Power system risk assessment method and device, equipment and storage medium

Abstract

The invention relates to the technical field of risk assessment, and discloses an electric power system risk assessment method, device and equipment and a storage medium, and the method comprises the steps: obtaining multi-source operation related data of an electric power system, employing united Kalman filtering based on a variational information bottleneck mechanism, carrying out the noise filtering and data fusion of the multi-source operation related data, and obtaining the risk assessment result of the electric power system. Obtaining a system state feature vector; inputting the system state feature vector into a risk assessment model to obtain a risk assessment result; the risk assessment model is constructed based on a Fourier neural operator and a fuzzy neural network, the Fourier neural operator is used for extracting physical prior features of a system state feature vector, and the fuzzy neural network is used for performing risk assessment on the physical prior features and the system state feature vector. And the fuzzy neural network is obtained by training with an improved whale optimization algorithm. According to the invention, real-time monitoring and robust evaluation of the operation state of the power grid can be realized, and the operation safety and stability of the power system are further ensured.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, and in particular to a method, apparatus, equipment and storage medium for risk assessment of power systems. Background Technology

[0002] The output of new energy sources such as wind and solar power in new power systems is significantly intermittent and fluctuating, which greatly increases the sensitivity of the power grid to weather changes. Extreme weather events can lead to accidents such as equipment damage and large-scale power outages, thus posing a serious threat to the safe and stable operation of the power system. At present, the main technologies for risk research on power systems are divided into traditional physical modeling methods and statistical analysis methods, but both have certain limitations.

[0003] While traditional physical modeling methods have certain physical significance in specific scenarios, they generally do not consider all factors and are difficult to fully incorporate dynamic factors from multiple fields such as meteorology and load response into the evaluation system. Moreover, they are highly subjective, and the establishment of models and parameter settings rely on expert experience, making it difficult to meet the needs of real-time evaluation in the face of occasional extreme weather and increased power grid complexity.

[0004] Statistical analysis methods primarily employ regression models or principal component analysis to assess power outage risks by analyzing historical data. However, for new power systems or novel extreme weather events, sufficient data support is often lacking. Furthermore, the interaction between power systems and meteorological systems is highly nonlinear and dynamically coupled, making it difficult for traditional statistical methods to accurately characterize this complex relationship. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a power system risk assessment method, apparatus, equipment, and storage medium. By employing a joint Kalman filter based on a variational information bottleneck mechanism and a physics-guided risk assessment model, it is possible to accurately assess the operational risks of the power system, thereby improving the safety and stability of power system operation.

[0006] In a first aspect, the present invention provides a method for power system risk assessment, the method comprising: Multi-source operation-related data of the power system are acquired, and joint Kalman filtering based on variational information bottleneck mechanism is used to filter noise and fuse data to obtain system state feature vector; The system state feature vector is input into a pre-constructed risk assessment model to obtain a risk assessment result. The risk assessment model is constructed based on Fourier neural operators and fuzzy neural networks. The Fourier neural operators are used to extract the physical prior features of the system state feature vector. The fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector. The fuzzy neural network is trained using an improved whale optimization algorithm.

[0007] Furthermore, the step of employing joint Kalman filtering based on a variational information bottleneck mechanism to filter noise and fuse data from the multi-source operational data to obtain the system state feature vector includes: Multiple sub-filters are used to process the multi-source operational data collected by their respective sensors; The state vector output from each sub-filter is input into a pre-trained variational information bottleneck encoder to obtain Gaussian pseudo-measurement values ​​and Gaussian pseudo-measurement noise matrix. Based on the Gaussian pseudo-measurement value and the Gaussian pseudo-measurement noise matrix, the measurement value and measurement noise matrix in each sub-filter are updated, and the state information of each sub-filter is generated based on the updated measurement value and measurement noise matrix. The state information of each sub-filter is fused using the main filter to obtain the system state feature vector.

[0008] Furthermore, the risk assessment model consists of cascaded Fourier neural operators and fuzzy neural networks; The Fourier neural operator is used to physically guide the input system state feature vector to obtain a risk precursor field and physical prior features. The risk precursor field is the physical prior vector output by the last layer of the Fourier neural operator, and the physical prior features are obtained by feature extraction from the risk precursor field. The fuzzy neural network is used to fuzzify the input concatenated feature vector to obtain the risk assessment result. The concatenated feature vector is obtained by concatenating the system state feature vector and the physical prior features.

[0009] Furthermore, the center and standard deviation of the membership function of the fuzzy neural network are dynamically adjusted and determined based on the risk precursor field.

[0010] Furthermore, the training steps of the fuzzy neural network include: Population encoding and initialization are performed on the model parameters of the fuzzy neural network, and initial population positions are randomly generated; According to the preset search strategy, the whale optimization algorithm is used to optimize the position of the initial population to obtain the global optimal position; Based on the global optimal position, the iterative stagnation method is used to determine whether a local optimum has been reached; If a local optimum is reached, the global optimum position is taken as the current elite solution, and an elite reverse learning strategy is used to generate an elite reverse solution. The fitness value of the elite reverse solution is compared with the fitness value of the current elite solution, and the global optimal position is updated according to the comparison result. Determine whether the iteration stopping condition has been met. If the iteration stopping condition has been met, then the global optimal position is used as the optimal model parameter of the fuzzy neural network. If the iteration stopping condition is not met, the whale optimization algorithm is used to continue the position search until the iteration stopping condition is met.

[0011] Furthermore, the step of using the whale optimization algorithm to optimize the initial population position according to a preset search strategy to obtain the globally optimal position includes: Determine if the norm of the coefficient vector is less than the norm threshold. If it is, use a local search strategy; otherwise, use a global search strategy. The initial population position is optimized according to the selected search strategy to obtain the global optimal position.

[0012] Furthermore, the step of optimizing the initial population position according to the selected search strategy to obtain the globally optimal position includes: In response to the use of a local search strategy for location optimization, it is determined whether the probability of strategy selection is less than the probability threshold. If so, a shrinking encirclement strategy is adopted to update the individual's position; otherwise, a bubble net predation strategy is adopted to update the individual's position. In response to employing a global search strategy for position optimization, an average optimal position is generated through quantum behavior, and the individual position is updated based on the average optimal position; The updated global optimal position is obtained based on the updated fitness value of the individual position.

[0013] Secondly, the present invention provides a power system risk assessment device, the device comprising: The data fusion module is used to acquire multi-source operation-related data of the power system. It uses a joint Kalman filter based on the variational information bottleneck mechanism to perform noise filtering and data fusion on the multi-source operation-related data to obtain the system state feature vector. The risk assessment module is used to input the system state feature vector into a pre-constructed risk assessment model to obtain a risk assessment result. The risk assessment model is constructed based on Fourier neural operators and fuzzy neural networks. The Fourier neural operators are used to extract the physical prior features of the system state feature vector. The fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector. The fuzzy neural network is trained using an improved whale optimization algorithm.

[0014] Thirdly, embodiments of the present invention also provide a computer device, including 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-described method.

[0015] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0016] This invention provides a method, apparatus, device, and storage medium for power system risk assessment. It achieves power system operational risk assessment under extreme weather scenarios through an integrated link of multi-source information fusion, physical prior injection, fuzzy inference, global optimization, and risk output. This invention embeds variational information bottlenecks in the joint filtering of information forms for denoising, improving the robustness of data fusion. By embedding physical guidance into the risk assessment model, it enhances the perception capability and generalization ability of the model in extreme scenarios. Furthermore, by optimizing model parameters through an improved whale optimization algorithm, it improves the model's assessment accuracy. This invention enables real-time monitoring and robust assessment of the power grid's operating status, thereby ensuring the safety and stability of power system operation. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the power system risk assessment method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the power system risk assessment device in an embodiment of the present invention; Figure 3 This is an internal structural diagram of the computer device in an embodiment of the present invention.

[0018] Figure label: 10. Data fusion module; 20. Risk assessment module. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 The first embodiment of the present invention proposes a power system risk assessment method, including steps S10 to S20: Step S10: Obtain multi-source operation-related data of the power system, and use joint Kalman filtering based on variational information bottleneck mechanism to perform noise filtering and data fusion on the multi-source operation-related data to obtain system state feature vector; Step S20: Input the system state feature vector into a pre-constructed risk assessment model to obtain a risk assessment result; wherein, the risk assessment model is constructed based on Fourier neural operators and fuzzy neural networks, the Fourier neural operators are used to extract the physical prior features of the system state feature vector, the fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector, and the fuzzy neural network is trained using an improved whale optimization algorithm.

[0021] This invention provides a method for assessing the risk of a power system. In actual operation, a power system is affected by various factors, such as external weather and internal equipment. For example, extreme weather or physical damage to critical internal equipment may lead to various fault risks. In order to accurately assess the current operating status of the power system, this invention conducts a comprehensive analysis of multi-source data related to the operation of the power system to assess the operating risk of the power system.

[0022] This embodiment acquires multi-source heterogeneous data related to power system operation through multiple sensors and meteorological monitoring stations installed in the power system. This includes weather data and operational data. Operational data includes node voltages, currents, and load values ​​on the power source, grid, and load sides. Weather data includes wind speed, rainfall / snowfall, and duration of sustained high temperatures. Specific data can be flexibly set according to the actual site conditions, and no specific limitations are imposed here. Because the collected data has different types and structures, data processing is required before risk assessment. This processing includes data filtering and fusion, such as outlier handling, interpolation filling, data alignment, and format standardization. Kalman filtering, particle swarm filtering, or statistical learning methods are then used for fusion to obtain the fused characteristics of the multi-source data.

[0023] In a preferred embodiment, to accurately characterize the coupling relationship between weather data and power grid operating status, this invention provides a method for noise filtering and data fusion of multi-source data based on a variational information bottleneck mechanism using joint Kalman filtering. The specific steps include: Multiple sub-filters are used to process the multi-source operational data collected by their respective sensors; The state vector output from each sub-filter is input into a pre-trained variational information bottleneck encoder to obtain Gaussian pseudo-measurement values ​​and Gaussian pseudo-measurement noise matrix. Based on the Gaussian pseudo-measurement value and the Gaussian pseudo-measurement noise matrix, the measurement value and measurement noise matrix in each sub-filter are updated, and the state information of each sub-filter is generated based on the updated measurement value and measurement noise matrix. The state information of each sub-filter is fused using the main filter to obtain the system state feature vector.

[0024] In this embodiment, the joint Kalman filter algorithm is a distributed structure comprising a master filter and multiple local sub-filters. The conventional processing steps of this algorithm are that each sub-filter independently processes its own sensor data, and then the processed state information is fused with the master filter (time update and measurement update) to obtain the globally optimal estimate. For the i-th sub-filter, the discrete model of the joint Kalman filter can be expressed as: In the formula, k represents the time step. It is the state vector of the i-th sub-filter at time k, which integrates the key state information of the power system. and These are the state transition matrix and control input matrix at time k, respectively. It is the process noise vector at time k. It is the observation vector of the i-th sub-filter at time k, which is the value directly measured by the sensor, i.e., the measurement value. Let be the observation matrix of the i-th sub-filter at time k. Let represent the observation noise vector of the i-th sub-filter at time k.

[0025] Each sub-filter independently calculates its own information contribution, and the resulting information pairs are obtained after local fusion and updating. And send it to the main filter: In the formula, This is the state vector of the i-th sub-filter at time k included in this measurement. Information matrix, This is the state vector of the i-th sub-filter at time k included in this measurement. Information vector, Let represent the measurement noise matrix of the i-th sub-filter at time k, which is the covariance matrix of the observation noise vector. Let be the observation matrix of the i-th sub-filter at time k, which can be identified by the least squares method.

[0026] Joint filters are particularly convenient for handling multi-sensor fusion problems because information from different sensors can be fused through simple matrix addition. In the main filter, to obtain the information matrix of the main filter at time k... and information vector To make a prediction, the state at time k can be predicted based on the posterior information matrix and posterior information vector at time k-1. The specific formula is as follows: In the formula, This represents the posterior information matrix of the master filter at time k-1. The inverse of the covariance matrix of the process noise is represented. This represents the posterior information vector of the master filter at time k-1. Let k represent the prior information vector at time k. Let represent the prior information matrix of the master filter at time k.

[0027] The prior estimate of the main filter at time k is obtained. Then, the measured values ​​are used to correct the prior estimate obtained in the prediction stage, thereby obtaining a more accurate posterior estimate at time k. This completes the information fusion of a distributed, multi-source, heterogeneous, large-scale system. The specific formula is as follows: in, It is the posterior information matrix of the master filter at time k. It is the posterior information vector of the master filter at time k. This is the state vector of the i-th sub-filter at time k included in this measurement. Information matrix, This is the state vector of the i-th sub-filter at time k included in this measurement. The information vector, where N is the total number of sub-filters.

[0028] As can be seen, after receiving information from all sub-filters, the main filter accumulates it with the predicted information to achieve global optimal fusion. The joint Kalman filter described above can be used to fuse multi-source operational data.

[0029] To further filter redundant information, extract key information, and improve the effectiveness of fused features in risk assessment, this embodiment embeds a variational information bottleneck (VIB) mechanism in the joint Kalman filtering process. VIB is a deep learning regularization method based on information bottleneck theory, which aims to optimize the intermediate representation of the model so that it retains information useful for the prediction task while reducing its dependence on redundant input information.

[0030] Specifically, during the training phase, for the state vector of the i-th sub-filter... The reparameterized sampling and classification losses are obtained using the encoder and decoder in VIB, respectively. The network is then trained according to the VIB classification total loss formula, where the total loss function... It can be represented as: In the formula, E(*) represents the expectation function, and KL(*) represents the KL divergence. x As training samples, y For monitoring targets (classification labels or consistency / pseudo-labels). z This represents an information bottleneck. Let z be the posterior distribution of the encoder given the input x. This is the conditional probability distribution for the decoder to generate y given z. Let z be the prior distribution of z; , is the weighting factor; φ and θ All are distributed parameters.

[0031] The left-hand side of the total loss function makes the information bottleneck z effective for task discrimination, while the right-hand side forces z to discard and compress irrelevant information, thus achieving effective denoising. For a trained encoder-decoder, only the encoder needs to be retained during online deployment.

[0032] By inputting the state vector output from the sub-filter into the encoder of the VIB, we can obtain the Gaussian pseudo-measurement value and the Gaussian pseudo-measurement noise matrix: In the formula, This represents the state vector of the i-th sub-filter at time k. Under the condition that the information bottleneck of the i-th sub-filter at time k The posterior distribution of follows a normal distribution. and Let $\mathbf{i}$ be the mean and covariance of the information bottleneck of the $i$-th sub-filter at time $k$, and let $\mathbf{i}$ be a symmetric positive definite matrix.

[0033] The state vector of the i-th sub-filter at time k. Input the trained encoder This yields the mean and covariance corresponding to the information bottleneck. The mean is used as the Gaussian pseudo-measurement value, and the covariance is used as the Gaussian pseudo-measurement noise matrix. These two values ​​then replace the original measurement value and measurement noise matrix of the sub-filter. The original measurement value refers to the observation vector of the i-th sub-filter, and the original measurement noise matrix refers to the covariance matrix of the observation noise vector of the i-th sub-filter. The formula can be expressed as: As can be seen, this embodiment uses the covariance matrix automatically learned by the VIB encoder. Replace traditional measurement noise matrix The mean value automatically learned using VIB's encoder. Replace the measurements in a traditional Kalman filter The information of each sub-filter is compared using Gaussian pseudo-measurements and the Gaussian pseudo-measurement noise matrix. After the update, the main filter will fuse the updated information pairs to obtain the system state feature vector.

[0034] In this embodiment, the measurements of each sub-mode are first denoised and compressed using VIB (Variable Input Method) to obtain a pseudo-measurement characterization containing "measurement mean and uncertainty". Then, within the information form framework of joint Kalman filtering, the "information increments" of each sub-mode are converged to the main filter to complete the globally optimal fusion. This process achieves seamless integration without structural changes, while automatically reducing the weight of modes with high uncertainty during fusion, improving robustness and consistency under extreme weather and anomalous data, and maintaining auditability and simplicity of engineering implementation.

[0035] After obtaining the system state feature vector through data fusion, the vector is input into a pre-established risk assessment model to conduct risk assessment, thereby obtaining the final risk assessment result. The risk assessment model can be built based on a neural network model, such as a fuzzy neural network, a long short-term memory neural network, or a deep convolutional neural network.

[0036] In a preferred embodiment, since the fused data are all continuous time series, and the main physical laws controlling weather evolution are dominated by slowly changing low-frequency components, the rapidly fluctuating high-frequency components can generally be considered noise. Therefore, this embodiment introduces a physics-guided neural operator pre-model into the risk assessment model, incorporating physical prior features into the input of the fuzzy neural network to achieve synergy between global physical modeling and data-driven reasoning. Specifically, the risk assessment model consists of cascaded Fourier Neural Operators (FNOs) and Fuzzy Neural Networks (FNNs). The FNO is used to physically guide the input system state feature vector to obtain the risk precursor field and physical prior features. The risk precursor field is the physical prior vector output by the last layer of the Fourier Neural Operator. Feature extraction from the risk precursor field yields the physical prior features, which are then concatenated with the system state feature vector before being input into the FNN for risk assessment.

[0037] In the risk assessment model of this embodiment, the system state feature vector is first input into FNO. FNO performs dimensionality upscaling on the system state feature vector, and then transforms the upscaled vector to the frequency domain through Fourier transform. Then, a frequency domain linear transformation is performed on the transformed frequency domain data. During the frequency domain linear transformation, only the first K low-frequency modes are transformed. Then, the vector is transformed to the time domain through inverse Fourier transform. Finally, residual superposition and nonlinear processing are performed to obtain the output data, which is the physical prior vector obtained by physical guidance.

[0038] The physical prior vector output from the last layer of FNO is used as the risk precursor field. Then, feature extraction is performed on the risk precursor field to obtain physical prior features. Preferably, the mean, quantile, overthreshold time proportion, and low-frequency energy proportion in the frequency domain of the risk precursor field can be used as physical prior features.

[0039] After obtaining the physical prior features, they are concatenated with the original system state feature vector to obtain a concatenated feature vector. The concatenated feature vector is then input into the FNN for risk assessment. Compared with directly inputting the system state feature vector into the FNN, the concatenated vector can enhance the FNN's ability to perceive extreme scenarios while maintaining interpretability.

[0040] The fuzzy neural network can be divided into four layers. The first layer is the input layer, which in this embodiment receives the concatenated feature vector after physical information fusion and passes it to the next layer. The second layer is the fuzzification layer, which represents the fuzzy subset of fuzzy linguistic variables. It first uses a Gaussian membership function to calculate the degree to which the input variable belongs to a certain fuzzy set, then calculates the fitness according to each rule, combines the membership degrees of all input variables, calculates the T-norm of the membership degree of each rule, and finally normalizes the trigger strength of each rule. The third layer is the fuzzy inference layer, which is the hidden layer of the BP neural network. This layer calculates the output of each rule and combines it with the normalized trigger strength to map the fuzzy values ​​of the input variables to the fuzzy values ​​of the output variables. Finally, the output layer defuzzifies the fuzzy inference results and outputs the final risk assessment results of the power system operation process. Since a multi-layer neural network can approximate a nonlinear continuous function with arbitrary precision, this embodiment preferably uses a BP neural network with one hidden layer as the inference part of the fuzzy neural network. It should be noted that the model architecture and inference process of FNO and FNN in this embodiment can refer to the conventional structure of these two models, and will not be repeated here.

[0041] This embodiment addresses the issue that FNN still has insufficient adaptability to global physical laws and resolution changes. It introduces a neural operator for pre-physical guidance into FNN, which learns global dependencies and dominant low-frequency behaviors through low-rank convolution in the frequency domain to obtain physical prior features. This enables FNN to significantly enhance its perception and generalization ability to out-of-distribution extreme scenes while maintaining the interpretability of rules.

[0042] In a preferred embodiment, to further improve the accuracy of risk assessment results, the present invention uses a risk precursor field to adaptively adjust the center and width of the Gaussian membership function of the second layer of the FNN. The Gaussian membership function with adaptive adjustment mechanism can be expressed as: In the formula, This represents the m-th component of the input data. The membership degree of the nth membership function, where exp(*) represents the natural exponential function. and These are the initial parameters for the nth membership function, namely the original center value and the original width value. and It is the learnable weight matrix of the nth membership function. It is the physical prior vector. By adjusting the center and width of the membership function through the weight matrix and the physical prior vector, the fuzzy neural network can make adaptive adjustments according to different risk scenarios, thereby improving the accuracy of the FNN's evaluation results.

[0043] The risk assessment model in this embodiment can be trained using genetic algorithms, fuzzy annealing algorithms, or particle swarm optimization algorithms. However, these algorithms are prone to local optima and slow convergence in complex high-dimensional optimization problems. Therefore, in a preferred embodiment, this invention provides a hybrid whale optimization algorithm based on quantum behavior and elite reverse learning. The specific training steps include: Population encoding and initialization are performed on the model parameters of the fuzzy neural network, and initial population positions are randomly generated; According to the preset search strategy, the whale optimization algorithm is used to optimize the position of the initial population to obtain the global optimal position; Based on the global optimal position, the iterative stagnation method is used to determine whether a local optimum has been reached; If a local optimum is reached, the global optimum position is taken as the current elite solution, and an elite reverse learning strategy is used to generate an elite reverse solution. The fitness value of the elite reverse solution is compared with the fitness value of the current elite solution, and the global optimal position is updated according to the comparison result. Determine whether the iteration stopping condition has been met. If the iteration stopping condition has been met, then the global optimal position is used as the optimal model parameter of the fuzzy neural network. If the iteration stopping condition is not met, the whale optimization algorithm is used to continue the position search until the iteration stopping condition is met.

[0044] In this embodiment, the whale optimization algorithm is first initialized by determining algorithm parameters such as population size, maximum number of iterations, spatial dimension, and spiral shape constant. An initial population is randomly generated within the feasible solution range, and the initial population position, i.e., the position of each individual whale, is randomly generated; this position is a model parameter to be optimized. Then, the fitness value of each individual whale's position in the population is calculated based on the fitness function. The individual with the best fitness is selected from the initial population, and its position and fitness value are recorded as the globally optimal position. and global optimal fitness At the same time, each whale's individual historical best position is initialized to its current position.

[0045] Preferably, the present invention uses root mean square error as the fitness function, and in order to prevent overfitting, a regularization constraint term is added. Therefore, the fitness function can be expressed as: Where X represents the parameter vector, i.e., the individual position. f(X) The fitness of the parameter vector X is represented. This represents the true label value of the e-th sample in the validation set, where E represents the total number of samples. This represents the predicted risk value of the model for the e-th sample when using the parameter vector X. Let X represent the regularization coefficient, and G represent the total dimension of the parameters to be optimized, i.e., the length of the parameter vector X. w g This represents the g-th parameter to be optimized.

[0046] As can be seen, the first term of the fitness function is the root mean square error, which is used to evaluate the model's prediction accuracy; the smaller the value, the higher the model's accuracy. The second term is the regularization constraint, which is a constraint on the sum of squared parameters. It can suppress excessively large parameter values, effectively prevent overfitting, and enhance the model's generalization ability. Of course, other functions can also be used as the fitness function, such as the root mean square error function, etc. Here, only the preferred option is given, not a specific limitation.

[0047] After initialization, the algorithm enters the main loop, which determines the update strategy for the current s-th whale's position based on the norm of the coefficient vector A. The coefficient vector A can be represented as: In the formula, It is a random vector whose elements are uniformly distributed in the range [0,1]. It is a control parameter that varies with the number of iterations. Decreasing linearly from 2 to 0, It represents the maximum number of iterations.

[0048] As can be seen, the coefficient vector The range of variation is The control mechanism, specifically the size of its L1 norm, determines whether the algorithm performs local mining or global exploration. When the conditions are met, a local search strategy is used; otherwise, a global search strategy is used. The local search strategy is further divided into a shrinking encirclement strategy and a bubble-net predation strategy, determined by comparing the strategy selection probability with a probability threshold.

[0049] Specifically, when When a local search strategy is adopted, and when the strategy selection probability is... At that time, a shrinking encirclement strategy is used for position updates: In the formula, D is the individual position of the s-th whale in the t-th iteration. and the current optimal solution The distance vector between them, where C represents a random vector and A represents a coefficient vector. It represents the position of the s-th whale in the (t+1)th iteration.

[0050] when and At this time, a bubble-net hunting strategy is employed, which simulates the behavior of humpback whales expelling spiral-shaped bubble nets to drive away prey. This is also a fine-grained localized hunting strategy, and the behavior is modeled using a logarithmic spiral function. The whale's position update formula is: in: It is the individual position of the s-th whale in the t-th iteration. Compared with the current optimal solution The distance vector between them. b is a constant used to define the shape of the logarithmic spiral, usually with a value of 1. It is a random number that is uniformly distributed in the interval [-1, 1].

[0051] As can be seen, in the local search strategy, whales will choose the shrinking encirclement strategy with a 50% probability and the bubble net hunting strategy with a 50% probability.

[0052] when When the algorithm reaches a point where it needs to explore a large area, quantum behavior is activated to escape the current region. Since the standard whale optimization algorithm relies only on the current optimal solution, information exchange is insufficient, and the whale optimization algorithm itself lacks a mechanism to escape local optima, making it prone to premature convergence. Therefore, this embodiment introduces the concept of average optimal position through quantum behavior. This represents the collective wisdom and experience of the entire population, and the average optimal position can be expressed as: In the formula, This represents the average optimal position at the t-th iteration. Let S represent the historical best position of the s-th whale in the first t iterations, and let S represent the total number of whales.

[0053] The individual whale positions are then updated according to the following equation: In the formula, denoted as the contraction / expansion coefficient, and u represents a random number.

[0054] This embodiment replaces the inefficient random prey-hunting stage of the standard whale optimization algorithm with a global search based on quantum behavior. By introducing quantum behavior, the whale's motion becomes uncertain; its position is no longer a fixed point, but rather appears with a certain probability within a quantum attractive potential well centered on an "average optimal position." This means that each whale can potentially appear anywhere in the search space, but has a higher probability of appearing near historical optimal solutions. This approach significantly enhances the algorithm's ability to escape local optima. It is the contraction-expansion coefficient, a hyperparameter used to control the scale of the search step. u It is a random number within the interval (0,1). The ln(1 / u) term is key, when... u When this value approaches 0, it produces a very large value, enabling the whale to perform quantum leaps, greatly enhancing its global exploration capabilities and its ability to escape local optima. ± indicates that the addition or subtraction operation is randomly selected with a 50% probability, increasing the diversity of the search.

[0055] After performing boundary checks and corrections on the updated individual whale positions, the individual position, the individual's historical best position, and the global best position are updated by calculating the fitness value of the individual position. Then, the iterative operation continues until the preset iteration stopping condition is reached, such as reaching the preset number of iterations, so as to obtain the final global best position, which is also the optimal parameters of the model.

[0056] Furthermore, to accelerate parameter optimization convergence and address the slow convergence speed and evolutionary stagnation problem inherent in the standard whale optimization algorithm, this invention introduces an elite back-learning mechanism into the whale optimization algorithm. This back-learning addresses evolutionary stagnation and accelerates convergence. Specifically, this embodiment adds a counter at the end of each iteration. This counter counts the number of times the global optimum (i.e., the optimal solution) remains unchanged during the iteration process. If the optimal solution does not improve within a pre-set number of consecutive iterations, the algorithm may be trapped in a local optimum. In this case, the elite back-learning strategy is activated. During elite back-learning, the dynamic boundary is first calculated based on the extreme values ​​of the current population in each dimension to provide a search range for generating back-learning solutions. Then, for the global optimum, elite back-learning solutions are generated according to the dimensions. In the formula, This represents the parameter vector X at the t-th iteration. g Elite reverse solution in each dimension Let g represent a random number that is uniformly distributed in the interval [0,1] along dimension g. and They represent the t-th iteration and the t-th iteration respectively. gThe lower and upper bounds of each dimension. Represents the position of the globally optimal position. g Values ​​in each dimension.

[0057] Elite inverse solutions in various dimensions constitute the elite inverse solution. Then, the fitness value of the elite reverse solution is calculated and compared with the fitness value of the current global optimum. If the fitness value of the elite reverse solution is higher than the fitness value of the global optimum in the current iteration, the elite reverse solution is used as the optimal solution for the next iteration and the calculator is reset. Otherwise, the optimal solution of the previous iteration is still used as the optimal solution for the next iteration.

[0058] As can be seen, if the reverse solution is better, it points to a better direction for the next round of evolution of the population, thus effectively breaking the deadlock of evolutionary stagnation. The main loop terminates when the number of iterations reaches the maximum stopping value. The algorithm outputs the final global optimal position; this vector represents the optimal parameter configuration of the fuzzy neural network.

[0059] The improved whale optimization algorithm provided in this embodiment inherits the excellent local mining capabilities of the standard whale optimization algorithm. By introducing a stagnation breakthrough mechanism of quantum behavior global exploration and elite reverse learning, it constructs a more powerful optimization framework, effectively avoiding premature model convergence and accelerating the convergence speed of parameter optimization. This further improves the evaluation accuracy and timeliness of the risk assessment model under complex coupling conditions.

[0060] For a well-trained risk assessment model, inputting the system state feature vector obtained through noise filtering and data fusion into the model can predict the risk level of the power system under its current operating state. For example, the system's operating risk can be pre-classified into five levels: high risk, relatively high risk, medium risk, relatively low risk, and low risk. The corresponding risk level can then be directly obtained through the model. The risk level predicted by the model can serve as the basic data for power system operation control and maintenance, enabling adjustments to grid control strategies to ensure the stable operation of the grid under extreme weather conditions.

[0061] This embodiment provides a power system risk assessment method that achieves power system operation risk assessment under extreme weather scenarios through an integrated link of multi-source information fusion, physical prior injection, fuzzy inference, global optimization, and risk output. This embodiment embeds variational information bottlenecks in the joint filtering of information forms for denoising, improving the robustness of data fusion. By embedding physical guidance in the risk assessment model, the perception capability of extreme scenarios and the generalization capability of the model are enhanced. Furthermore, the improved whale optimization algorithm is used to optimize model parameters, improving the model's assessment accuracy. This embodiment realizes real-time monitoring and robust assessment of the power grid operation status, further ensuring the safety and stability of power system operation.

[0062] Please see Figure 2 Based on the same inventive concept, the second embodiment of the present invention proposes a power system risk assessment device, comprising: Data fusion module 10 is used to acquire multi-source operation-related data of the power system, and adopts joint Kalman filtering based on variational information bottleneck mechanism to perform noise filtering and data fusion on the multi-source operation-related data to obtain system state feature vector; The risk assessment module 20 is used to input the system state feature vector into a pre-constructed risk assessment model to obtain a risk assessment result. The risk assessment model is constructed based on Fourier neural operators and fuzzy neural networks. The Fourier neural operators are used to extract the physical prior features of the system state feature vector. The fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector. The fuzzy neural network is trained using an improved whale optimization algorithm.

[0063] The technical features and effects of the power system risk assessment device proposed in this embodiment are the same as those of the method proposed in this embodiment, and will not be repeated here. Each module in the above-mentioned power system risk assessment device can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0064] Furthermore, embodiments of the present invention also propose a computer device, including 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-described method.

[0065] Please see Figure 3The diagram illustrates the internal structure of a computer device in one embodiment. This computer device can specifically be a terminal or a server. The computer device includes a processor, memory, network interface, display, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a power system risk assessment method. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input devices of the computer device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse, etc.

[0066] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0067] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0068] In summary, the present invention provides a power system risk assessment method, apparatus, device, and storage medium. The method acquires multi-source operational data of the power system, employs a joint Kalman filter based on a variational information bottleneck mechanism to filter noise and fuse the multi-source operational data, obtaining a system state feature vector. The system state feature vector is then input into a pre-constructed risk assessment model to obtain a risk assessment result. The risk assessment model is constructed based on Fourier neural operators and a fuzzy neural network. The Fourier neural operators are used to extract the physical prior features of the system state feature vector, and the fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector. Furthermore, the fuzzy neural network is trained using an improved whale optimization algorithm. This invention embeds variational information bottlenecks in information-form joint filtering for denoising, thereby improving the robustness of data fusion. By embedding physical guidance in the risk assessment model, it enhances the ability to perceive extreme scenarios and the model's generalization ability. Furthermore, by using an improved whale optimization algorithm to optimize model parameters, it improves the model's evaluation accuracy. This invention enables real-time monitoring and robust evaluation of the power grid's operating status, further ensuring the safety and stability of the power system.

[0069] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0070] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for risk assessment of power systems, characterized in that, include: Multi-source operation-related data of the power system are acquired, and joint Kalman filtering based on variational information bottleneck mechanism is used to filter noise and fuse data to obtain system state feature vector; The system state feature vector is input into a pre-constructed risk assessment model to obtain a risk assessment result. The risk assessment model is constructed based on Fourier neural operators and fuzzy neural networks. The Fourier neural operators are used to extract the physical prior features of the system state feature vector. The fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector. The fuzzy neural network is trained using an improved whale optimization algorithm.

2. The power system risk assessment method according to claim 1, characterized in that, The steps of using joint Kalman filtering based on a variational information bottleneck mechanism to filter noise and fuse data from the multi-source operational data to obtain the system state feature vector include: Multiple sub-filters are used to process the multi-source operational data collected by their respective sensors; The state vector output from each sub-filter is input into a pre-trained variational information bottleneck encoder to obtain Gaussian pseudo-measurement values ​​and Gaussian pseudo-measurement noise matrix. Based on the Gaussian pseudo-measurement value and the Gaussian pseudo-measurement noise matrix, the measurement value and measurement noise matrix in each sub-filter are updated, and the state information of each sub-filter is generated based on the updated measurement value and measurement noise matrix. The state information of each sub-filter is fused using the main filter to obtain the system state feature vector.

3. The power system risk assessment method according to claim 1, characterized in that, The risk assessment model consists of cascaded Fourier neural operators and fuzzy neural networks; The Fourier neural operator is used to physically guide the input system state feature vector to obtain a risk precursor field and physical prior features. The risk precursor field is the physical prior vector output by the last layer of the Fourier neural operator, and the physical prior features are obtained by feature extraction from the risk precursor field. The fuzzy neural network is used to fuzzify the input concatenated feature vector to obtain the risk assessment result. The concatenated feature vector is obtained by concatenating the system state feature vector and the physical prior features.

4. The power system risk assessment method according to claim 3, characterized in that, The center and standard deviation of the membership function of the fuzzy neural network are determined by dynamic adjustment based on the risk precursor field.

5. The power system risk assessment method according to claim 1, characterized in that, The training steps of the fuzzy neural network include: Population encoding and initialization are performed on the model parameters of the fuzzy neural network, and initial population positions are randomly generated; According to the preset search strategy, the whale optimization algorithm is used to optimize the position of the initial population to obtain the global optimal position; Based on the global optimal position, the iterative stagnation method is used to determine whether a local optimum has been reached; If a local optimum is reached, the global optimum position is taken as the current elite solution, and an elite reverse learning strategy is used to generate an elite reverse solution. The fitness value of the elite reverse solution is compared with the fitness value of the current elite solution, and the global optimal position is updated according to the comparison result. Determine whether the iteration stopping condition has been met. If the iteration stopping condition has been met, then the global optimal position is used as the optimal model parameter of the fuzzy neural network. If the iteration stopping condition is not met, the whale optimization algorithm is used to continue the position search until the iteration stopping condition is met.

6. The power system risk assessment method according to claim 5, characterized in that, The step of using the whale optimization algorithm to optimize the initial population position according to a preset search strategy to obtain the globally optimal position includes: Determine if the norm of the coefficient vector is less than the norm threshold. If it is, use a local search strategy; otherwise, use a global search strategy. The initial population position is optimized according to the selected search strategy to obtain the global optimal position.

7. The power system risk assessment method according to claim 6, characterized in that, The step of optimizing the initial population position according to the selected search strategy to obtain the globally optimal position includes: In response to the use of a local search strategy for location optimization, it is determined whether the probability of strategy selection is less than the probability threshold. If so, a shrinking encirclement strategy is adopted to update the individual's position; otherwise, a bubble net predation strategy is adopted to update the individual's position. In response to employing a global search strategy for position optimization, an average optimal position is generated through quantum behavior, and the individual position is updated based on the average optimal position; The updated global optimal position is obtained based on the updated fitness value of the individual position.

8. A power system risk assessment device, characterized in that, include: The data fusion module is used to acquire multi-source operation-related data of the power system. It uses a joint Kalman filter based on the variational information bottleneck mechanism to perform noise filtering and data fusion on the multi-source operation-related data to obtain the system state feature vector. The risk assessment module is used to input the system state feature vector into a pre-constructed risk assessment model to obtain a risk assessment result. The risk assessment model is constructed based on Fourier neural operators and fuzzy neural networks. The Fourier neural operators are used to extract the physical prior features of the system state feature vector. The fuzzy neural network is used to perform risk assessment on the physical prior features and the system state feature vector. The fuzzy neural network is trained using an improved whale optimization algorithm.

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 the method according to any one of claims 1 to 7.

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 1 to 7.

Images