Sample generation method, system, device and medium for mountain fire trip prediction model

Abstract

The application discloses a sample generation method, system, device and medium for a mountain fire tripping prediction model, and the method comprises the following steps: acquiring an electric wire sample set when a mountain fire occurs; extracting environmental features of a real tripping sample from the electric wire sample set, and then determining a physical prior condition vector; inputting the generated random noise vector and the physical prior condition vector into a generated adversarial network, outputting an initial synthetic tripping sample through a generator, calculating a physical constraint loss according to the initial synthetic tripping sample, calculating a generator loss value based on the physical constraint loss and an adversarial loss, iteratively training until the generator loss value meets a preset convergence condition, and obtaining a trained generator; generating a target synthetic tripping sample based on the generator, and merging the target synthetic tripping sample with the real tripping sample to construct a training data set, so as to be used for training of a tripping prediction model. Therefore, the application can quickly generate a large number of tripping samples with few real tripping cases.

Description

Technical Field

[0001] This invention relates to the field of sample generation, and more particularly to a sample generation method, system, device, and medium for wildfire trip prediction models. Background Technology

[0002] In the field of power system disaster prevention and mitigation, transmission line tripping caused by wildfires has become a major hidden danger threatening the safe operation of the power grid. With the increasing frequency of extreme weather events, wildfires near transmission line corridors can easily cause ground-to-ground or phase-to-phase discharge tripping, and even regional power outages, through the dual physical mechanisms of increased conductor sag due to high temperatures and reduced air insulation strength due to combustion products. To improve disaster prevention capabilities, the industry has widely introduced data-driven methods such as deep learning to build tripping risk prediction models by integrating multi-dimensional features such as meteorology, terrain, vegetation, and power grid operation. However, in actual operation records, these models have millions of safe samples, while actual tripping events caused by wildfires are extremely rare (the ratio of positive to negative samples often exceeds 1:1000). This extreme imbalance causes traditional machine learning models to develop severe majority class bias during training, that is, they tend to classify all inputs as "no tripping" in exchange for artificially high accuracy, thus losing sensitivity to real danger signals and causing frequent missed detections.

[0003] To address the aforementioned extremely small sample problem, existing technologies mainly rely on oversampling methods based on interpolation (such as SMOTE, Borderline-SMOTE, and ADASYN). These methods synthesize new samples by performing linear interpolation between minority class samples and their nearest neighbors. While simple to implement and friendly to traditional classifiers, their generation mechanism depends solely on the distance relationship in the feature space and completely lacks guidance from physical mechanisms. This can easily lead to the interpolation of physically invalid pseudo-samples between two real dangerous samples. Furthermore, when the number of original minority samples is extremely small and their distribution is sparse, the interpolation direction is easily dominated by noise points, resulting in insufficient engineering reliability of the generated results. Summary of the Invention

[0004] This invention provides a method, system, device, and medium for generating samples for wildfire tripping prediction models, which can quickly generate a large number of tripping samples using very few real tripping cases.

[0005] This invention provides a sample generation method for a wildfire tripping prediction model, comprising: Obtain a sample set of power lines at the time of wildfires; Environmental features of real tripping samples are extracted from the wire sample set, and physical prior condition vectors are determined based on the environmental features. The generated random noise vector is concatenated with the physical prior condition vector to obtain a concatenated vector. The concatenated vector is then input into the constructed generative adversarial network (GAN). The generator outputs an initial synthetic tripping sample. The physical constraint loss is calculated based on the initial synthetic tripping sample. The generator loss value is calculated based on the physical constraint loss and the adversarial loss. The training is iterative until the generator loss value meets a preset convergence condition, resulting in a trained generator. The GAN includes the generator and a discriminator, and the adversarial loss is calculated by the discriminator based on the initial synthetic tripping sample. The generator generates target synthetic tripping samples, and the target synthetic tripping samples are merged with the real tripping samples to construct a training dataset for training the tripping prediction model.

[0006] This invention extracts environmental features from real tripping samples and calculates physical prior condition vectors accordingly. It injects the thermodynamics of the power grid and the physical laws of insulation breakdown into the generator input in explicit carrier form, providing a clear physical baseline for a very small number of real samples. This allows the generator to learn complex multi-physics coupling relationships from scratch under small sample conditions, significantly compressing the search range of the generation space and significantly improving the convergence speed and sample generation efficiency in the early stages of training. Secondly, by concatenating random noise vectors and physical prior condition vectors into the generator, it retains the nonlinear fitting ability of generative adversarial networks to data distribution. Furthermore, the physical baseline constraint transforms the generation task from "blindly fitting limited data" to "performing residual modulation within the physically feasible domain," effectively avoiding the risk of overfitting or mode collapse in purely data-driven models under very small sample conditions. This ensures that each forward propagation generates physically reasonable candidate samples, providing structural guarantees for rapid iterative generation. Finally, after the generator outputs the initial synthesized tripping samples, the physical constraint loss is calculated simultaneously. The generator's total loss is formed by combining the adversarial loss fed back from the discriminator. The physical constraint loss performs real-time physical compliance checks on the generated samples from two dimensions: meteorological thermodynamic consistency and insulation breakdown critical conditions. The adversarial loss ensures that the generated samples approximate real data at the statistical distribution level. The two form a dual traction of "physical correctness" and "data authenticity," guiding the generator to optimize towards both objectives simultaneously in each iteration. This avoids the pseudo-sample problem that is easily generated by relying solely on adversarial training and significantly reduces invalid generation and repeated trial and error caused by physical verification failures. As a result, the preset convergence condition can be reached within a very short training cycle. Finally, the trained generator outputs target synthetic tripping samples in batches and merges them with the original real tripping samples to construct a balanced training dataset. This allows the downstream tripping prediction model to fully learn the disaster-causing feature patterns with the support of sufficient and physically reliable minority class samples. This achieves an order-of-magnitude expansion of the number of tripping samples with extremely low real data annotation costs, significantly improving the early warning model's sensitivity to high-risk tripping events and its engineering practicality.

[0007] Further, determining the physical prior condition vector based on the environmental features includes: The initial distance between the conductor and the ground, the increase in conductor thermal expansion sag, the line operating voltage, and the live wire intensity characteristics are obtained from the environmental characteristics; wherein, the increase in conductor thermal expansion sag is determined based on the ambient temperature and the line load current. Calculate the first difference between the initial distance of the conductor to the ground and the increase in sag due to thermal expansion of the conductor; Calculate the ratio of the first difference to the line operating voltage. Based on the ratio and the live wire strength characteristics, calculate the insulation breakdown margin under the current state, and use the insulation breakdown margin as the physical prior condition vector.

[0008] By using thermodynamic and insulation physics formulas to calculate the insulation breakdown margin as a prior condition, the purely random generation is transformed into "physical baseline + residual modulation", which significantly compresses the generation space, enabling the generator to converge stably even with extremely scarce samples, reducing training difficulty and improving generation efficiency.

[0009] Further, the step of outputting the initial synthesized trip sample through the generator includes: The generator maps the concatenated vector into a query matrix, a key matrix, and a value matrix; Calculate the Manhattan distance between the query matrix and the key matrix; The Manhattan distance is scaled to obtain the scaled result; The scaling result is converted into attention weights using a normalized exponential function; Multiplying the attention weights by the value matrix yields a weighted feature representation; Forward propagation is performed based on the weighted feature representation to generate the initial synthetic trip sample.

[0010] This method of using Manhattan distance instead of dot product to calculate attention weights is naturally robust to dimensional differences and outliers, effectively suppressing the dominance of high-dimensional features (such as voltage) on attention, accurately focusing on low-dimensional disaster-causing factors such as sudden changes in humidity and vegetation drying, and improving the quality and credibility of generated samples.

[0011] Further, the calculation of physical constraint loss based on the initial synthetic tripping sample includes: The generated fire line intensity features and ambient humidity features are extracted from the initial synthetic trip sample. The first product of the fire line intensity features and the ambient humidity is calculated. The first product is subtracted from the preset meteorological consistency threshold to obtain the second difference. The second difference is processed by a linear rectification function to obtain the meteorological thermodynamic consistency loss. The generated trip tag and insulation breakdown margin are extracted from the initial synthetic trip sample. When the trip tag indicates that a trip has occurred and the insulation breakdown margin is greater than the preset safety critical threshold, a second penalty value is generated as the critical penalty loss for insulation breakdown. The meteorological thermodynamic consistency loss and the insulation breakdown critical penalty loss are combined as the physical constraint loss.

[0012] By designing dual physical penalty terms of meteorological consistency and insulation breakdown, the generation process is constrained from both feature combination and causal logic perspectives. This forces the elimination of anti-physical samples such as "high humidity fire" and "safety tripping," ensuring that the generated data is rigorous and reliable in engineering and avoiding misleading subsequent prediction models.

[0013] Further, the calculation of physical constraint loss based on the initial synthetic tripping sample includes: The initial synthetic tripping sample and the real tripping sample are input into the discriminator to output a realism score, and the adversarial loss is calculated based on the realism score. Calculate the second product of the meteorological thermodynamic consistency loss and the first weighting coefficient; Calculate the third product of the critical insulation breakdown penalty loss and the second weighting coefficient; The generator loss value is obtained by summing the adversarial loss, the second product, and the third product.

[0014] This approach combines the adversarial loss and the two physical penalty losses in a weighted sum as the generator's total loss, enabling the generator to simultaneously pursue statistical realism and physical plausibility. The weighting coefficients can be flexibly adjusted to adapt to the physical constraint strength requirements of different scenarios, enhancing the method's versatility.

[0015] Further, the iterative training until the generator loss value satisfies a preset convergence condition to obtain a trained generator includes: The network weights of the discriminator are updated in reverse based on the adversarial loss calculated in each iteration; After updating the network weights of the discriminator, the current weights of the discriminator are fixed, and the network weights of the generator are updated in reverse according to the generator loss value. Repeat the steps of updating the discriminator weights and updating the generator weights until the preset convergence condition is met, and the trained generator is obtained.

[0016] This adversarial training strategy, which alternates between updating the discriminator and generator, combined with the gradient penalty of WGAN-GP, effectively improves training stability and prevents mode collapse. Even under conditions of insufficient positive samples, it can still generate diverse and realistic high-quality tripping samples.

[0017] Furthermore, the acquisition of the power line sample set during a wildfire includes: Obtain an initial power line sample set at the time of a wildfire, wherein the initial power line sample set includes meteorological data, terrain data, wildfire data, and line operation parameters; The initial wire sample set is cleaned to obtain the data cleaning results; The data cleaning results are normalized to obtain the normalized result; The normalization result is then subjected to feature engineering to obtain the processed wire sample set.

[0018] This process of standardizing multi-source heterogeneous raw data through data cleaning, normalization, and feature engineering eliminates noise and dimensional differences, providing clean and standardized input for subsequent physical prior calculations and network training, thereby improving the reliability and robustness of the overall generation process.

[0019] Another embodiment of the present invention provides a sample generation system for a wildfire trip prediction model, comprising: the sample generation system being used to perform the sample generation method for the wildfire trip prediction model as described in the present invention.

[0020] Another embodiment of the present invention provides a terminal device, comprising: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the sample generation method for the wildfire trip prediction model as described in the present invention.

[0021] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the sample generation method of the wildfire trip prediction model as described in the present invention. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an embodiment of the sample generation method for the wildfire tripping prediction model provided in this application; Figure 2 This is a schematic diagram of the structure of the generative adversarial network provided in this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0029] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0030] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0031] See Figure 1To quickly generate a large number of tripping samples using very few real tripping cases, an embodiment of the present invention provides a sample generation method for a wildfire tripping prediction model, including steps S101 to S104: Step S101: Obtain a sample set of power lines at the time of the wildfire; In some embodiments, obtaining the power line sample set at the time of a wildfire includes: obtaining an initial power line sample set at the time of a wildfire, wherein the initial power line sample set includes meteorological data, terrain data, wildfire data, and line operation parameters; performing data cleaning on the initial power line sample set to obtain a data cleaning result; performing normalization processing on the data cleaning result to obtain a normalization processing result; and performing feature engineering on the normalization processing result to obtain a processed power line sample set.

[0032] In some embodiments, historical multidimensional data related to transmission lines are collected across systems, covering four data dimensions: meteorological data (temperature, humidity, wind speed and direction, etc.), terrain data (elevation, slope, etc.), wildfire data (fire line intensity FRP, distance from fire point, etc.), and line operating parameters (operating voltage, load current, etc.). Then, the initial sample set is cleaned to remove missing values ​​and abnormal noise, resulting in cleaned data. Subsequently, the cleaned results are normalized to eliminate the influence of dimensional differences between dimensions (e.g., temperature is only on the order of tens, while voltage can reach tens of thousands of volts), ensuring that features from different sources and scales are within the same order of magnitude. Based on this, feature engineering is performed on the normalized results to extract or construct effective features related to wildfire tripping mechanisms, ultimately constructing a power line sample set X=[ ],in, Represents meteorological characteristic vectors (temperature, humidity, wind, etc.). This represents the vector of wildfire and terrain features (fire line intensity, distance from the conductor to the fire point, etc.). The characteristic vector of the power grid line (operating voltage, initial sag, etc.) is represented by y∈{0,1}, which indicates whether a trip has occurred.

[0033] It should be noted that the collected historical multidimensional data comes from multiple business platforms, including the power grid dispatch system, meteorological monitoring station network, satellite remote sensing fire point monitoring platform, and geographic information system.

[0034] It should be noted that in actual working conditions, power outages caused by wildfires are extremely rare events. The number of normally operating samples in this dataset is far greater than the number of outage samples, resulting in a highly unbalanced data distribution. This provides a data foundation for the subsequent extraction of physical prior conditions and the training of the generation network.

[0035] This process of standardizing multi-source heterogeneous raw data through data cleaning, normalization, and feature engineering eliminates noise and dimensional differences, providing clean and standardized input for subsequent physical prior calculations and network training, thereby improving the reliability and robustness of the overall generation process.

[0036] Step S102: Extract environmental features of real tripped samples from the wire sample set, and determine the physical prior condition vector based on the environmental features; In some embodiments, environmental features of real tripping samples are extracted from the wire sample set. Specifically, firstly, in the constructed wire sample set X=[ Based on this, when extracting environmental features from real tripping samples, the operation method is to first rely on the tripping label. The samples were screened to locate the A very small number of records have a value of 1, representing actual historical wildfire power outages. For these selected records, their category labels are removed. The remaining feature tensors are separated into three components: meteorological feature vectors. Wildfire and terrain feature vectors and power grid line feature vectors This separation process essentially extracts the complete feature space of minority class samples from the highly imbalanced original data, providing the necessary input parameters for subsequent calculations of physical prior conditions. Key physical quantities are derived from the feature vectors obtained in this extraction step.

[0037] It should be noted that, since real tripped samples account for a very small percentage of the dataset (the ratio of positive to negative samples can be as high as 1:1000 or even more), this extraction step only filters out a small number of valid records, thus supporting the subsequent construction of physical priors and training of the generative network based on these scarce real samples.

[0038] In some embodiments, determining the physical prior condition vector based on the environmental characteristics includes: obtaining the initial distance between the conductor and ground, the increase in conductor thermal expansion sag, the line operating voltage, and the live wire strength characteristics from the environmental characteristics; wherein the increase in conductor thermal expansion sag is determined based on the ambient temperature and the line load current; calculating a first difference between the initial distance between the conductor and ground and the increase in conductor thermal expansion sag; calculating the ratio of the first difference to the line operating voltage; calculating the insulation breakdown margin under the current state based on the ratio and the live wire strength characteristics; and using the insulation breakdown margin as the physical prior condition vector.

[0039] In some embodiments, after extracting the environmental features of real tripping samples, the construction of the physical prior condition vector requires selecting four key physical parameters, namely the initial conductor-to-ground / tree distance. Wire strength FRP, operating voltage and the increase in conductor sag due to thermal expansion After obtaining the above four parameters, first calculate the initial distance between the conductor and the ground. Increase in sag due to thermal expansion of the conductor The first difference between This difference represents the actual clearance distance of the conductor under the current thermal state; Then divide the first difference by the line operating voltage. Get the ratio This is used to evaluate the insulation gap level under unit voltage; Finally, subtract the fire wire strength FRP and free coefficient from this ratio. product That is, according to the formula The insulation breakdown margin IBM under the current state is calculated, and the insulation breakdown margin IBM is output as the physical prior condition vector c.

[0040] It should be noted that the increase in conductor sag due to thermal expansion It is not a static value read directly from a historical database, but rather needs to be adjusted based on the current ambient temperature. With line load current The physical meaning of dynamic calculation is that the high temperature environment generated by wildfires will cause the conductor to expand due to heat, which in turn will increase the sag and reduce the air gap between the conductor and the ground or trees.

[0041] It should be noted that the insulation breakdown margin IBM, as the output of the physical prior condition vector c, quantifies the line's ability to withstand breakdown under the current environmental and power grid conditions—the smaller the value, the more insufficient the insulation margin and the higher the risk of tripping.

[0042] By using thermodynamic and insulation physics formulas to calculate the insulation breakdown margin as a prior condition, the purely random generation is transformed into "physical baseline + residual modulation", which significantly compresses the generation space, enabling the generator to converge stably even with extremely scarce samples, reducing training difficulty and improving generation efficiency.

[0043] Step S103: The generated random noise vector is concatenated with the physical prior condition vector to obtain a concatenated vector. The concatenated vector is input into the constructed generative adversarial network (GAN). The generator outputs an initial synthetic tripping sample. The physical constraint loss is calculated based on the initial synthetic tripping sample. The generator loss value is calculated based on the physical constraint loss and the adversarial loss. The training is iterative until the generator loss value meets the preset convergence condition to obtain the trained generator. The GAN includes the generator and the discriminator. The adversarial loss is calculated by the discriminator based on the initial synthetic tripping sample. In some embodiments, the generated random noise vector is concatenated with the physical prior condition vector to obtain a concatenated vector. Specifically, at the start of each training iteration, the generator first samples a pure random noise vector from a standard normal distribution or a uniform distribution. This vector serves as a random seed in the generation process, introducing diversity in the data distribution. Then, this random noise vector z is horizontally concatenated with the calculated physical prior condition vector c along the feature dimension to obtain the concatenated vector. =[z,c].

[0044] This splicing operation is not a simple numerical stacking, but rather uses physical priors as the deterministic input basis of the generator and random noise as an uncertain modulation signal, together forming the complete input of the generator.

[0045] Please refer to Figure 2 To rapidly generate samples, this invention proposes a Physics-Guided GAN with Manhattan Attention (PGAN-MA), specifically designed to address the issues of extremely small sample generation and data imbalance caused by wildfire-induced power grid tripping. In the overall architecture of PGAN-MA, extremely imbalanced historical wildfire and power grid data are collected as input to the discriminator D. Unlike traditional GANs that rely on pure random noise, this invention's generator G incorporates both physics-based conditions and random noise as input. During network optimization, the discriminator updates its weights using only a data-driven adversarial loss, while the generator's optimization relies on a composite loss function. This function not only includes adversarial loss but also innovatively introduces multiphysics consistency loss (Physics Loss). After training, the generator can synthesize realistic minority class (tripping) samples that conform to the common sense of power grid physics, thereby transforming the unbalanced dataset into a balanced dataset for use in subsequent wildfire tripping models.

[0046] In some embodiments, the step of outputting the initial synthetic trip sample through the generator includes: the generator mapping the concatenated vector to a query matrix, a key matrix, and a value matrix; calculating the Manhattan distance between the query matrix and the key matrix; scaling the Manhattan distance to obtain a scaling result; converting the scaling result into attention weights using a normalized exponential function; multiplying the attention weights by the value matrix to obtain a weighted feature representation; and performing forward propagation based on the weighted feature representation to generate the initial synthetic trip sample.

[0047] In some embodiments, the spliced ​​vector The input is a constructed generative adversarial network (GAN), and the generator G in the GAN receives the concatenated vector. Then, the input features are first mapped to the query matrix Q, the key matrix K, and the value matrix through a linear transformation layer. All three belong to , among which, among which The length of the feature sequence. The dimension is the feature vector. Then, the Manhattan distance between the query matrix Q and the key matrix K is calculated. (i.e., L1 norm).

[0048] Next, the calculated Manhattan distance is negativeized and then scaled by a scaling factor. Then take the negative value to get the scaling result. Then, the scaling result is converted into an attention weight matrix through the normalized exponential function (Softmax). Its physical meaning is: the smaller the Manhattan distance (the more similar the features), the higher the attention weight is assigned between feature pairs.

[0049] Next, the attention weights and value matrix are... Multiplication yields a weighted feature representation. Finally, the generator uses this weighted feature representation to perform subsequent forward propagation (including several fully connected layers and activation functions), progressively decoding and ultimately outputting the initial synthesized tripped sample. Throughout the computation, the mathematical expression for the attention mechanism is: .

[0050] It should be noted that weighted feature representation integrates the correlation between global features in the feature dimension. At the same time, since the L1 norm is naturally robust to extreme outliers, it can effectively filter the interference of differences in dimensions.

[0051] This method of using Manhattan distance instead of dot product to calculate attention weights is naturally robust to dimensional differences and outliers, effectively suppressing the dominance of high-dimensional features (such as voltage) on attention, accurately focusing on low-dimensional disaster-causing factors such as sudden changes in humidity and vegetation drying, and improving the quality and credibility of generated samples.

[0052] In some embodiments, calculating the physical constraint loss based on the initial synthetic trip sample includes: extracting generated live wire intensity features and ambient humidity features from the initial synthetic trip sample; calculating a first product of the live wire intensity features and the ambient humidity; subtracting a preset meteorological consistency threshold from the first product to obtain a second difference; processing the second difference using a linear rectification function to obtain the meteorological thermodynamic consistency loss; extracting generated trip tags and insulation breakdown margins from the initial synthetic trip sample; generating a second penalty value as the insulation breakdown critical penalty loss when the trip tag indicates a trip and the insulation breakdown margin is greater than a preset safety critical threshold; and using the meteorological thermodynamic consistency loss and the insulation breakdown critical penalty loss together as the physical constraint loss.

[0053] In some embodiments, after the initial synthetic trip sample is generated, two independent physical consistency checks need to be performed simultaneously. The first check targets meteorological thermodynamic logic: extracting the generated fire line intensity features from the initial synthetic trip sample. With environmental humidity characteristics Calculate the first product of the two. The physical meaning of this product is that intense wildfires (high FRP value) are usually not accompanied by extremely high relative humidity, and high humidity environment has a significant inhibitory effect on flame combustion.

[0054] Then, subtract the preset meteorological consistency threshold from the first product. (This threshold is determined based on reasonable physical experience values ​​from historical statistics), thus obtaining the second difference. The difference is then processed using a linear rectified unit (ReLU), meaning a positive loss is only generated when the product exceeds a threshold, otherwise the loss is zero, thus yielding the meteorological thermodynamic consistency loss. .

[0055] The second verification step addresses the causal logic of insulation breakdown: extracting the generated trip tag from the initial synthetic trip sample. and insulation breakdown margin When the trip label indicates that a trip has occurred (i.e., → 1) and insulation breakdown margin Greater than the preset safety threshold When this occurs, it means the sample is marked as tripped, but the physical status shows the line is still within the safe insulation margin range. This clearly violates the electrical and physical causal logic that "tripping must be accompanied by insufficient insulation margin." In this case, a second penalty value is generated as the critical penalty loss for insulation breakdown; conversely, if no tripping or If the value is less than or equal to the critical threshold, the penalty is 0. The mathematical expression for the critical penalty loss for insulation breakdown is: Finally, the meteorological thermodynamic consistency loss and the insulation breakdown critical penalty loss are combined as physical constraint losses.

[0056] It should be noted that the meteorological consistency threshold This application does not impose any restrictions based on reasonable physical experience values ​​determined by historical statistics.

[0057] By designing dual physical penalty terms of meteorological consistency and insulation breakdown, the generation process is constrained from both feature combination and causal logic perspectives. This forces the elimination of anti-physical samples such as "high humidity fire" and "safety tripping," ensuring that the generated data is rigorous and reliable in engineering and avoiding misleading subsequent prediction models.

[0058] In some embodiments, calculating the generator loss value based on the physical constraint loss and the adversarial loss includes: inputting the initial synthetic trip sample and the real trip sample into the discriminator to output a fidelity score, and calculating the adversarial loss based on the fidelity score; calculating the second product of the meteorological thermodynamic consistency loss and the first weighting coefficient; calculating the third product of the insulation breakdown critical penalty loss and the second weighting coefficient; and summing the adversarial loss, the second product, and the third product to obtain the generator loss value.

[0059] In some embodiments, during each training iteration, the initial synthetic tripping samples output by the generator and the real tripping samples extracted from the original dataset are jointly input into the discriminator D. The discriminator outputs a ground truth score (i.e., a Wasserstein distance estimate) for each input sample. This score reflects the discriminator's confidence that the input sample originates from the real data distribution. Based on the difference between the discriminator's scores for synthetic and real samples, the adversarial loss is calculated, and its mathematical form is: The physical meaning of this is to encourage the generator to synthesize samples that can "deceive" the discriminator and make it give a high score.

[0060] At the same time, the previously calculated meteorological thermodynamic consistency loss will be... Multiplied by the first weighting coefficient The second product is obtained. The critical penalty loss for insulation breakdown. Multiply by the second weighting factor The third product is obtained. .in and All of these are hyperparameters greater than zero, used to adjust the relative strength of the two types of physical constraints in the generator's total loss. The specific values ​​can be determined through cross-validation or empirical tuning based on the actual application scenario.

[0061] Finally, the adversarial loss, the second product, and the third product are summed to obtain the overall loss function of the generator. The total loss value is used as the basis for gradient backpropagation, and the network weights of the generator are updated through the backpropagation algorithm.

[0062] This composite loss structure ensures that the generator is simultaneously constrained by the pressure of fitting the data distribution and the pressure of complying with physical laws during the training process. The generated samples not only approximate the real tripping data in terms of statistical characteristics, but also satisfy the causal logic of thermodynamics and insulation breakdown in terms of physical mechanism.

[0063] This approach combines the adversarial loss and the two physical penalty losses in a weighted sum as the generator's total loss, enabling the generator to simultaneously pursue statistical realism and physical plausibility. The weighting coefficients can be flexibly adjusted to adapt to the physical constraint strength requirements of different scenarios, enhancing the method's versatility.

[0064] In some embodiments, the iterative training until the generator loss value satisfies a preset convergence condition to obtain a trained generator includes: updating the network weights of the discriminator in reverse according to the adversarial loss calculated in each iteration; after updating the network weights of the discriminator, fixing the current weights of the discriminator, and updating the network weights of the generator in reverse according to the generator loss value; repeating the above steps of updating the discriminator weights and updating the generator weights until the preset convergence condition is met to obtain a trained generator.

[0065] In some embodiments, during each training iteration, the initial synthetic tripping samples output by the generator and the real tripping samples are first input into the discriminator. The discriminator outputs a realism score for the samples and calculates the adversarial loss. The objective function of the discriminator is then calculated based on this adversarial loss. The third term is a gradient penalty term used to satisfy the Lipschitz continuity constraint of Wasserstein GAN.

[0066] Subsequently, gradient backpropagation is performed based on the objective function, updating the network weights of the discriminator DD separately, while the generator weights remain completely unchanged during this process. After updating the discriminator weights, the current weight parameters of the discriminator are fixed so that they do not change in subsequent steps; then, the overall loss value of the generator is used as the basis for the update. Gradient backpropagation is performed to update the network weights of the generator G. In this step, the fixed weights of the discriminator provide the generator with a constant evaluation criterion for the realism of the samples, and the generator adjusts its parameters based on this evaluation criterion and the joint feedback from the two physical losses.

[0067] After completing the two update steps described above, training for the current epoch ends. Then, it is checked whether the training has reached the preset maximum number of iterations, or whether the generator loss and discriminator loss have reached stable convergence. If the convergence condition is not met, the process returns to resampling a new random noise vector z and continues with the next round of discriminator and generator weight updates. This process is repeated, and the discriminator's discrimination ability and the generator's synthesis ability gradually improve through alternating competition. When the preset convergence condition is reached, iterative training stops, and the resulting generator possesses the ability to synthesize tripping samples that conform to the physical principles of the power grid under extremely small sample conditions.

[0068] This adversarial training strategy, which alternates between updating the discriminator and generator, combined with the gradient penalty of WGAN-GP, effectively improves training stability and prevents mode collapse. Even under conditions of insufficient positive samples, it can still generate diverse and realistic high-quality tripping samples.

[0069] Step S104: Generate target synthetic tripping samples based on the generator, and merge the target synthetic tripping samples with the real tripping samples to construct a training dataset for training the tripping prediction model.

[0070] In some embodiments, after training, the weight parameters of the generator G are frozen. At this point, the generator has the ability to synthesize tripping samples that conform to the physical common sense of the power grid under extremely small sample conditions. The generator with frozen weights is then fed in batches of random noise vectors z and corresponding physical prior condition vectors c for testing. Based on these inputs, the generator generates target synthetic tripping samples in batches. Subsequently, all the newly generated target synthetic tripping samples are merged with the real tripping samples extracted from the original dataset. After the above merging operation, the ratio of positive to negative samples in the newly constructed training dataset is adjusted to 1:1, achieving absolute balance. Each synthetic sample in this balanced dataset has rigorous physical meaning and can be directly output to downstream modules for training wildfire tripping prediction classification models or early warning models.

[0071] It should be noted that each of these target synthetic trip samples has undergone mandatory verification of physical constraint loss during the training phase, and satisfies rigorous physical meaning at both the meteorological thermodynamic logic and insulation breakdown causal logic levels. There are no anti-physical contradictions such as "high humidity accompanied by a large fire" or "tripping under a safe state".

[0072] This invention extracts environmental features from real tripping samples and calculates physical prior condition vectors accordingly. It injects the thermodynamics of the power grid and the physical laws of insulation breakdown into the generator input in explicit carrier form, providing a clear physical baseline for a very small number of real samples. This allows the generator to learn complex multi-physics coupling relationships from scratch under small sample conditions, significantly compressing the search range of the generation space and significantly improving the convergence speed and sample generation efficiency in the early stages of training. Secondly, by concatenating random noise vectors and physical prior condition vectors into the generator, it retains the nonlinear fitting ability of generative adversarial networks to data distribution. Furthermore, the physical baseline constraint transforms the generation task from "blindly fitting limited data" to "performing residual modulation within the physically feasible domain," effectively avoiding the risk of overfitting or mode collapse in purely data-driven models under very small sample conditions. This ensures that each forward propagation generates physically reasonable candidate samples, providing structural guarantees for rapid iterative generation. Finally, after the generator outputs the initial synthesized tripping samples, the physical constraint loss is calculated simultaneously. The generator's total loss is formed by combining the adversarial loss fed back from the discriminator. The physical constraint loss performs real-time physical compliance checks on the generated samples from two dimensions: meteorological thermodynamic consistency and insulation breakdown critical conditions. The adversarial loss ensures that the generated samples approximate real data at the statistical distribution level. The two form a dual traction of "physical correctness" and "data authenticity," guiding the generator to optimize towards both objectives simultaneously in each iteration. This avoids the pseudo-sample problem that is easily generated by relying solely on adversarial training and significantly reduces invalid generation and repeated trial and error caused by physical verification failures. As a result, the preset convergence condition can be reached within a very short training cycle. Finally, the trained generator outputs target synthetic tripping samples in batches and merges them with the original real tripping samples to construct a balanced training dataset. This allows the downstream tripping prediction model to fully learn the disaster-causing feature patterns with the support of sufficient and physically reliable minority class samples. This achieves an order-of-magnitude expansion of the number of tripping samples with extremely low real data annotation costs, significantly improving the early warning model's sensitivity to high-risk tripping events and its engineering practicality.

[0073] Compared with existing technologies, this invention can generate high-quality minority class samples under conditions where real samples of wildfire tripping are extremely scarce, effectively alleviating the problem of severe imbalance between positive and negative samples. By constructing a generative network for extremely small sample scenarios, the original extremely imbalanced dataset is expanded into a more balanced training dataset, thereby significantly improving the downstream tripping prediction model's ability to identify minority class dangerous events, reducing the false negative rate, and increasing early warning sensitivity. It is more suitable for the practical needs of "better to report false alarms than to miss alarms" in power grid disaster early warning scenarios.

[0074] Compared with existing technologies, this invention introduces physical prior conditions and physical consistency constraints during the generation process, enabling the generated samples to possess statistical authenticity. This effectively avoids the anti-physical pseudo-sample problems that easily occur in existing technologies, such as high humidity and heavy rainfall accompanied by severe fires, or lines with sufficient safety margins being marked as tripped. This improves the engineering credibility and practical usability of the generated data, making the generated samples more suitable for training real-world power grid business models.

[0075] Compared with existing technologies, this invention addresses the significant differences in the dimensionality of multi-source heterogeneous features from meteorological, geographical, vegetation, and power grid operation sources by designing a feature-level Manhattan attention mechanism. This mechanism uses a robust distance metric for feature weight allocation, effectively reducing the dominant role of high-dimensional variables in network training. It enhances the identification of key disaster-causing factors such as humidity changes, sudden wind direction shifts, and decreased vegetation moisture content, while avoiding the attention bias problem inherent in traditional attention mechanisms. This improves the quality of generated samples and the accuracy and stability of subsequent tripping prediction models.

[0076] Compared with existing technologies, this invention is the first to unify and integrate minimal sample generation, physical mechanism constraints, and heterogeneous feature attention modeling, forming an integrated technical solution for wildfire power outage scenarios. This solution simultaneously improves the number of samples, ensures physical rationality, and enhances key feature extraction capabilities, thereby comprehensively improving the generalization performance, engineering adaptability, and deployment value of wildfire power outage prediction systems, and possessing stronger integrated application advantages.

[0077] Based on the above method embodiments, corresponding apparatus embodiments are provided; An embodiment of the present invention provides a sample generation system for a wildfire trip prediction model, comprising: the sample generation system being used to perform the sample generation method for the wildfire trip prediction model as described in the present invention.

[0078] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can implement the sample generation method of the wildfire trip prediction model provided by any of the above-described method embodiments of the present invention.

[0079] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0080] Based on the above embodiments of the sample generation method for the wildfire tripping prediction model, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the sample generation method for the wildfire tripping prediction model of any embodiment of the present invention.

[0081] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0082] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0083] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0084] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium, including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the sample generation method of the wildfire tripping prediction model described in any of the above-described method embodiments of the present invention.

[0085] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0086] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for generating samples for a wildfire tripping prediction model, characterized in that, include: Obtain a sample set of power lines at the time of wildfires; Environmental features of real tripping samples are extracted from the wire sample set, and physical prior condition vectors are determined based on the environmental features. The generated random noise vector is concatenated with the physical prior condition vector to obtain a concatenated vector. The concatenated vector is then input into the constructed generative adversarial network (GAN). The generator outputs an initial synthetic tripping sample. The physical constraint loss is calculated based on the initial synthetic tripping sample. The generator loss value is calculated based on the physical constraint loss and the adversarial loss. The training is iterative until the generator loss value meets a preset convergence condition, resulting in a trained generator. The GAN includes the generator and a discriminator, and the adversarial loss is calculated by the discriminator based on the initial synthetic tripping sample. The generator generates target synthetic tripping samples, and the target synthetic tripping samples are merged with the real tripping samples to construct a training dataset for training the tripping prediction model.

2. The sample generation method for the wildfire tripping prediction model according to claim 1, characterized in that, The determination of the physical prior condition vector based on the environmental features includes: The initial distance between the conductor and the ground, the increase in conductor thermal expansion sag, the line operating voltage, and the live wire intensity characteristics are obtained from the environmental characteristics; wherein, the increase in conductor thermal expansion sag is determined based on the ambient temperature and the line load current. Calculate the first difference between the initial distance of the conductor to the ground and the increase in sag due to thermal expansion of the conductor; Calculate the ratio of the first difference to the line operating voltage. Based on the ratio and the live wire strength characteristics, calculate the insulation breakdown margin under the current state, and use the insulation breakdown margin as the physical prior condition vector.

3. The sample generation method for the wildfire tripping prediction model according to claim 1, characterized in that, The step of outputting the initial synthesized trip sample through the generator includes: The generator maps the concatenated vector into a query matrix, a key matrix, and a value matrix; Calculate the Manhattan distance between the query matrix and the key matrix; The Manhattan distance is scaled to obtain the scaled result; The scaling result is converted into attention weights using a normalized exponential function; Multiplying the attention weights by the value matrix yields a weighted feature representation; Forward propagation is performed based on the weighted feature representation to generate the initial synthetic trip sample.

4. The sample generation method for the wildfire tripping prediction model according to claim 1, characterized in that, The calculation of physical constraint loss based on the initial synthetic tripping sample includes: The generated fire line intensity features and ambient humidity features are extracted from the initial synthetic trip sample. The first product of the fire line intensity features and the ambient humidity is calculated. The first product is subtracted from the preset meteorological consistency threshold to obtain the second difference. The second difference is processed by a linear rectification function to obtain the meteorological thermodynamic consistency loss. The generated trip tag and insulation breakdown margin are extracted from the initial synthetic trip sample. When the trip tag indicates that a trip has occurred and the insulation breakdown margin is greater than the preset safety critical threshold, a second penalty value is generated as the critical penalty loss for insulation breakdown. The meteorological thermodynamic consistency loss and the insulation breakdown critical penalty loss are combined as the physical constraint loss.

5. The sample generation method for the wildfire tripping prediction model according to claim 4, characterized in that, The calculation of physical constraint loss based on the initial synthetic tripping sample includes: The initial synthetic tripping sample and the real tripping sample are input into the discriminator to output a realism score, and the adversarial loss is calculated based on the realism score. Calculate the second product of the meteorological thermodynamic consistency loss and the first weighting coefficient; Calculate the third product of the critical insulation breakdown penalty loss and the second weighting coefficient; The generator loss value is obtained by summing the adversarial loss, the second product, and the third product.

6. The sample generation method for the wildfire tripping prediction model according to claim 1, characterized in that, The iterative training continues until the generator loss value satisfies a preset convergence condition, resulting in a trained generator, including: The network weights of the discriminator are updated in reverse based on the adversarial loss calculated in each iteration; After updating the network weights of the discriminator, the current weights of the discriminator are fixed, and the network weights of the generator are updated in reverse according to the generator loss value. Repeat the steps of updating the discriminator weights and updating the generator weights until the preset convergence condition is met, and the trained generator is obtained.

7. The sample generation method for the wildfire tripping prediction model according to any one of claims 1-6, characterized in that, The acquisition of the power line sample set at the time of the wildfire includes: Obtain an initial power line sample set at the time of a wildfire, wherein the initial power line sample set includes meteorological data, terrain data, wildfire data, and line operation parameters; The initial wire sample set is cleaned to obtain the data cleaning results; The data cleaning results are normalized to obtain the normalized result. The normalization result is then subjected to feature engineering to obtain the processed wire sample set.

8. A sample generation system for a wildfire tripping prediction model, characterized in that, include: The sample generation system is used to execute the sample generation method of the wildfire tripping prediction model as described in any one of claims 1-7.

9. A terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the sample generation method for the wildfire trip prediction model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, the device containing the computer-readable storage medium controls the steps of the sample generation method for the wildfire trip prediction model as described in any one of claims 1-7 to perform the steps.

Images