Fault scene simulation method based on deep learning fusion model and terminal

Abstract

The invention discloses a fault scene simulation method based on a deep learning fusion model and a terminal. According to the method, an encoder is introduced into a generative adversarial network comprising a generator and a discriminator, so that the generative adversarial network which is difficult to capture time sequence data features originally can effectively extract important features of the time sequence data. Conditional variables of data are introduced into a generative adversarial network comprising a generator and a discriminator, so that the generator can control characteristics of simulated fault data based on the conditional variables. Meanwhile, mutual confrontation training is carried out through the generator and the discriminator, and the generator generates simulation fault data as vivid as possible, so that the discriminator is difficult to distinguish the generated data and the real data, and the discriminator accurately distinguishes the generated data and the real data. The generator continuously learns how to generate simulated fault data closer to real data in the iterative training process, so that the accuracy of the generated data is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and particularly to a fault scenario simulation method and a terminal based on a deep learning fusion model. Background Art

[0002] With the continuous expansion of the scale and the increase in complexity of power systems, fault diagnosis and prediction play a crucial role in ensuring the stable operation of power systems. Traditional fault diagnosis methods mainly rely on rule-based expert systems and statistical analysis methods, which often show limitations when dealing with large-scale, multi-variable, and non-linear data.

[0003] Traditional rule bases or statistical models are usually based on the abstraction and simplification of partial fault mechanisms, and it is difficult to comprehensively and accurately depict the dynamic evolution process of complex systems under different fault scenarios. Once the fault scenario exceeds the applicable range of the original modeling assumptions or rule bases, the obtained fault data is not real enough or cannot be generated. Moreover, traditional expert systems or statistical models often use linear approximations, fixed rules, or a small number of random variables to describe complex systems, and cannot fully depict the non-linear coupling, non-equilibrium states, and random factors (such as load fluctuations, uncertainty of new energy output, etc.) in power systems. When there are significant deviations between the actual fault conditions and the model assumptions, the fault data that these methods can generate or infer is far from the real situation, and it is difficult to meet the requirements of large-scale and diverse fault scenarios.

[0004] In addition, with the transformation of the power grid to a smart grid, a large number of new types of equipment and technologies (such as distributed power sources, energy storage systems, electric vehicle charging and discharging facilities, flexible DC transmission equipment, etc.) are continuously connected. The dynamic response process of these equipment under fault conditions and the interaction mechanism with the traditional power grid often do not have mature models or rule bases. The traditional methods are lagging in updating, resulting in the generated data being unable to reflect the diverse fault characteristics under new technologies, new equipment, and new operation modes. Moreover, traditional methods usually conduct fault research on power systems with relatively small scales or relatively fixed structures, and have limited data processing capabilities and algorithm iteration speeds. Facing the huge data volume of modern power systems and the requirements for real-time online diagnosis or prediction, traditional methods are difficult to iteratively update the fault data generation model flexibly and efficiently, thus it is difficult to reflect the real-time operation characteristics of the power grid and unable to generate fast and diverse fault data.

[0005] In summary, traditional methods have deficiencies in generating diverse and real fault data and are difficult to comprehensively simulate complex fault scenarios. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a fault scenario simulation method and terminal based on a deep learning fusion model, which can improve the authenticity and diversity of generated data, enhance the ability to capture complex fault patterns, and thus optimize the power system to achieve the effect of fault simulation.

[0007] To solve the above technical problem, a technical solution adopted by the present invention is: A fault scenario simulation method based on a deep learning fusion model, including: Inputting the fault training data of the power system under different conditional variables into a preset fusion generative adversarial network model, where the fusion generative adversarial network model includes an encoder, a generator, and a discriminator; Compressing the fault training data through the encoder to obtain a hidden state vector; Generating data through the generator using the hidden state vector and the conditional variable corresponding to the hidden state vector to obtain simulated fault data under the conditional variable; Distinguishing the simulated fault data under the conditional variable and the historical fault data through the discriminator to obtain a first loss function of the generator and a second loss function of the discriminator; Iteratively training the fusion generative adversarial network model according to the first loss function and the second loss function to obtain a fault scenario simulation model; Generating target fault data of the power system under target conditional variables through the fault scenario simulation model.

[0008] To solve the above technical problem, another technical solution adopted by the present invention is: A fault scenario simulation terminal based on a deep learning fusion model, including a memory, a processor, and a computer program stored on the memory and running on the processor, and when the processor executes the computer program, each step in the above-mentioned fault scenario simulation method based on a deep learning fusion model is implemented.

[0009] The beneficial effects of the present invention are as follows: The fault training data of the power system belongs to time-series data. An encoder is introduced into the generative adversarial network including a generator and a discriminator. The time-series data is compressed into a low-dimensional hidden state vector through the encoder, enabling the generative adversarial network that originally had difficulty capturing the characteristics of time-series data to effectively extract important features such as the time dependence, period, and trend of the time-series data. Moreover, in the generative adversarial network including a generator and a discriminator, a conditional variable of the data is introduced, enabling the generator to control the characteristics of the simulated fault data based on the conditional variable, effectively enhancing the scenario pertinence and type diversity of the fault data. At the same time, the generative adversarial network conducts mutual adversarial training through the generator and the discriminator. The generator generates simulated fault data as realistic as possible, making it difficult for the discriminator to distinguish the generated data from the real data. The discriminator accurately distinguishes the generated data from the real data, enabling the generator to continuously learn how to generate simulated fault data closer to the real data during the iterative training process, thereby improving the accuracy of the generated data. By integrating the encoder and the conditional variable in the generative adversarial network, the generative adversarial network can accurately generate time-series data under multiple fault scenarios based on the conditional variable, thereby optimizing the effect of fault simulation in the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flowchart of the steps of a fault scenario simulation method based on a deep learning fusion model provided by an embodiment of the present invention; Figure 2 It is a training flowchart of a fusion generative adversarial network model provided by an embodiment of the present invention; Figure 3 It is a schematic structural diagram of a fault scenario simulation terminal based on a deep learning fusion model provided by an embodiment of the present invention; Reference Signs Explanation: 100. A fault scenario simulation terminal based on a deep learning fusion model; 101. A memory; 102. A processor. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] To describe in detail the technical content, achieved objectives, and effects of the present invention, the following is described in conjunction with the embodiments and with reference to the accompanying drawings.

[0012] An embodiment of the present invention provides a fault scenario simulation method based on a deep learning fusion model, including: Inputting the fault training data of the power system under different conditional variables into a preset fusion generative adversarial network model, where the fusion generative adversarial network model includes an encoder, a generator, and a discriminator; Compressing the fault training data through the encoder to obtain a hidden state vector; Generate simulated fault data under the conditional variable by inputting the hidden state vector and the conditional variable corresponding to the hidden state vector into the generator. Distinguish the simulated fault data under the conditional variable and the historical fault data through the discriminator to obtain the first loss function of the generator and the second loss function of the discriminator. Iteratively train the fusion generative adversarial network model according to the first loss function and the second loss function to obtain a fault scenario simulation model. Generate target fault data of the power system under the target conditional variable through the fault scenario simulation model.

[0013] As can be seen from the above description, the beneficial effects of the present invention are as follows: The fault training data of the power system belongs to time series data. An encoder is introduced into the generative adversarial network including a generator and a discriminator. The time series data is compressed into a low-dimensional hidden state vector through the encoder, enabling the generative adversarial network that originally had difficulty capturing the characteristics of time series data to effectively extract important features such as the time dependence, period, and trend of time series data. Moreover, in the generative adversarial network including a generator and a discriminator, the conditional variable of the data is introduced, enabling the generator to control the characteristics of the simulated fault data based on the conditional variable, effectively enhancing the scenario pertinence and type diversity of the fault data. At the same time, the generative adversarial network conducts mutual adversarial training through the generator and the discriminator. The generator generates simulated fault data as realistic as possible, making it difficult for the discriminator to distinguish between the generated data and the real data. The discriminator accurately distinguishes between the generated data and the real data, enabling the generator to continuously learn how to generate simulated fault data closer to the real data during the iterative training process, thereby improving the accuracy of the generated data. By integrating the encoder and the conditional variable in the generative adversarial network, the generative adversarial network can accurately generate time series data under multiple fault scenarios based on the conditional variable, thereby optimizing the effect of the power system in realizing fault simulation.

[0014] Further, before inputting the fault training data of the power system under different conditional variables into the preset fusion generative adversarial network model, it further includes: Obtain the historical operation data of the power system under different conditional variables, where the historical operation data includes historical non-fault data and historical fault data. Decompose and combine the historical operation data through a variational mode model to obtain fault training data under different conditional variables.

[0015] As can be seen from the above description, before inputting the historical operation data into the fusion generative adversarial network model for training, the variational mode model is used to decompose and combine the historical operation data to obtain the fault training data under different conditional variables, which can more accurately extract the features related to the fault and provide high-quality input data for the subsequent training of the generative adversarial network model, so as to improve the model's ability to capture complex fault patterns and enhance the authenticity and accuracy of the generated data.

[0016] Further, decomposing and combining the historical operation data through the variational mode model to obtain the fault training data under different conditional variables includes: Decomposing the historical operation data into multiple modal components and the corresponding central frequencies of each modal component through the variational mode model; Selecting the key modal components related to the conditional variable from the multiple modal components according to the central frequency; Setting the combination weights according to the contribution degree of the key modal components to the conditional variable; Recombining the key modal components according to the combination weights to obtain the fault training data under different conditional variables.

[0017] As can be seen from the above description, the variational mode model generates the fault training data through the decomposition of modal components, the selection of key modal components, and the recombination of modal components. This process can effectively extract the key features related to the conditional variables, perform weighted processing according to their contribution degrees, further optimize the quality of the fault training data, and also increase the diversity of the training data through the recombination method. The fault training data generated in this way is more targeted and representative, which helps to improve the training effect of the fusion generative adversarial network model, enables it to more accurately generate the fault data under different conditional variables, and improves the accuracy and reliability of the power system fault simulation.

[0018] Further, the encoder includes a first long short-term memory network layer; Compressing the fault training data through the encoder to obtain the hidden state vector includes: Extracting the key fault features from the fault training data through the first long short-term memory network layer and compressing the key fault features into the hidden state vector.

[0019] As described above, the encoder uses the first long short-term memory network layer to extract key fault features and compress them into the hidden state vector. The long short-term memory network layer (LSTM) has powerful time series data processing capabilities and can effectively capture long-term dependencies in the data. This encoder structure design not only improves the efficiency and quality of data compression but also provides the generator with richer time series feature information, helping the generator generate relevant time series data more accurately and further enhancing the performance and effectiveness of the entire fault data generation method.

[0020] Further, the generator includes a second long short-term memory network layer and a first fully connected layer using the hyperbolic tangent activation function; Generating data from the hidden state vector and the conditional variable corresponding to the hidden state vector through the generator to obtain the simulated fault data under the conditional variable includes: Converting the conditional variable corresponding to the hidden state vector into a conditional vector through one-hot encoding; Combining the conditional vector with the obtained random noise vector to obtain a target vector; Combining the hidden state vector and the target vector to obtain an input vector; Fusing the fault features of the input vector through the second long short-term memory network layer to obtain generated data; Mapping the generated data to the target data dimension through the first fully connected layer to obtain the simulated fault data under the conditional variable.

[0021] As described above, the conditional variable is converted into a conditional vector through one-hot encoding, which facilitates combining the conditional vector with the random noise vector and then inputting it into the generator. The random noise vector containing the conditional vector and the hidden state vector are used for fault feature fusion and data mapping through the LSTM and the fully connected layer. This generator structure design makes full use of the long-term dependence capture ability of the LSTM and the non-linear mapping characteristics of the hyperbolic tangent activation function, and can generate simulated fault data that more accurately matches the conditional variable, providing more reliable simulated data support for fault analysis and prediction in the power system and helping to improve the accuracy and effectiveness of fault handling.

[0022] Further, the discriminator includes a third long short-term memory network layer and a second fully connected layer using the sigmoid activation function; Differentiating the simulated fault data under the conditional variable and the historical fault data through the discriminator to obtain the first loss function of the generator and the second loss function of the discriminator includes: Extract the joint features between the conditional variables and the time series data from the simulated fault data and the historical fault data respectively through the third long short-term memory network layer; Map the joint features to scalar data through the second fully connected layer to obtain the probability values that the simulated fault data and the historical fault data are judged to be real data; Calculate the first loss function of the generator and the second loss function of the discriminator according to the probability values.

[0023] As can be seen from the above description, the discriminator uses the third long short-term memory network layer to extract the joint features of the conditional variables and the data, and maps out the probability value of judging real data through the fully connected layer with the sigmoid activation function. This discriminator structure design can effectively distinguish the simulated fault data and the historical fault data, providing accurate feedback information for the training of the generator and the discriminator. The discriminator can more accurately evaluate the authenticity of the generated data, thereby guiding the generator to continuously optimize the generation process and improve the quality and diversity of the generated data.

[0024] Further, iteratively training the fusion generative adversarial network model according to the first loss function and the second loss function to obtain a fault scenario simulation model includes: Alternately update the training parameters of the generator and the discriminator according to the first loss function and the second loss function until the first loss function reaches the maximum and the second loss function reaches the minimum to obtain the target parameters; Configure the generator and the discriminator according to the target parameters to obtain a fault scenario simulation model.

[0025] As can be seen from the above description, alternately updating the training parameters of the generator and the discriminator according to the first loss function and the second loss function until the optimal target parameters are reached can ensure the balance between the generator and the discriminator during the adversarial training process, thereby improving the stability and convergence speed of the model. The fault scenario simulation model trained in this way can more accurately generate the target fault data under the target conditional variables, providing a more reliable tool for the fault simulation and analysis of the power system, and helping to improve the accuracy and efficiency of fault handling.

[0026] Further, after alternately updating the training parameters of the generator and the discriminator according to the first loss function and the second loss function, it further includes: If it is detected that the update times of the generator and the discriminator reach the target test, generate the verification fault data of the power system under the test conditional variables through the fusion generative adversarial network model; Evaluate the verification fault data through the real fault data corresponding to the test conditional variables to obtain a quality evaluation result; Update the training parameters of the generator and the discriminator according to the quality assessment result.

[0027] As can be seen from the above description, based on iterative training, the quality of the fault data generated by the model is evaluated regularly. The generated verification fault data of the model is compared with the real fault data under the same test condition variables, and the training parameters of the generator and the discriminator are further optimized. This verification and optimization mechanism can ensure the performance and reliability of the model in practical applications, avoid overfitting or underfitting, and enable the model to generate simulation data closer to real fault data.

[0028] Furthermore, the verification fault data is evaluated by the real fault data corresponding to the test condition variables, and the quality assessment result includes: Quantify the data similarity between the real fault data and the verification fault data by mean square error, and quantify the data diversity of the verification fault data by Shannon index; Generate a quality assessment result according to the data similarity and the data diversity.

[0029] As can be seen from the above description, quantifying the data similarity and diversity of the data generated by the model by mean square error and Shannon index respectively can comprehensively and accurately evaluate the quality of the generated data, ensure that the generated data not only has a high similarity with the real data, but also can accurately generate data for a variety of fault scenarios, avoid ineffective repeated generation of data, provide more reliable simulation data support for the fault analysis and processing of the power system, and help improve the accuracy and effectiveness of fault handling.

[0030] Another embodiment of the present invention provides a fault scenario simulation terminal based on a deep learning fusion model, including a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the computer program, each step in the above-mentioned fault scenario simulation method based on a deep learning fusion model is implemented.

[0031] As can be seen from the above description, the beneficial effects of the present invention are as follows: The fault training data of the power system belongs to time series data. An encoder is introduced into the generative adversarial network including a generator and a discriminator. The time series data is compressed into a low-dimensional hidden state vector through the encoder, enabling the generative adversarial network, which was originally difficult to capture the characteristics of time series data, to effectively extract important features such as the time dependence, period, and trend of the time series data. Moreover, in the generative adversarial network including a generator and a discriminator, a conditional variable of the data is introduced, enabling the generator to control the characteristics of the simulated fault data based on the conditional variable, effectively enhancing the scenario pertinence and type diversity of the fault data. At the same time, the generative adversarial network conducts mutual adversarial training through the generator and the discriminator. The generator generates simulated fault data as realistic as possible, making it difficult for the discriminator to distinguish the generated data from the real data, and the discriminator accurately distinguishes the generated data from the real data, enabling the generator to continuously learn how to generate simulated fault data closer to the real data during the iterative training process, thereby improving the accuracy of the generated data. By integrating the encoder and the conditional variable in the generative adversarial network, the generative adversarial network can accurately generate time series data under multiple fault scenarios based on the conditional variable, thereby optimizing the effect of the power system in realizing fault simulation.

[0032] The above-mentioned method and terminal for simulating fault scenarios based on a deep learning fusion model of the present invention are applicable to the power grid fault reconstruction scenario, can improve the authenticity and diversity of the fault simulation data, enhance the ability to capture complex fault patterns, and thus optimize the effect of the power system in realizing fault simulation. The following is described through specific embodiments: Please refer to Figures 1 to 2 , the first embodiment of the present invention is: A method for simulating fault scenarios based on a deep learning fusion model, including: S1. Input the fault training data of the power system under different conditional variables into a preset fusion generative adversarial network model, where the fusion generative adversarial network model includes an encoder, a generator, and a discriminator.

[0033] Among them, the conditional variables include the fault type to which the data belongs, the acquisition time, and environmental factors, etc. In this embodiment, the conditional variable is the fault type of the data.

[0034] S2. Compress the fault training data through the encoder to obtain a hidden state vector.

[0035] Specifically, the encoder includes a first long short-term memory network layer, and step S2 includes: S21. Extract key fault features from the fault training data through the first long short-term memory network layer, and compress the key fault features into a hidden state vector. In this way, the input multi-modal time series data can be compressed into a hidden state vector in the latent space by the encoder.

[0036] S3. Generate data for the conditional variable through the generator using the hidden state vector and the conditional variable corresponding to the hidden state vector, to obtain simulated fault data under the conditional variable.

[0037] Specifically, the generator includes a second long short-term memory network layer and a first fully connected layer using a hyperbolic tangent activation function, then step S3 includes: S31. Convert the conditional variable corresponding to the hidden state vector into a conditional vector through one-hot encoding.

[0038] S32. Combine the conditional vector with the obtained random noise vector to obtain a target vector.

[0039] S33. Combine the hidden state vector and the target vector to obtain an input vector.

[0040] S34. Fuse the fault features of the input vector through the second long short-term memory network layer to obtain generated data.

[0041] S35. Map the generated data to the target data dimension through the first fully connected layer to obtain simulated fault data under the conditional variable. Among them, the hyperbolic tangent activation function is used to ensure that the value of the output simulated fault data is within a reasonable range to match the range of actual data.

[0042] S4. Distinguish the simulated fault data under the conditional variable and the historical fault data through the discriminator, to obtain the first loss function of the generator and the second loss function of the discriminator.

[0043] Specifically, the discriminator includes a third long short-term memory network layer and a second fully connected layer using a sigmoid activation function, then step S4 includes: S41. Extract the joint features between the conditional variable and the time series data from the simulated fault data and the historical fault data respectively through the third long short-term memory network layer.

[0044] S42. Map the joint features to scalar data through the second fully connected layer to obtain the probability values that the simulated fault data and the historical fault data are judged as real data. Among them, the sigmoid activation function is used to map the output data to probability values, and the range of probability values is from 0 to 1.

[0045] S43. Calculate the first loss function of the generator and the second loss function of the discriminator according to the probability value.

[0046] In some embodiments, the default number of network layers for the first long short-term memory network layer, the second long short-term memory network layer, and the third long short-term memory network layer is two, and the number of hidden units included in each layer of the network is 128.

[0047] In some embodiments, both the first loss function and the second loss function use cross-entropy loss. The first loss function can be specifically expressed as , where L G represents the first loss function, E represents the expected value, z~p z (z) represents sampling noise p z (z) from the prior distribution z , G(z|c) represents the generator G receiving the noise z and the conditional variable c to generate data, D represents the discriminator, logD(G(z|c)) represents the log probability of the discriminator's discrimination result of the generated data. The goal of the generator is to maximize the first loss function so that the discriminator believes the generated data is real. The second loss function can be specifically expressed as , where L D represents the second loss function, E represents the expected value, x~p data (x) represents sampling the real data x from the real data distribution p data (x) , D(x|c) represents the discrimination result output by the discriminator D after receiving the real data x and the conditional variable c, logD(x|c) represents the log probability of the discriminator's discrimination result of the real data, log(1 - D(G(z|c))) represents the log probability of the discriminator's discrimination result of the generated data. The goal of the discriminator is to minimize the second loss function so that the discriminator can correctly distinguish between real data and generated data.

[0048] In addition, in order to ensure that the data generated by the generator is consistent with the real data in the feature space, a content loss function needs to be added to the first loss function. The content loss function is specifically expressed as: ; where L content represents the content loss function,n Indicates the number of feature vectors, x i represents the feature vector of the i-th real data. The objective of the content loss function is to minimize the difference between the generated data and the real data in the feature space, ensuring that the generated data is similar to the real data in content.

[0049] S5. Iteratively train the fusion generative adversarial network model according to the first loss function and the second loss function to obtain a fault scenario simulation model.

[0050] Specifically, step S5 includes: S51. Alternately update the training parameters of the generator and the discriminator according to the first loss function and the second loss function until the first loss function reaches the maximum and the second loss function reaches the minimum to obtain target parameters.

[0051] In some embodiments, the generator and the discriminator are alternately trained based on the data batch size and the number of iterations in each predetermined training cycle. For each step of training, the first loss function and the second loss function need to be calculated, and the corresponding training parameters are updated. Monitor the loss of the dataset during training. If the loss does not decrease for several consecutive training cycles, early stopping is implemented to prevent overfitting.

[0052] S52. Configure the generator and the discriminator according to the target parameters to obtain a fault scenario simulation model.

[0053] Wherein, after alternately updating the training parameters of the generator and the discriminator according to the first loss function and the second loss function, it further includes: S501. If it is detected that the update times of the generator and the discriminator reach the target test, generate verification fault data of the power system under the test condition variables through the fusion generative adversarial network model.

[0054] S502. Evaluate the verification fault data through the real fault data corresponding to the test condition variables to obtain a quality evaluation result.

[0055] Specifically, step S502 includes: S5021. Quantify the data similarity between the real fault data and the verification fault data through the mean square error, and quantify the data diversity of the verification fault data through the Shannon index.

[0056] In some embodiments, specifically quantifying the data similarity between the real fault data and the verification fault data through the mean square error is: ; Among them, MSE represents the mean square error, n represents the total number of data, and Y i represents the true fault data, and represents the verified fault data.

[0057] In some embodiments, quantifying the data diversity of the verified fault data by the Shannon index specifically is: Dversity Index = , where Dversity Index represents the data diversity, n represents the total number of sample types, p i represents the relative abundance of the samples of the i-th type, that is, the ratio of the number of samples of the i-th type to the total number of samples.

[0058] S5022. Generate a quality assessment result according to the data similarity and the data diversity.

[0059] In some embodiments, the fault training data includes a training data set and a test data set. Among them, the training data set is used to alternately update the training parameters of the generator and the discriminator, and the test data set is used to evaluate the quality assessment result of the simulated data output by the model.

[0060] In some embodiments, the quality of the fusion generative adversarial network model can also be evaluated, which specifically includes: 1. Analyze the specific output of the model to find any possible problems, such as mode collapse (the generator generates very similar or repetitive outputs), or overfitting (the generated data is too similar to the training data to generalize). 2. Perform k-fold cross-validation on the training data set to evaluate the performance of the model on different data subsets to ensure the generalization ability of the model. 3. Apply the generated simulated fault data to actual fault detection, diagnosis, or prediction tasks to evaluate its effect and contribution in actual applications, and ensure that the generated data can effectively improve the overall performance of the system.

[0061] S503. Update the training parameters of the generator and the discriminator according to the quality assessment result.

[0062] In some embodiments, the training parameters include the learning rate of the model, the number of layers and units of the long short-term memory network layer. By updating the training parameters, a parameter combination that optimizes the model performance is determined. For example, by modifying the structures of the generator and the discriminator, that is, increasing or decreasing the number of layers of the long short-term memory network layer, the complexity of the generator and the discriminator is adjusted, so as to reduce problems such as mode collapse or overfitting.

[0063] After determining the fault scenario simulation model, the true fault data of the power system can be continuously collected to optimize the training of the fault scenario simulation model to adapt to new fault modes that may appear in the power system.

[0064] S6. Generate target fault data of the power system under target condition variables through the fault scenario simulation model.

[0065] In some embodiments, step S6 is specifically as follows: 1. Set specific condition variables to simulate different fault scenarios. The condition variables include fault type, occurrence time, environmental factors, grid load status, etc., to ensure that each fault scenario has precise condition variables. 2. Input the condition variables under different fault scenarios into the fault scenario simulation model to generate simulated fault data under the corresponding scenarios. 3. Check the consistency and accuracy of the generated data through automated scripts and statistical analysis tools to ensure that all data meet the set conditions. 4. Recombine the generated modal components to form complete time series data, and label detailed fault scenarios and condition variables for each data sample to ensure that the distribution of the model output is closer to the real data distribution.

[0066] Embodiment 2 of the present invention is as follows: A fault scenario simulation method based on a deep learning fusion model, which is different from Embodiment 1 in that before step S1, it further includes: S101. Obtain historical operation data of the power system under different condition variables, where the historical operation data includes historical non-fault data and historical fault data.

[0067] It should be noted that the historical operation data of the power system is composed of multiple sets of time series data. For example, power data such as current, voltage, and frequency of the power system are collected and observed in chronological order. Since the historical operation data of the power system may come from different database sources, there are certain differences in the data formats of the historical operation data. Before decomposing the historical operation data into signals, it is necessary to process the data formats to ensure that the formats of all historical operation data are consistent for subsequent processing.

[0068] In some embodiments, before step S102, it further includes preprocessing the obtained historical operation data. The preprocessing process is specifically as follows: 1. Delete or correct significantly incorrect historical operation data to ensure the continuity and integrity of the time series data. 2. Identify outliers in the historical operation data and correct or remove them. In an alternative embodiment, the quartile rule can be used to identify outliers in the data. When the data x < Q1 - 1.5×(Q3 - Q1) or the data x > Q3 + 1.5×(Q3 - Q1), the data x is an outlier, where Q1 represents the first quartile and Q3 represents the third quartile. 3. Perform normalization and standardization processing on the historical operation data to reduce numerical instability in model training.

[0069] S102. Decompose and combine the historical operation data through a variational mode model to obtain fault training data under different conditional variables.

[0070] Among them, step S102 includes: S1021. Decompose the historical operation data into multiple modal components and corresponding central frequencies for each modal component through the variational mode model.

[0071] In some embodiments, determine the number of modes K to be decomposed according to the complexity of the time series data and the feature levels to be captured, and decompose each time series data into K modal components and corresponding central frequencies through the variational mode model.

[0072] S1022. Select key modal components related to the conditional variable from the multiple modal components according to the central frequencies.

[0073] In some embodiments, analyze the frequency characteristics (such as central frequency and bandwidth) of each modal component, and screen out key modal components related to the fault type from them.

[0074] S1023. Set combination weights according to the contribution degree of the key modal components to the conditional variable.

[0075] S1024. Recombine the key modal components according to the combination weights to obtain fault training data under different conditional variables.

[0076] It should be noted that in order for the generator and discriminator to process multi-modal data, the long short-term memory network layers in the generator and discriminator first process different modal components in parallel, and then fuse their different modal features to ensure that the information of each modal component can be effectively utilized in the generation and discrimination processes.

[0077] In some embodiments, after obtaining the fault training data under different conditional variables, regard the decomposed key modal components as independent features for dataset division, so as to divide the fault training data into a training dataset, a validation dataset, and a test dataset. During the division process, it is necessary to ensure that all key modal components decomposed from the same time series data appear in the same dataset simultaneously to maintain data consistency. At the same time, each key modal component retains the conditional variable corresponding to the original time series data. Among them, the training dataset is used to determine the first loss function and the second loss function to alternately update the training parameters of the generator and discriminator, and the validation dataset and the test dataset are used to determine the quality evaluation results to update the training parameters of the generator and discriminator. In addition, data augmentation methods can be adopted for each modal component to increase data diversity and improve the generalization ability of the training model.

[0078] Please refer to Figure 3 , the third embodiment of the present invention is as follows: A fault scenario simulation terminal 100 based on a deep learning fusion model, including a memory 101, a processor 102, and a computer program stored on the memory 101 and running on the processor 102. When the processor 102 executes the computer program, each step in a fault scenario simulation method based on a deep learning fusion model in the above-mentioned first or second embodiment is implemented.

[0079] In summary, the present invention provides a fault scenario simulation method and terminal based on a deep learning fusion model. The historical operation data of the power system is decomposed and recombined through a variational mode model. While effectively extracting the multi-scale features of the signal, the diversity of the data is increased through recombination. Processing the historical operation data based on the variational mode model provides rich and more decoupled feature-based data for subsequent model training. In addition, compared with the traditional generative adversarial network, an encoder and a conditional variable are simultaneously introduced in the fusion generative adversarial network. First, the encoder compresses the time series data into a low-dimensional hidden state vector, enabling the generative adversarial network that originally had difficulty capturing the features of time series data to effectively extract important features such as the time dependence, period, and trend of the time series data. And the generator can control the characteristics of the simulated fault data based on the conditional variable, effectively enhancing the scenario pertinence and type diversity of the fault data. At the same time, in the fusion generative adversarial network, the generator and the discriminator are trained against each other. The generator generates simulated fault data as realistic as possible, making it difficult for the discriminator to distinguish between the generated data and the real data. The discriminator accurately distinguishes between the generated data and the real data, enabling the generator to continuously learn how to generate simulated fault data closer to the real data during the iterative training process, thereby improving the accuracy of the generated data. Combining the variational mode model and the fusion generative adversarial network model to train the fault scenario simulation model not only improves the authenticity and diversity of the simulated fault data, but also enhances the model's ability to capture complex fault patterns, significantly improving the effect of power system fault simulation and reconstruction.

[0080] The above are only the embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in the related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A fault scenario simulation method based on a deep learning fusion model, characterized in that Including: Inputting the fault training data of the power system under different condition variables into a preset fusion generative adversarial network model, where the fusion generative adversarial network model includes an encoder, a generator, and a discriminator; Compressing the fault training data through the encoder to obtain a hidden state vector; Generating data through the generator using the hidden state vector and the corresponding condition variable to obtain simulated fault data under the condition variable; Distinguishing the simulated fault data under the condition variable and the historical fault data through the discriminator to obtain a first loss function of the generator and a second loss function of the discriminator; Iteratively training the fusion generative adversarial network model according to the first loss function and the second loss function to obtain a fault scenario simulation model; Generating target fault data of the power system under the target condition variable through the fault scenario simulation model.

2. The method according to claim 1, characterized in that Before inputting the fault training data of the power system under different condition variables into a preset fusion generative adversarial network model, it further includes: Obtaining historical operation data of the power system under different condition variables, where the historical operation data includes historical non-fault data and historical fault data; Decomposing and combining the historical operation data through a variational mode model to obtain fault training data under different condition variables.

3. The method according to claim 2, wherein Decomposing and combining the historical operation data through a variational mode model to obtain fault training data under different condition variables includes: Decomposing the historical operation data through the variational mode model into multiple mode components and the corresponding central frequency for each mode component; Selecting key mode components related to the condition variable from the multiple mode components according to the central frequency; Setting combination weights according to the contribution degree of the key mode components to the condition variable; Recombining the key mode components according to the combination weights to obtain fault training data under different condition variables.

4. The method according to claim 1, wherein The encoder includes a first long short-term memory network layer; Compressing the fault training data through the encoder to obtain a hidden state vector includes: Extracting key fault features from the fault training data through the first long short-term memory network layer and compressing the key fault features into the hidden state vector.

5. The method according to claim 1, wherein The generator includes a second long short-term memory network layer and a first fully connected layer using a hyperbolic tangent activation function; Generating data through the generator using the hidden state vector and the corresponding condition variable to obtain simulated fault data under the condition variable includes: Converting the condition variable corresponding to the hidden state vector into a condition vector through one-hot encoding; Combining the condition vector with a randomly obtained noise vector to obtain a target vector; Combining the hidden state vector and the target vector to obtain an input vector; Fusing fault features of the input vector through the second long short-term memory network layer to obtain generated data; Map the generated data to the target data dimension through the first fully connected layer to obtain the simulated fault data under the conditional variable.

6. The method according to claim 1, wherein The discriminator includes a third long short-term memory network layer and a second fully connected layer using a sigmoid activation function; Distinguish the simulated fault data and the historical fault data under the conditional variable through the discriminator to obtain the first loss function of the generator and the second loss function of the discriminator, including: Extract the joint features between the conditional variable and the time series data from the simulated fault data and the historical fault data respectively through the third long short-term memory network layer; Map the joint features to scalar data through the second fully connected layer to obtain the probability values that the simulated fault data and the historical fault data are determined to be real data; Calculate the first loss function of the generator and the second loss function of the discriminator according to the probability values.

7. The method according to claim 1, wherein Iteratively train the fusion generative adversarial network model according to the first loss function and the second loss function to obtain a fault scenario simulation model, including: Alternately update the training parameters of the generator and the discriminator according to the first loss function and the second loss function until the first loss function reaches the maximum and the second loss function reaches the minimum to obtain the target parameters; Configure the generator and the discriminator according to the target parameters to obtain a fault scenario simulation model.

8. The method according to claim 7, wherein After alternately updating the training parameters of the generator and the discriminator according to the first loss function and the second loss function, it further includes: If it is detected that the update times of the generator and the discriminator reach the target test, generate the verification fault data of the power system under the test conditional variable through the fusion generative adversarial network model; Evaluate the verification fault data through the real fault data corresponding to the test conditional variable to obtain a quality evaluation result; Update the training parameters of the generator and the discriminator according to the quality evaluation result.

9. The method according to claim 8, wherein Evaluating the verification fault data through the real fault data corresponding to the test conditional variable to obtain a quality evaluation result includes: Quantify the data similarity between the real fault data and the verification fault data through the mean square error, and quantify the data diversity of the verification fault data through the Shannon index; Generate a quality evaluation result according to the data similarity and the data diversity.

10. A fault scenario simulation terminal based on a deep learning fusion model, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements each step in a fault scenario simulation method based on a deep learning fusion model according to any one of claims 1-9.

Images