the transient energy function to obtain the theoretical transient energy margin. A positive theoretical transient energy margin indicates system stability, while a negative margin indicates system instability; the absolute value reflects the degree of stability. The theoretical power angle, theoretical speed, and theoretical transient energy margin are used as physical benchmarks for model training to constrain the physical consistency of the model output. Model network structure: The neural network model includes three parallel feature extraction branches and a fusion output module during the training phase. The fusion output module outputs transient stability prediction results, stability margin prediction values, and confidence scores, and also includes two auxiliary output heads for outputting power angle prediction values ​​and speed prediction values, respectively. After the transient stability model training is completed, the auxiliary output heads are deleted, and only the transient stability discrimination results, stability margin index, and confidence scores are retained for the inference phase. The total loss function of physical regularization constraints is constructed as a weighted sum of data-driven loss terms and physical regularization constraint terms. The data-driven loss term is calculated by the cross-entropy between the transient stability prediction results output by the neural network model for the training samples and the actual transient stability results corresponding to the training samples. The physical regularization constraint terms include a trajectory regularization term and an energy consistency regularization term. Specifically, the motion trajectory regularization term is constructed as follows: For each generator in each training sample, the power angle prediction sequence output by the neural network model during training is compared with the theoretical power angle sequence calculated by the generator rotor motion equation at each time point, and the mean square error between the two is calculated; similarly, the mean square error between the speed prediction sequence and the theoretical speed sequence is calculated; the mean square error of the power angle and speed of all generators is summed and averaged to obtain the motion trajectory regularization term; During training, multimodal data of each training sample is input into the neural network model. After forward propagation, the classification head calculates a two-dimensional probability vector using a normalized exponential function, where the sum of the stable class probability and the unstable class probability is one. The class corresponding to the larger probability value is taken as the transient stability prediction result. The cross-entropy of this prediction result and the actual transient stability result corresponding to the training sample is calculated to obtain the data-driven loss term. The specific calculation of the motion trajectory regularization term is as follows: based on the operating conditions before the fault, the fault type, the fault location, and the fault clearing time, the generator rotor motion equation is used for numerical integration to calculate the theoretical power angle and theoretical speed of each generator at each integration time point. The predicted power angle and predicted speed values ​​at the corresponding time points are obtained from the auxiliary output head. The mean square error between the predicted power angle and the theoretical power angle, as well as the mean square error between the predicted speed and the theoretical speed, are calculated for each generator. The mean square errors of the power angle and the mean square errors of the speed of all generators are summed and averaged to obtain the motion trajectory regularization term. Specifically, the energy consistency regularization term is constructed as follows: the stability margin prediction value output by the neural network model is compared with the theoretical transient energy margin calculated through the transient energy function, and the mean square error between the two is calculated; wherein, the physical meaning of the stability margin prediction value is consistent with that of the theoretical transient energy margin: a positive value indicates system stability, and the larger the absolute value, the higher the stability; a negative value indicates system instability, and the larger the absolute value, the more severe the instability; the energy consistency regularization term ensures that the stability margin index output by the model and the theoretical margin calculated by the transient energy function have consistent sign and dimension relationship; Specifically, the calculation of the energy consistency regularization term includes: based on the pre-fault operating conditions, fault type, fault location, and fault clearing time corresponding to the training samples, calculating the theoretical transient energy margin using the transient energy function at the fault clearing time; obtaining the stability margin prediction value from the regression head; calculating the mean square error between the stability margin prediction value and the theoretical transient energy margin to obtain the energy consistency regularization term; a positive theoretical transient energy margin indicates system stability, a negative value indicates system instability, and the absolute value reflects the degree of stability; Model parameter update: During training, the gradient of the total loss function with respect to the parameters of each layer of the neural network model is calculated using the backpropagation algorithm. The weights and bias parameters of the model are iteratively updated according to the gradient direction, so that the value of the total loss function gradually decreases. The above process is repeated until the preset number of iterations is reached, and the trained neural network model is output as a transiently stable model. Specifically, by embedding the above-mentioned physical regularization constraints into the training process, the problem of pure data-driven model output violating the physical common sense of power systems is solved; the motion trajectory regularization term forces the intermediate features learned by the model to conform to the motion law of the generator rotor, and the energy consistency regularization term ensures that the stability margin index of the model output is consistent with the theoretical margin calculated by the transient energy function, so that the model can still output physically reliable evaluation results under unknown operating conditions, thereby enhancing the model's generalization ability and decision reliability. As an example, a single-phase ground fault occurs on a transmission line. Real-time phasor measurement data shows that the voltage amplitude drops, data acquisition and monitoring control system data shows that the circuit breaker of the line trips, and fault waveform data shows that the peak fault current increases significantly. Furthermore, the process of feature fusion and inference using a pre-defined transient stability model combined with multimodal data is as follows: The preset transient stability model includes three parallel feature extraction branches: a graph neural network branch extracts the power grid topology spatial features from phasor measurement data; a temporal neural network branch extracts the temporal evolution features from SCADA operation data; and a convolutional neural network branch extracts the transient waveform features from fault recording data. The graph neural network branch maps voltage amplitude, voltage phase angle, active power, and reactive power from phasor measurement data to node features in the power grid topology graph, and transmission lines to edge features. It outputs power grid topology spatial features through message passing and aggregation in a graph convolutional network. The temporal neural network branch arranges SCADA operation data by time steps, extracts temporal dependencies through a long short-term memory network, and outputs temporal evolution features. The convolutional neural network branch treats instantaneous voltage and current waveforms from fault recording data as a one-dimensional time series, and outputs transient waveform features through a one-dimensional convolutional neural network and pooling operations. The three branches compute in parallel, outputting feature vectors with the same three dimensions. A weighted fusion feature vector is output by weighting and fusing the power grid topology spatial features, temporal evolution features, and transient waveform features through a trainable attention mechanism. Specifically, the following steps are implemented: Each of the power grid topology spatial features, temporal evolution features, and transient waveform features is configured with a set of learnable query parameters and key parameters; the inner product of each set of query parameters and key parameters is calculated to obtain the original attention score corresponding to each feature; each original attention score is input into a single-layer nonlinear mapping layer for transformation, and the transformed original attention scores are converted into three attention weights with a sum of 1 using a normalized exponential function; the power grid topology spatial features, temporal evolution features, and transient waveform features are multiplied by their respective attention weights to obtain a three-way weighted feature vector; the three-way weighted feature vectors are then summed element-wise to output a fusion feature vector with a unified dimension. A set of learnable query parameters and key parameters are configured for each of the power grid topology spatial features, temporal evolution features, and transient waveform features. The inner product of each set of query parameters and key parameters is calculated to obtain the original attention score corresponding to each feature. The original attention scores are input into a single-layer nonlinear mapping layer for linear transformation and nonlinear activation processing. The transformed original attention scores are converted into three attention weights with a sum of one through a normalized exponential function. The three feature vectors are multiplied by their respective attention weights to obtain three-way weighted feature vectors. The three-way weighted feature vectors are summed element-wise to output a fused feature vector with a unified dimension. The fused feature vector is processed by a normalized exponential function to output stable and unstable class probabilities. These probabilities are compared, and the class with the larger probability is identified as the transient stability determination result. The probability value corresponding to this class is used as the confidence score. Simultaneously, a stability margin index is obtained using a linear regression algorithm with the fused feature vector as input. The fused feature vector is then input into a classification head, which contains a fully connected layer followed by a normalized exponential function. The normalized exponential function outputs stable and unstable class probabilities, and their sum is one. The stable and unstable class probabilities are compared, and the class with the larger probability is identified as the transient stability determination result. The probability value corresponding to this class is used as the confidence score. As an example, after inputting the above multimodal data into the transient stability model, the classification head outputs a low probability of the stable category and a high probability of the unstable category. Therefore, the transient stability judgment result is the unstable category, and the confidence score is taken as the higher probability value. The regression head outputs a negative stability margin index, indicating that the system is in an unstable state. S2. Based on the preset transient stability model, a fault transmission causal evidence chain starting from the fault point is generated through feature attribution analysis, and a transient stability assessment result is generated including the causal evidence chain, transient stability discrimination result, stability margin index and confidence score. Based on the transient stability assessment result, a graded early warning level is generated through preset adaptive grading rules. Specifically, the preset transient stability model generates a fault propagation causal evidence chain starting from the fault point through feature attribution analysis, including: The phasor measurement data, SCADA operation data, and fault recording data in the multimodal data are used as the input feature set. Using the feature values ​​corresponding to the pre-fault steady-state operation data as the baseline, the contribution of each feature in the input feature set to the transient stability discrimination result is calculated using the integral gradient method. The integral gradient method integrates along a straight path from the baseline to the current input, calculating the average gradient of each feature to the model output, and fairly distributing the contribution. The sum of the contributions of all input features corresponding to each power equipment node is calculated as the node contribution. The sum of the contributions of all input features corresponding to each power equipment node to the transient stability discrimination result is calculated as the node contribution. Starting from the power equipment node corresponding to the fault point, the contributions are sorted from high to low. Combining the electrical connection relationships in the power system topology, adjacent equipment nodes with direct electrical connection to the current power equipment node and whose contributions exceed a preset contribution threshold are searched level by level, and these adjacent equipment nodes are included in the search path. An ordered sequence is formed expanding outward from the fault point, and this ordered sequence is used as the causal evidence chain for fault propagation. The ordered sequence is arranged in a hierarchical order expanding outward from the fault point. Equipment nodes within the same level are sorted from high to low contribution, including the equipment node, the name of the input feature with the highest contribution under the equipment node, and the contribution value corresponding to the input feature. Specifically, by generating a fault transmission causal evidence chain starting from the fault point, the problem that the early warning conclusion in the existing technology only outputs an instability label and cannot provide a causal transmission path is solved. The dispatcher can directly view the order in which the evidence chain extends from the fault node to the surrounding equipment nodes and the key sensitive features of each node, understand the physical propagation mechanism of transient instability, and improve the trust and acceptance rate of the model evaluation results. As an example, the evidence chain generated for a certain fault is as follows: The first layer (fault point) includes the bus node B1, where the input feature with the highest contribution is the voltage amplitude; the second layer includes the line node L1 and the generator node G1, where the input feature with the highest contribution under the line node L1 is the active power, and the input feature with the highest contribution under the generator node G1 is the power angle; the third layer includes the bus node B2, where the input feature with the highest contribution is the voltage phase angle. Furthermore, the process of generating graded early warning levels based on the transient stability assessment results and using preset adaptive grading rules includes: The preset adaptive grading rule is based on the statistical distribution of the stability margin index and confidence score output by the transient stability model under historical operating conditions. The confidence threshold, the first margin threshold and the second margin threshold are dynamically determined by the quantile method, and the first margin threshold is greater than the second margin threshold. Specifically, confidence scores of historical stable samples are collected, and preset quantiles are used as confidence thresholds; stability margin indices of historical unstable samples are collected. Since the stability margin indices of unstable samples are all negative, the upper quartile is used as the first margin threshold, and the lower quartile is used as the second margin threshold. The preset quantiles, upper quartiles, and lower quartiles are all statistics dynamically calculated based on the distribution of historical data and do not depend on fixed values. The confidence threshold is dynamically determined using the quantile method based on the statistical distribution of confidence scores for stable samples output by the transient stability model under historical operating conditions. The first margin threshold and the second margin threshold are dynamically determined using the quantile method based on the statistical distribution of stability margin indices for unstable samples output by the transient stability model under historical operating conditions, with the first margin threshold being greater than the second margin threshold. A level four warning is triggered when the transient stability determination result is in the stable category and the confidence score is less than the confidence threshold. When the transient stability determination result is in the unstable category, a level three warning is triggered if the stability margin index is greater than or equal to the first margin threshold, a level two warning is triggered if the stability margin index is greater than or equal to the second margin threshold and less than the first margin threshold, and a level one warning is triggered if the stability margin index is less than the second margin threshold. The classification rules are as follows: When the transient stability determination result is in the stable category and the confidence score is greater than or equal to the confidence threshold, there is no warning level; when the transient stability determination result is in the stable category and the confidence score is less than the confidence threshold, it corresponds to a level four warning level; when the transient stability determination result is in the unstable category and the stability margin index is greater than or equal to the first margin threshold, it corresponds to a level three warning level; when the transient stability determination result is in the unstable category and the stability margin index is greater than or equal to the second margin threshold and less than the first margin threshold, it corresponds to a level two warning level; when the transient stability determination result is in the unstable category and the stability margin index is less than the second margin threshold, it corresponds to a level one warning level. The priority of the tiered early warning levels, from highest to lowest, is as follows: Level 1, Level 2, Level 3, and Level 4. Specifically, by using a dynamic adaptive grading rule based on the quantile method, the problem of poor adaptability to changes in operating mode and the coexistence of false alarms and missed alarms in the traditional static threshold method is solved. The threshold is dynamically updated with the distribution of historical data, which can adapt to the model output characteristics under different seasons and different load levels, thereby effectively suppressing false alarms while ensuring the sensitivity of the early warning, and enabling dispatchers to allocate emergency resources reasonably according to the early warning level. In the example, the current dynamic threshold has a relatively high confidence threshold, a small negative absolute value for the first margin threshold, and a relatively large negative absolute value for the second margin threshold. If an assessment result indicates instability, and the stability margin index falls between the second and first margin thresholds, a level-two warning is triggered, and a corresponding warning signal is output. S3. Calculate the sensitivity of each power equipment node in the causal evidence chain to the stability margin index through sensitivity decoupling analysis, and locate the core instability influencing factors after sorting the sensitivity in descending order. Specifically, the sensitivity analysis for calculating the impact sensitivity of each power equipment node in the causal evidence chain on the stability margin index includes: Each power equipment node in the causal evidence chain is taken as the analysis object, and the input features corresponding to each power equipment node are extracted. The input features include the voltage amplitude, voltage phase angle, active power, reactive power, and power angle value of the power equipment node. The input features corresponding to each power equipment node are taken as independent variables, and the stability margin index is taken as the dependent variable. The first-order partial derivative of each independent variable with respect to the dependent variable is calculated by automatic differentiation. The first-order partial derivative reflects the instantaneous direction and intensity of the small change of the feature on the stability margin index. The absolute values ​​of all the first-order partial derivatives corresponding to each power equipment node are summed to output the sensitivity of each power equipment node to the stability margin index. Each power equipment node in the causal evidence chain is taken as the analysis object, and its voltage amplitude, voltage phase angle, active power, reactive power, and power angle value are extracted as input features. The stability margin index is taken as the dependent variable, and the first-order partial derivative of each input feature with respect to the stability margin index is calculated using the automatic differentiation method. The absolute values ​​of all first-order partial derivatives corresponding to the same power equipment node are summed to obtain the sensitivity of the node to the stability margin index. All power equipment nodes are sorted in descending order of their sensitivity. Nodes with sensitivity greater than a preset sensitivity threshold are extracted as core instability influencing factors. All power equipment nodes are sorted in descending order based on the impact sensitivity, generating a descending sequence of equipment nodes; the equipment nodes with impact sensitivity greater than a preset sensitivity threshold are extracted as core instability influencing factors. Specifically, by using sensitivity decoupling analysis to locate the core instability influencing factors, the key equipment nodes with the greatest impact on stability margin were further screened from the evidence chain. Not only was the instability propagation path given, but the contribution of each node in the path to stability margin was also quantified, enabling dispatchers to focus on the few most critical pieces of equipment and providing clear control targets for subsequent precise prevention and control. For example, a causal evidence chain contains multiple device nodes. The influence sensitivity of each node is obtained by automatically differentiating and calculating the sum of the absolute values ​​of the first-order partial derivatives of each node. A preset sensitivity threshold is taken as the average value of the sensitivity of all nodes. Nodes with sensitivity greater than this average value are extracted as core instability influencing factors. For example, a generator node and a line node are identified as core instability influencing factors. S4. Combining the core instability influencing factors and graded early warning levels, a transient stability prevention and control strategy for the current power system operating condition is generated by matching the preset prevention and control strategy mapping table. Specifically, the step of generating transient stability control strategies for the current power system operating conditions by matching a preset control strategy mapping table includes: The preset prevention and control strategy mapping table is pre-constructed based on transient stability control data under historical operating conditions. It includes multiple mapping entries, each recording a graded early warning level, a set of core instability influencing factors, and a set of corresponding transient stability control measures. The graded early warning level and core instability influencing factors are used as a joint matching index to search for the mapping entry corresponding to the joint matching index in the prevention and control strategy mapping table. When the mapping entry is found, the corresponding transient stability control measure is extracted from the mapping entry, and the transient stability control measure is used as the transient stability prevention and control strategy. When the mapping entry is not found, a default control measure is matched based on the graded early warning level. The default control measure is as follows: Level 1 early warning level corresponds to rapid load shedding through preset load reduction priority; Level 2 early warning level corresponds to generator power redistribution; Level 3 early warning level corresponds to switching on and off the dynamic reactive power compensation device; and Level 4 early warning level corresponds to issuing only an alarm. The preset load reduction priority is pre-configured based on the importance level of the load and the contribution of each load node to transient instability. Specifically, by jointly matching core instability influencing factors with graded early warning levels to generate differentiated prevention and control strategies, the system has achieved an improvement from general generator and load shedding to precise targeted control. The mapping table mechanism fully utilizes the interpretable information generated in the preceding steps, transforming early warning levels and dominant instability factors into specific control commands. This ensures transient stability while minimizing the impact on normal power supply, thereby improving the renewable energy absorption capacity and the economic efficiency of grid operation. For example, in a certain assessment, the result is a level 2 warning level, and the core instability influencing factor is generator power angle swing. The corresponding entry is found in the prevention and control strategy mapping table, and the control measure is extracted as generator power redistribution, specifically reducing the output of the generator with the leading power angle and increasing the output of the generator with the lagging power angle. If no matching entry is found, the generator power redistribution corresponding to the level 2 warning level in the default control measures will be implemented. Furthermore, after the transient stability control strategy is executed, system operation data is continuously collected. This data includes the actual value of the stability margin index, the change in the sensitivity of each power equipment node, and the result of whether the fault was successfully suppressed. This data is added to the historical operating condition database as input samples for the next threshold calculation in the adaptive grading rules. Based on the updated historical statistical distribution, the confidence threshold, the first margin threshold, and the second margin threshold are recalculated using the quantile method to achieve dynamic adaptive adjustment of the thresholds. Simultaneously, successfully matched entries and their execution effects are recorded in the control strategy mapping table. This is used to periodically optimize the transient stability control measure parameters in the mapping table, forming a complete closed loop from assessment, early warning, control to effect feedback, continuously improving the accuracy and adaptability of intelligent transient stability assessment and early warning. This embodiment embeds the generator rotor motion equation and transient energy function into the neural network loss function by using motion trajectory regularization and energy consistency regularization terms. This ensures that the transient stability discrimination results and stability margin indices strictly follow physical laws, solving the problem that pure data-driven models may violate physical common sense and improving the credibility and generalization ability of the evaluation results. By using the integral gradient method combined with the power system topology to search step by step to generate a fault propagation causal evidence chain starting from the fault point, it solves the problem that the early warning conclusions cannot trace the causal propagation path, enabling dispatchers to intuitively understand the instability propagation mechanism and improving the acceptability of early warning information. By dynamically determining the confidence threshold and margin threshold using the quantile method based on historical statistical distribution, and combining the transient stability discrimination results, confidence scores, and stability margin indices to generate a four-level adaptive early warning level, it solves the problem of high false alarm and false negative rates of static threshold methods and realizes adaptive dynamic calibration of early warning levels. By using sensitivity decoupling analysis to locate the core instability influencing factors and matching the prevention and control strategy mapping table, it generates a precise prevention and control strategy for the current operating conditions, realizing physical credibility, causal traceability, hierarchical adaptation, and precise prevention and control of power system transient stability assessment and early warning.

[0013] Example 2 This embodiment details the application of a power system transient stability intelligent assessment and early warning method based on multi-source data fusion in a power system dispatch center environment. The actual regional power system includes substations, thermal power units, wind farms, and photovoltaic power stations at multiple voltage levels, interconnected with the external power grid through multiple transmission lines. The dispatch center deploys an online security analysis platform, which includes synchronous phasor measurement units, data acquisition and monitoring control system servers, and fault recording devices deployed in each substation and power plant, as well as a data aggregation server, graphics processor server, and early warning display workstation located in the dispatch center. All devices are connected through the power dispatch data network. In the actual system, the data aggregation server of the dispatch center receives in real time the voltage amplitude, voltage phase angle, active power, reactive power, and frequency data uploaded by the synchronous phasor measurement unit; it also receives the circuit breaker status, disconnector switch position, transformer tap position, line power flow, and generator output data uploaded by the data acquisition and monitoring control system server; and it receives the instantaneous voltage and current waveforms uploaded by the fault recording device. All data is sent to the graphics processor server after high-precision time synchronization. Before model deployment, the transient stability model has been trained offline using historical data of the actual regional power system. In the training, the generator rotor motion equation is used to calculate the theoretical power angle and theoretical speed, and the transient energy function is used to calculate the theoretical transient energy margin, which is embedded as a physical regularization constraint loss function. After training, the transient stability model is deployed on the graphics processor server. On a certain day, a single-phase ground fault occurred in the actual system. The data aggregation server collected data in real time: the synchronous phasor measurement units on both sides of the fault point showed a significant drop in voltage amplitude; the data acquisition and monitoring control system showed that the circuit breaker of the line tripped; and the fault waveform recording device recorded a significant increase in the peak value of the fault current. The above data was input into the transient stability model. The graph neural network branch of the transient stability model extracted the topological spatial features of the power grid, the temporal neural network branch extracted the temporal evolution features of the data acquisition and monitoring control system, and the convolutional neural network branch extracted the fault waveform features. The attention mechanism performed weighted fusion of the three features. After fusion, the classification head output that the probability of the instability category is higher than the probability of the stable category, and it is determined to be an instability category. The corresponding probability value is used as the confidence score. The regression head outputs a negative stability margin index. Based on the transient stability model, the contribution of each input feature to the discrimination result is calculated using the integral gradient method; the feature value corresponding to the steady-state operation data before the fault is used as the baseline to calculate the contribution of each feature; the node contribution of each power equipment node is calculated, and starting from the power equipment node corresponding to the fault point, the adjacent equipment nodes with direct electrical connection are searched level by level in combination with the node contribution ranking and the preset contribution threshold to generate an ordered sequence that expands outward from the fault point as the fault transmission causal evidence chain; Specifically, the fault propagation causal evidence chain is arranged hierarchically: the first layer is the bus node where the fault point is located, and its highest contributing input feature is voltage amplitude; the second layer includes the faulty line node and the adjacent generator node, the former's highest contributing feature is active power, and the latter's highest contributing feature is power angle; the third layer includes the remote bus node, and its highest contributing feature is voltage phase angle; the fault propagation causal evidence chain is displayed in real time on the dispatching screen, allowing dispatchers to intuitively see the path of instability propagation from the fault point to the generator; The system adopts an adaptive grading rule, which is based on the statistical distribution of the stability margin index and confidence score output by the transient stability model under historical operating conditions. The system dynamically determines the confidence threshold, the first margin threshold, and the second margin threshold using the quantile method, and the first margin threshold is greater than the second margin threshold. When the transient stability determination result is in the stable category and the confidence score is greater than or equal to the confidence threshold, there is no warning level; when the transient stability determination result is in the stable category and the confidence score is less than the confidence threshold, a level four warning level is triggered; when the transient stability determination result is in the unstable category and the stability margin index is greater than or equal to the first margin threshold, a level three warning level is triggered; when the transient stability determination result is in the unstable category and the stability margin index is greater than or equal to the second margin threshold and less than the first margin threshold, a level two warning level is triggered; when the transient stability determination result is in the unstable category and the stability margin index is less than the second margin threshold, a level one warning level is triggered; the current assessment result is in the unstable category, and the stability margin index is between the second margin threshold and the first margin threshold, triggering a level two warning level; the warning display workstation issues a corresponding warning signal; Sensitivity decoupling analysis is performed on each power equipment node in the fault propagation causal evidence chain. The voltage amplitude, voltage phase angle, active power, reactive power, and power angle value of each node are extracted as input features. The first-order partial derivatives of the stability margin index with respect to each feature are automatically calculated using differentiation. The absolute values ​​of all partial derivatives for each node are summed to obtain the node's impact sensitivity. All nodes are sorted in descending order of their impact sensitivity. Nodes with impact sensitivity greater than a preset sensitivity threshold are extracted as core instability influencing factors. The analysis results show that the impact sensitivity of a certain generator node and a certain line node is significantly higher than that of other nodes, and they are identified as the core instability influencing factors. The system automatically marks the positions of these two devices in the power grid wiring diagram, providing clear targets for subsequent prevention and control. Using the graded early warning level (Level 2) and the core instability influencing factors (generator power angle swing and line overload) as a joint matching index, the prevention and control strategy mapping table is queried. The corresponding entries in the prevention and control strategy mapping table record corresponding transient stability control measures. The system automatically generates a specific strategy: adjusting generator power distribution, reducing the output of generators with leading power angles and increasing the output of generators with lagging power angles, and transferring part of the load from the faulty line to parallel lines. After confirmation by the dispatcher, the strategy is issued and executed through the automatic generation control system. After execution, the stability margin index changes from negative to positive, the system returns to stability, and a potential instability accident is avoided. The dispatch center continuously collects system operation data after execution, including the actual value of the stability margin index after it turns from negative to positive, the change in the sensitivity of the generator and line to the impact, and the result of successful fault suppression. This data is added to the historical operating condition database as input samples for the next adaptive hierarchical rule threshold calculation. Based on the updated historical statistical distribution, the confidence threshold, the first margin threshold, and the second margin threshold are recalculated using the quantile method to achieve dynamic adaptive adjustment of the thresholds. Simultaneously, the matching entries of the secondary warning level with generator power angle swing and line overload in the prevention and control strategy mapping table, along with their successful execution effects, are recorded for periodic optimization of control measure parameters under these entries. Thus, the method forms a complete closed loop from data collection, evaluation, early warning, prevention and control to effect feedback, continuously improving the accuracy and adaptability of transient stability intelligent assessment and early warning. This embodiment processed multiple fault warning events during the continuous operation of the actual regional power system. Compared with the original assessment system based on static thresholds, the accuracy of the warning is significantly improved, while the false alarm rate and missed alarm rate are greatly reduced. Through the fault propagation causal evidence chain and the sensitivity decoupling analysis, the decision-making time of dispatchers is significantly shortened. Through precise prevention and control strategies, unnecessary generator and load shedding operations are reduced, and the renewable energy absorption capacity is improved. This embodiment verifies the effectiveness of the method described in Embodiment 1 in the actual power system, realizing the physical reliability, causal traceability, hierarchical self-adaptation, and precise prevention and control of transient stability assessment and early warning. The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A power system transient stability intelligent evaluation and early warning method based on multi-source data fusion, characterized in that, Includes the following steps: S1. Acquire multimodal data during the operation of the power system in real time; perform feature fusion and inference by combining the multimodal data with a preset transient stability model, and output transient stability discrimination results, stability margin index and confidence score; The preset transient stability model is a neural network model trained by embedding a data-driven loss function with physical regularization constraints, using historical multimodal data of the power system as training samples. S2. Based on the preset transient stability model, a fault transmission causal evidence chain starting from the fault point is generated through feature attribution analysis, and a transient stability assessment result is generated including the causal evidence chain, transient stability discrimination result, stability margin index and confidence score. Based on the transient stability assessment result, a graded early warning level is generated through preset adaptive grading rules. S3. Calculate the sensitivity of each power equipment node in the causal evidence chain to the stability margin index through sensitivity decoupling analysis, and locate the core instability influencing factors after sorting the sensitivity in descending order. S4. Combining the core instability influencing factors and graded early warning levels, a transient stability prevention and control strategy for the current power system operating condition is generated by matching the preset prevention and control strategy mapping table. 2.The power system transient stability intelligent evaluation and early warning method based on multi-source data fusion of claim 1, wherein, The process of training a neural network model by embedding a data-driven loss function using physical regularization constraints includes: Construct a weighted sum of the total loss function as the data-driven loss term and the physical regularization constraint term; The data-driven loss term is calculated by the cross-entropy between the transient stability prediction result output by the neural network model for the training samples and the actual transient stability result corresponding to the training samples; the physical regularization constraint term includes a motion trajectory regularization term and an energy consistency regularization term. The motion trajectory regularization term is constructed by the mean square error between the theoretical power angle and theoretical speed calculated from the generator rotor motion equation and the predicted power angle and speed output by the neural network model during training; the energy consistency regularization term is constructed by the mean square error between the theoretical transient energy margin calculated based on the transient energy function and the predicted stability margin output by the neural network model. The theoretical power angle, theoretical rotational speed, and theoretical transient energy margin are used as the physical benchmarks for training the neural network model. During the training process, the total loss function is optimized through the backpropagation algorithm, and the parameters of the neural network model are iteratively updated until the preset number of iterations is reached. The trained neural network model is then output as a transiently stable model. 3.The power system transient stability intelligent evaluation and early warning method based on multi-source data fusion of claim 1, wherein, The process of feature fusion and inference using a preset transient stability model combined with multimodal data includes: The multimodal data includes: phasor measurement data, SCADA operation data, and fault recording data; The preset transient stability model includes three parallel feature extraction branches: a graph neural network branch extracts the power grid topology spatial features from phasor measurement data; a temporal neural network branch extracts the temporal evolution features from SCADA operation data; and a convolutional neural network branch extracts the transient waveform features from fault recording data. The power grid topology features, temporal evolution features, and transient waveform features are weighted and fused using a trainable attention mechanism to output a fused feature vector. The fused feature vector is processed by a normalized exponential function to output stable class probabilities and unstable class probabilities. The stable class probabilities are compared with the unstable class probabilities, and the class with the larger probability value is determined as the transient stability discrimination result. The probability value corresponding to the class is used as the confidence score. At the same time, the stability margin index is obtained by using the fused feature vector as input through a linear regression algorithm.

4. The power system transient stability intelligent evaluation and early warning method based on multi-source data fusion according to claim 3, characterized in that, The process of weighted fusion of power grid topology features, temporal evolution features, and transient waveform features using a trainable attention mechanism includes: Configure a set of learnable query parameters and key parameters for each of the power grid topology spatial features, temporal evolution features, and transient waveform features; calculate the inner product of each set of query parameters and key parameters to obtain the original attention score corresponding to each feature; Each original attention score is input into a single-layer nonlinear mapping layer for transformation, and the transformed original attention scores are converted into three attention weights with a sum of 1 through a normalized exponential function. The power grid topology features, time-series evolution features, and transient waveform features are multiplied by their respective attention weights to obtain a three-way weighted feature vector; the three-way weighted feature vectors are then summed element-wise to output a unified fusion feature vector.

5. The intelligent assessment and early warning method for power system transient stability based on multi-source data fusion according to claim 1, characterized in that, The preset transient stability model generates a fault propagation causal evidence chain starting from the fault point through feature attribution analysis, including: The phasor measurement data, SCADA operation data, and fault recording data in the multimodal data are used as the input feature set; the contribution of each feature in the input feature set to the transient stability discrimination result is calculated by the integral gradient method. Starting from the power equipment node corresponding to the fault point, the contribution is sorted from high to low. Combining the electrical connection relationship in the power system topology, the adjacent equipment nodes that have a direct electrical connection with the current power equipment node and whose contribution exceeds the preset contribution threshold are searched level by level, and the adjacent equipment nodes are included in the search path; an ordered sequence is formed that expands outward from the fault point, and the ordered sequence is used as the causal evidence chain of fault transmission. The ordered sequence is arranged in a hierarchical order that expands outward from the fault point. The device nodes within the same level are sorted from high to low contribution, including the device node, the name of the input feature with the highest contribution under the device node, and the contribution value corresponding to the input feature.

6. The intelligent assessment and early warning method for power system transient stability based on multi-source data fusion according to claim 1, characterized in that, The process of generating graded early warning levels based on transient stability assessment results and using preset adaptive grading rules includes: The preset adaptive grading rule is based on the statistical distribution of the stability margin index and confidence score output by the transient stability model under historical operating conditions. The confidence threshold, the first margin threshold and the second margin threshold are dynamically determined by the quantile method, and the first margin threshold is greater than the second margin threshold. When the transient stability determination result is in the stable category and the confidence score is greater than or equal to the confidence threshold, there is no warning level. When the transient stability determination result is in the stable category and the confidence score is less than the confidence threshold, it corresponds to a level four warning level; When the transient stability determination result is instability category and the stability margin index is greater than or equal to the first margin threshold, it corresponds to a level three warning level; When the transient stability determination result is instability category, and the stability margin index is greater than or equal to the second margin threshold and less than the first margin threshold, it corresponds to a level two warning level; When the transient stability determination result is instability category and the stability margin index is less than the second margin threshold, it corresponds to a first-level warning level; The priority of the graded early warning levels, from highest to lowest, is as follows: Level 1, Level 2, Level 3, and Level 4.

7. The intelligent assessment and early warning method for power system transient stability based on multi-source data fusion according to claim 1, characterized in that, The sensitivity analysis for calculating the impact sensitivity of each power equipment node in the causal evidence chain on the stability margin index includes: Each power equipment node in the causal evidence chain is taken as the analysis object, and the input features corresponding to each power equipment node are extracted. The input features include the voltage amplitude, voltage phase angle, active power, reactive power and power angle value of the power equipment node. The input characteristics corresponding to each power equipment node are used as independent variables, and the stability margin index is used as the dependent variable. The first-order partial derivative of each independent variable with respect to the dependent variable is calculated by automatic differentiation. The absolute values ​​of all first-order partial derivatives corresponding to each power equipment node are summed, and the sensitivity of each power equipment node to the stability margin index is output. All power equipment nodes are sorted in descending order of their impact sensitivity to generate a descending sequence of equipment nodes. The equipment node sequence is then traversed, and power equipment nodes with impact sensitivity greater than a preset sensitivity threshold are extracted as core instability influencing factors.

8. The intelligent assessment and early warning method for transient stability of power systems based on multi-source data fusion according to claim 1, characterized in that, The process combines core instability influencing factors with graded early warning levels, and generates transient stability control strategies for the current power system operating conditions through a pre-set control strategy mapping table. These strategies include: The preset prevention and control strategy mapping table is pre-constructed based on transient stability control data under historical operating conditions. It includes multiple mapping entries, each of which records a graded early warning level, a set of core instability influencing factors, and a set of corresponding transient stability control measures. The graded early warning level and the core instability influencing factors are used as a joint matching index to find the mapping entry corresponding to the joint matching index in the prevention and control strategy mapping table. When the mapping entry is found, the corresponding transient stability control measures in the mapping entry are extracted, and the transient stability control measures are used as the transient stability prevention and control strategy; When the mapping entry is not found, the default control measures are matched based on the graded early warning level. The default control measures are: the first-level early warning level corresponds to the rapid shedding of load through the preset load reduction priority; the second-level early warning level corresponds to the redistribution of generator power; the third-level early warning level corresponds to the switching on and off of the dynamic reactive power compensation device; and the fourth-level early warning level corresponds to only issuing an alarm. The preset load reduction priority is pre-configured based on the importance level of the load and the contribution of each load node to transient instability.

Images