Multi-modal data fusion driven protection setting value dynamic verification system, method and device and storage medium

Abstract

The invention discloses a multi-modal data fusion-driven protection setting value dynamic verification system, method and device and a storage medium, and the method comprises the steps: integrating static structured data, dynamic time sequence data, image data and text data in a power system, combining a large model and a deep learning technology, extracting the data features of each modal, and carrying out the dynamic verification of a protection setting value; a mixed fusion strategy is adopted, calculation is carried out through a Bayesian rule combination method, a final decision result is output, and real-time monitoring, accurate positioning and root cause analysis of constant value abnormity are achieved; the system is divided into four core hierarchies including a data layer, a fusion layer, an application layer and a presentation layer, and the hierarchies work cooperatively to improve the accuracy of constant value verification and the interpretability of anomaly diagnosis, and improve the real-time performance and the intelligent level of constant value operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and in particular to a multimodal data fusion-driven dynamic verification system, method, device and storage medium for protection settings. Background Technology

[0002] Currently, monitoring of abnormal settings in power system equipment relies on single data such as historical settings and real-time power flow, which suffers from limitations in data dimensions and a lack of spatiotemporal correlation. Monitoring is based solely on electrical quantities such as current and voltage, ignoring the influence of non-electrical quantities such as environment and equipment status. Manual analysis relies on experience, resulting in a lag in response to complex multi-factor coupled anomalies. Data sources from protection devices, SCADA systems, and meteorological monitoring are scattered and lack a unified integration framework. Traditional rule bases are difficult to cover dynamic scenarios such as changes in power grid topology and the integration of new energy sources. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a dynamic verification system for protection settings driven by multimodal data fusion, which solves the problems of existing power system protection setting verification relying on a single data mode and historical experience, low verification accuracy, and lack of interpretability in the diagnostic process; Another purpose of this invention is to provide a method, device, and storage medium for dynamic verification of protection settings driven by multimodal data fusion.

[0004] Technical Solution: The multimodal data fusion-driven dynamic verification system for protection settings of this invention includes a data layer, a fusion layer, an application layer, and a representation layer. The data layer is used to collect and preprocess multimodal raw data, including static modality, dynamic modality, auxiliary modality, and text modality. The fusion layer adopts a hybrid fusion strategy combining feature fusion and decision fusion to extract high-dimensional features of each modality data and outputs the final decision result through a Bayesian rule combination method. The application layer performs dynamic verification of settings, anomaly diagnosis, and adjustment scheme generation based on the final decision result output by the fusion layer. The representation layer provides a system visualization interface based on alarm and visualization services to support data visualization display, alarm push, and operation and maintenance.

[0005] Furthermore, in the data layer, four types of modal data are collected and standardized preprocessed: the static modal data includes protection device parameters, equipment ledgers, setting sheets, and network topology information; the dynamic modal data includes three-phase current / voltage waveforms, active / reactive power, switch status timing, and fault recordings; the auxiliary modal data includes ambient temperature and humidity and the operating conditions of primary and secondary equipment; and the text modal data comes from equipment logs, maintenance records, and fault reports.

[0006] Furthermore, in the fusion layer, a multi-level feature extraction and hybrid fusion strategy is implemented: a convolutional neural network (CNN) is used to model the static modality data, spatial topological features are extracted through convolution operations, and a fixed-length static feature vector F is output after passing through the ReLU activation function and global average pooling. s The Transformer model is used to process dynamic modal data, and the temporal features F of the dynamic modality are extracted by concatenating the outputs of each attention head through a multi-head self-attention mechanism. t The static topological features and dynamic temporal features are concatenated and input into a two-layer neural network, outputting a fusion vector V1. The auxiliary modal data uses an improved feature fusion algorithm to extract high-dimensional features of the environment and working conditions, and the multimodal features are unified in dimensionality through a randomized fusion function, outputting a fusion vector V2. The text modal data uses the BERT model to convert the text into a low-dimensional vector, outputting a fusion vector V3.

[0007] Furthermore, using the Bayesian rule combination method, the calculation formula is as follows: ; Where Y represents the decision result, and V1, V2, and V3 represent the fusion vectors of static-dynamic, auxiliary, and textual modalities, respectively. For the final decision result, For the joint likelihood of features, For class prior probabilities, This is an evidence factor.

[0008] Furthermore, the application layer performs dynamic verification, anomaly diagnosis, and decision output dynamic verification based on the fusion features. This is achieved by calculating the Euclidean distance between the fusion feature vector and the benchmark value. When the distance exceeds the threshold, it is marked as a potential anomaly. Based on the fusion decision result Y, the dynamic threshold range of the protection setting under the current power grid operating conditions is predicted.

[0009] The multimodal data fusion-driven dynamic verification method for protection settings of the present invention is implemented through the aforementioned multimodal data fusion-driven dynamic verification system for protection settings, and the method includes: (1) Preprocess the input multimodal data features and perform feature data verification; (2) Perform multimodal fusion decision analysis on the verified multimodal data features, and calculate the probability of anomaly occurrence based on Bayesian probability theory and the multimodal fusion results; (3) Decision threshold judgment: The power grid state feature vector, environmental impact feature vector and fault hazard feature vector extracted from the text modality after decision fusion are calculated. The Euclidean distance between each feature vector and the benchmark value is calculated. The calculated probability is compared with the preset threshold. If the result is judged that the probability is not higher than the threshold, it is judged that the set value is normal and the process ends; if the probability is higher than the threshold, it is judged that the set value is abnormal and the process proceeds to step (4). (4) Locate abnormal set values, perform root cause analysis, trigger alarm mechanisms to provide abnormal warnings, generate set value abnormality reports, and determine set value adjustment schemes.

[0010] Furthermore, the multimodal data fusion-driven dynamic verification method for protection settings also includes: (5) Verify whether the setting scheme effectively resolves the anomaly. If the setting is effective, the scheme successfully repairs the anomaly and the process ends. If the setting is ineffective, the scheme fails to resolve the anomaly and returns to step (1).

[0011] Furthermore, in the preprocessing of the input multimodal data features, static modal data is filled with mean based on equipment type to complete missing values, and outliers are identified and removed using the Z-score method; dynamic modal data is filled with linear interpolation to complete discontinuities and unify data from different sampling rates to the same time interval; analog quantities such as current and voltage are standardized from 0 to 1 to eliminate the influence of dimensional differences on model input; auxiliary modal data is based on timestamps and grid topology coordinates, aligning multi-source data in time and space; and equipment names and fault types are pre-extracted from text data to construct a knowledge dictionary for the power field.

[0012] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0013] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0014] Beneficial effects: Compared with the prior art, the significant advantages of this invention are: by deeply integrating multimodal data and dynamically analyzing large models, this invention breaks through the bottleneck of data simplification and experience-based decision-making in traditional protection setting operation and maintenance, and can be directly applied to scenarios such as smart substations and new energy grid-connected systems, providing core technical support for the digital transformation of the power grid. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The multimodal data fusion-driven dynamic verification system for protection settings described in this invention comprises four core layers: a data layer, a fusion layer, an application layer, and a representation layer. By integrating static structured data, dynamic time-series data, image data, and text data from the power system, and combining large-scale models and deep learning techniques, it extracts features from each modality of data, adopts a hybrid fusion strategy, and outputs the final decision result through a Bayesian rule combination method to achieve real-time monitoring, precise location, and root cause analysis of setting anomalies.

[0017] The data layer is responsible for collecting and preprocessing multimodal data to ensure data consistency and availability. Multimodal data includes static modalities, dynamic modalities, auxiliary modalities, and text modalities.

[0018] The static modality is obtained by collecting static structured data such as protection device parameters, equipment ledger information, setting sheets and network topology from substation monitoring systems and power dispatching platforms. Missing values ​​are filled by using the mean based on equipment type, and outliers are identified and removed by Z-score method.

[0019] The dynamic mode refers to real-time high-sampling-rate data such as three-phase current / voltage waveforms, active / reactive power, switch state timing, and fault recordings collected by synchronous phasor measurement devices and protection devices. Linear interpolation is used to complete discontinuities, unifying data from different sampling rates to the same time interval. Analog quantities such as current and voltage are standardized from 0 to 1 to eliminate the influence of dimensional differences on the model input.

[0020] The auxiliary mode refers to the temperature and humidity of the equipment operating environment and the operating conditions of primary and secondary equipment. Based on timestamps and power grid topology coordinates, it realizes the spatiotemporal alignment of multi-source data.

[0021] The text modality refers to unstructured text data derived from equipment operation logs, maintenance records, fault reports, etc. Equipment names and fault types are extracted from the text to construct a power industry knowledge dictionary to assist in diagnosis.

[0022] The fusion layer refers to the fusion strategy that combines feature fusion and decision fusion to fuse the input multimodal features. Feature fusion involves extracting the topological features of static modal data through a CNN model and extracting the temporal features of dynamic modal data through a Transformer model. The two types of features are concatenated and input into a two-layer neural network. Feature association is achieved through hidden layer weight learning, and the fusion vector V1 is output.

[0023] The static modal feature extraction is based on input static modal data. Spatial features are extracted using convolutional layers, with height H, width W, and number of channels C. ,in For convolution kernel weights, For bias, The ReLU activation function is used. Global average pooling is applied to the convolutional features to obtain a fixed-length topological feature vector. , For static feature dimensions, Reduce spatial dimensions Flattening characteristics.

[0024] The dynamic modal feature extraction is based on input dynamic modal data. Time step T Feature Dimension E The Transformer feature extraction process is as follows: Add position encoding To preserve timing information, , pass h The attention head learning time-series dependency relationship, the first i The attention output of the size is: ; in, , , , Let be the projection matrix. The multi-head attention results are concatenated and passed through a linear layer: ; The time-series feature vector is output through two layers of linear transformation and residual connection. Then, take the feature from the last time step as the global time series feature: .

[0025] Static topological features Fs and dynamic temporal features Ft are concatenated, and feature associations are learned through a two-layer neural network to output a fusion vector V1.

[0026] ; in, The first hidden layer weight matrix has a bias. b1 , The weight matrix for the second output layer, biased. b2 .

[0027] An improved feature fusion algorithm is used to extract features from auxiliary modal data. and working conditions High-dimensional features are unified in dimensionality through a randomized fusion function, outputting a fusion vector V2. The weight vector is generated by uniformly distributed sampling. , .

[0028] ; The text modal data is converted into a low-dimensional vector using the BERT model, and the output is a fused vector V3.

[0029] ; in, , , These are word vectors, text vectors, and position vectors, respectively, with T representing the number of tokens.

[0030] The decision fusion refers to taking the output power grid state feature vector V1, environmental impact feature vector V2, and fault hazard feature vector V3 extracted from the text modality as inputs, and using the Bayesian rule combination method to calculate the final decision result through the following formula.

[0031] ; Where Y represents the decision result, and V1, V2, and V3 represent the fusion vectors of static-dynamic, auxiliary, and textual modalities, respectively. For the final decision result, For the joint likelihood of features under a given category, The prior probability of a category can be set based on historical data statistics. This is an evidence factor.

[0032] Finally, by comparing the posterior probabilities of the three candidate categories, the category with the highest probability is selected as the decision result Y.

[0033] The application layer refers to the dynamic verification, anomaly diagnosis, and decision output based on the fused features. The Euclidean distance between each feature vector and the baseline value is calculated from the fused power grid state feature vector V1, environmental impact feature vector V2, and fault hazard feature vector V3 extracted from the text modality. D Determine whether the abnormal threshold is exceeded. .

[0034] ; ; in For the first k Historical mean of the dimensional feature Marked as a potential anomaly.

[0035] Based on the fusion decision result Y, predict the dynamic threshold range of protection settings under the current power grid operating conditions. .

[0036] Determine if the current protection setting S is satisfied. If the conditions for triggering an anomaly alarm are not met, the cause of the anomaly is identified based on potential anomalies. Natural language generation technology is used to convert the diagnostic conclusions into a structured report. Based on the anomaly diagnosis results, a setpoint adjustment plan is automatically generated. The adjusted setpoints are re-entered, and their effectiveness is verified. If the requirements are not met, the fusion and verification process is re-triggered.

[0037] The multimodal data fusion-driven dynamic verification method for protection settings of the present invention is implemented through the aforementioned multimodal data fusion-driven dynamic verification system for protection settings, and the method includes: (1) Preprocess the input multimodal data features and perform feature data verification; (2) Perform multimodal fusion decision analysis on the verified multimodal data features, and calculate the probability of anomaly occurrence based on Bayesian probability theory and the multimodal fusion results; (3) Decision threshold judgment: The power grid state feature vector, environmental impact feature vector and fault hazard feature vector extracted from the text modality after decision fusion are calculated. The Euclidean distance between each feature vector and the benchmark value is calculated. The calculated probability is compared with the preset threshold. If the result is judged that the probability is not higher than the threshold, it is judged that the set value is normal and the process ends; if the probability is higher than the threshold, it is judged that the set value is abnormal and the process proceeds to step (4). (4) Locate abnormal set values, perform root cause analysis, trigger the alarm mechanism to issue an abnormal warning, generate a set value abnormality report, and determine the set value adjustment scheme; (5) Verify whether the setting scheme effectively resolves the anomaly. If the setting is effective, the scheme successfully repairs the anomaly and the process ends. If the setting is ineffective, the scheme fails to resolve the anomaly and returns to step (1).

[0038] Furthermore, in the preprocessing of the input multimodal data features, static modal data is filled with mean based on equipment type to complete missing values, and outliers are identified and removed using the Z-score method; dynamic modal data is filled with linear interpolation to complete discontinuities and unify data from different sampling rates to the same time interval; analog quantities such as current and voltage are standardized from 0 to 1 to eliminate the influence of dimensional differences on model input; auxiliary modal data is based on timestamps and grid topology coordinates, aligning multi-source data in time and space; and equipment names and fault types are pre-extracted from text data to construct a knowledge dictionary for the power field.

[0039] The computer device of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0040] The computer-readable storage medium of the present invention stores a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

Claims

1. A multimodal data fusion-driven dynamic verification system for protection settings, characterized in that, It includes a data layer, a fusion layer, an application layer, and a representation layer. The data layer is used to collect and preprocess multimodal raw data, including static modality, dynamic modality, auxiliary modality, and text modality. The fusion layer adopts a hybrid fusion strategy that combines feature fusion and decision fusion to extract high-dimensional features of each modality data and outputs the final decision result through a Bayesian rule combination method. The application layer performs dynamic value verification, anomaly diagnosis, and adjustment scheme generation based on the final decision results output by the fusion layer; the presentation layer provides a system visualization interface based on alarm and visualization services to support data visualization display, alarm push and operation and maintenance.

2. The protection setting dynamic verification system driven by multimodal data fusion according to claim 1, characterized in that, In the data layer, four types of modal data are collected and standardized preprocessed: the static modal data includes protection device parameters, equipment ledgers, setting sheets, and network topology information; the dynamic modal data includes three-phase current / voltage waveforms, active / reactive power, switch status timing, and fault recordings; the auxiliary modal data includes ambient temperature and humidity and the operating conditions of primary and secondary equipment; and the text modal data comes from equipment logs, maintenance records, and fault reports.

3. The protection setting dynamic verification system driven by multimodal data fusion according to claim 1, characterized in that, In the fusion layer, a multi-level feature extraction and hybrid fusion strategy is implemented: a convolutional neural network (CNN) is used to model the static modality data, spatial topological features are extracted through convolution operations, and a fixed-length static feature vector F is output after ReLU activation and global average pooling. s The Transformer model is used to process dynamic modal data, and the temporal features F of the dynamic modality are extracted by concatenating the outputs of each attention head through a multi-head self-attention mechanism. t The static topological features and dynamic temporal features are concatenated and input into a two-layer neural network, outputting a fusion vector V1. The auxiliary modal data uses an improved feature fusion algorithm to extract high-dimensional features of the environment and working conditions, and the multimodal features are unified in dimensionality through a randomized fusion function, outputting a fusion vector V2. The text modal data uses the BERT model to convert the text into a low-dimensional vector, outputting a fusion vector V3.

4. The protection setting dynamic verification system driven by multimodal data fusion according to claim 3, characterized in that, Using the Bayesian rule combination method, the calculation formula is as follows: ； Where Y represents the decision result, and V1, V2, and V3 represent the fusion vectors of static-dynamic, auxiliary, and textual modalities, respectively. For the final decision result, For the joint likelihood of features, For class prior probabilities, This is an evidence factor.

5. The protection setting dynamic verification system driven by multimodal data fusion according to claim 1, characterized in that, The application layer performs dynamic verification, anomaly diagnosis, and decision output dynamic verification based on fusion features. This is achieved by calculating the Euclidean distance between the fusion feature vector and the benchmark value. When the distance exceeds a threshold, it is marked as a potential anomaly. Based on the fusion decision result Y, the dynamic threshold range of the protection setting under the current power grid operating conditions is predicted.

6. A method for dynamic verification of protection settings driven by multimodal data fusion, characterized in that, This is achieved through the multimodal data fusion-driven dynamic verification system for protection settings as described in any one of claims 1-5, the method comprising: (1) Preprocess the input multimodal data features and perform feature data verification; (2) Perform multimodal fusion decision analysis on the verified multimodal data features, and calculate the probability of anomaly occurrence based on Bayesian probability theory and the multimodal fusion results; (3) Decision threshold judgment: The power grid state feature vector, environmental impact feature vector and fault hazard feature vector extracted from the text modality after decision fusion are calculated. The Euclidean distance between each feature vector and the benchmark value is calculated. The calculated probability is compared with the preset threshold. If the result is judged that the probability is not higher than the threshold, it is judged that the set value is normal and the process ends; if the probability is higher than the threshold, it is judged that the set value is abnormal and the process proceeds to step (4). (4) Locate abnormal set values, perform root cause analysis, trigger alarm mechanisms to provide abnormal warnings, generate set value abnormality reports, and determine set value adjustment schemes.

7. The method for dynamic verification of protection settings driven by multimodal data fusion according to claim 6, characterized in that, Also includes: (5) Verify whether the setting scheme effectively resolves the anomaly. If the setting is effective, the scheme successfully repairs the anomaly, and the process ends. If the set value adjustment is invalid: the solution fails to resolve the exception, return to step (1).

8. The method for dynamic verification of protection settings driven by multimodal data fusion according to claim 6, characterized in that, In the preprocessing of input multimodal data features, static modal data is filled with mean based on equipment type to complete missing values, and outliers are identified and removed using the Z-score method; dynamic modal data is filled with discontinuities using linear interpolation to unify data from different sampling rates to the same time interval; analog quantities such as current and voltage are standardized from 0 to 1 to eliminate the influence of dimensional differences on model input; auxiliary modal data is based on timestamps and grid topology coordinates, aligning multi-source data in time and space; equipment names and fault types are pre-extracted from text data to construct a power field knowledge dictionary.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any of the methods described in claims 6-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 6-8.

Images