Power failure prediction system and method

Abstract

The invention relates to the technical field of power failure prediction, and discloses a power failure prediction system and method, which is used for solving the problem that a failure sample is not representative during power failure prediction, and comprises the following steps: obtaining an initial historical sample sequence, judging whether the failure sample needs to be supplemented according to the initial historical sample sequence, and if yes, judging whether the failure sample needs to be supplemented; if it is judged that the fault samples need to be supplemented, generating a synthetic fault sample set, averagely dividing a historical operation cycle into a plurality of historical sub-cycles, obtaining sub-cycle sample enhancement evaluation parameters of each historical sub-cycle, performing evaluation to obtain a sample enhancement judgment index, and judging whether the fault samples need to be supplemented in the historical sub-cycles or not; if it is judged that the fault samples need to be supplemented in the historical sub-periods, the fault samples are supplemented in the historical sub-periods, all the historical sub-periods are traversed, a supplemented historical sample sequence is obtained, an actual power fault prediction model is constructed according to the supplemented historical sample sequence, and the accuracy of power fault prediction is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power fault prediction technology, and more specifically to a power fault prediction system and method. Background Technology

[0002] With the continuous development of intelligent sensing technology and digital operation and maintenance platforms, power systems have deployed various sensing and monitoring devices and data acquisition systems in key areas such as transmission and transformation, distribution networks, and user terminals to achieve real-time monitoring of electrical parameters, recording of fault information, and archiving of historical operating data. Based on the operating data collected by these systems, power companies have gradually built multi-source operating datasets covering indicators such as current, voltage, temperature, frequency, and active power, providing data support for application scenarios such as fault diagnosis, equipment maintenance, and risk warning.

[0003] In intelligent operation and maintenance scenarios for power equipment, data-driven fault prediction models are gradually becoming the mainstream technical approach. These models collect historical operating data from the equipment and combine it with fault label information to construct a training sample set for model learning. The trained prediction model is then applied to current data to output fault probability or warning level, assisting in scheduling decisions and equipment maintenance. Compared to traditional threshold-based discrimination logic, this type of data-driven model has stronger scenario adaptability and recognition sensitivity, and is therefore widely used in transmission main equipment, distribution switchgear, and power electronic terminals.

[0004] However, the above-mentioned technologies have at least the following technical problems: In actual power system operation, faults are low-probability events, and the vast majority of operational data represents normal conditions. Fault samples are extremely scarce, resulting in a significant class imbalance. Traditional model training methods often rely on random sampling or overall training strategies, causing the model to favor recognizing normal states during training. This makes it difficult to effectively learn the characteristics of minority class faults, leading to problems such as high false negative rates and insufficient fault identification capabilities. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a power fault prediction system and method to solve the problems existing in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A power fault prediction system includes: a historical data acquisition module for collecting and storing historical data of power equipment during historical operating cycles, and constructing an initial historical sample sequence; a sample supplementation determination module for determining whether fault samples need to be supplemented based on the initial historical sample sequence; if supplementation is required, a synthetic fault sample set is generated through perturbation and SMOTE, and the synthetic fault sample set is transmitted to the sample supplementation module; if no supplementation is required, the historical sample sequence is transmitted to a supplementary historical sample sequence construction module; and a sample supplementation module for dividing the historical operating cycle into several historical sub-cycles, obtaining sub-cycle sample enhancement evaluation parameters for each sub-cycle, including the number of fault samples, the labels of historical samples, and the output structure of the fault prediction model, and obtaining a sample enhancement determination based on the sub-cycle sample enhancement evaluation parameters. The system comprises four modules: an index for determining whether fault samples need to be supplemented within historical sub-cycles; a supplementary historical sample sequence construction module, which, if determined to require supplementation, supplements fault samples within historical sub-cycles based on the sample enhancement index, iterating through all historical sub-cycles to obtain a supplementary historical sample sequence; a fault prediction model construction module, used to construct an actual power fault prediction model based on the supplementary historical sample sequence; a real-time data acquisition module, used to collect real-time operating data within the current power equipment operating cycle, including electrical operating parameters and environmental parameters consistent with historical data structures; and a fault risk warning and assessment module, used to input real-time operating data into the actual power fault prediction model, output the predicted probability value at the current moment, compare the predicted probability value at the current moment with a set risk threshold, and generate corresponding power equipment fault warning information if the threshold is exceeded.

[0007] Preferably, the step of determining whether to supplement fault samples based on the initial historical sample sequence comprises: obtaining the total number of samples in the initial historical sample sequence, counting the number of fault samples in the initial historical sample sequence, calculating the ratio of the number of fault samples to the total number of samples to obtain the historical fault percentage; dividing the historical sample sequence into several sub-sample sequences, each sub-sample sequence containing a corresponding sample subset, obtaining the number of fault samples in each sub-sample sequence, and calculating the ratio of the number of fault samples to the total number of samples in the sub-sample sequence to obtain the sub-sequence fault percentage; calculating the variance of the sub-sequence fault percentages of all sub-sample sequences to obtain the fault distribution dispersion; calculating the sample enhancement start-up judgment index based on the historical fault percentage and the fault distribution dispersion; and determining whether to supplement fault samples based on the sample enhancement start-up judgment index.

[0008] Preferably, the step of determining whether to supplement faulty samples based on the sample enhancement start-up determination index is as follows: compare the sample enhancement start-up determination index with the start-up threshold; if the sample enhancement start-up determination index is greater than or equal to the start-up threshold, it is determined that faulty samples need to be supplemented; if the sample enhancement start-up determination index is less than the start-up threshold, it is determined that faulty samples do not need to be supplemented.

[0009] Preferably, the steps for obtaining the sample augmentation judgment index are as follows: Obtain the total number of sub-period samples in each historical sub-period; obtain the number of sub-period fault samples in each historical sub-period; calculate the ratio of the number of sub-period fault samples to the total number of sub-period samples to obtain the sub-period fault ratio; obtain the label of each historical sample in the model prediction log within each historical sub-period; evaluate the historical prediction misjudgment impact coefficient based on the label of each historical sample; obtain the output structure of the fault prediction model within each historical sub-period; evaluate the prediction gradient sensitivity impact coefficient based on the output structure of the fault prediction model; normalize the sub-period fault ratio, the historical prediction misjudgment impact coefficient, and the prediction gradient sensitivity impact coefficient; evaluate the sample augmentation judgment index based on the normalized local sample ratio, the historical prediction misjudgment impact coefficient, and the prediction gradient sensitivity impact coefficient. The specific steps are as follows: In the formula, This is represented as the sample augmentation determination index. This represents the percentage of sub-cycle failures after normalization. This is represented by the normalized historical prediction misjudgment impact coefficient. This is expressed as the normalized prediction gradient sensitivity coefficient. These are represented as the weighting coefficients for the proportion of sub-cycle faults after normalization, the weighting coefficients for the impact of historical prediction misjudgments after normalization, and the weighting coefficients for the impact of prediction gradient sensitivity after normalization.

[0010] Preferably, the steps for obtaining the historical prediction misjudgment impact coefficient are as follows: Obtain the model prediction logs for each historical sub-cycle within the historical operating period. The prediction logs include the prediction labels and corresponding actual labels of the historical samples in each historical sub-cycle. The labels include two categories: "fault" and "non-fault". For each historical sub-cycle, iterate through all historical samples within that sub-cycle, count the number of false positives and false negatives, and sum the number of false positives and false negatives to obtain the number of misjudged samples. For each historical sub-cycle, obtain the total number of samples within the historical sub-cycle, and calculate the ratio of the number of misjudged samples to the total number of samples to obtain the historical sub-cycle misjudgment rate. Obtain the prediction probability of all samples within the historical sub-cycle, calculate the standard deviation of the prediction probabilities of all samples to obtain the prediction confidence volatility. Calculate the historical prediction misjudgment impact coefficient based on the historical sub-cycle misjudgment rate and the prediction confidence volatility.

[0011] Preferably, the step of obtaining the predicted gradient sensitivity coefficient is as follows: obtaining the output layer weight matrix and activation function form of the fault prediction model structure deployed in the historical operating cycle; obtaining the historical sample sequence of each historical sub-cycle in the historical operating cycle to form the input feature matrix of the historical sub-cycle, which includes several input feature vectors; for each input feature vector in the historical sub-cycle, calculating the gradient of the output probability with respect to the input feature; arranging all gradients in chronological order to obtain a gradient vector sequence, and sequentially calculating the Euclidean distance between two adjacent gradient vectors to obtain a gradient change rate sequence; and performing an arithmetic mean calculation on the gradient change rate sequence to obtain the predicted gradient sensitivity coefficient.

[0012] Preferably, the step of determining whether faulty samples need to be supplemented in the historical sub-period based on the sample enhancement determination index is as follows: the sample enhancement determination index is compared with the enhancement determination threshold. If the sample enhancement determination index is greater than or equal to the enhancement determination threshold, it is determined that faulty samples need to be supplemented in the historical sub-period; if the sample enhancement determination index is less than the enhancement determination threshold, it is determined that faulty samples do not need to be supplemented in the historical sub-period.

[0013] Preferably, the step of supplementing fault samples in the historical sub-period based on the sample enhancement judgment index is as follows: calculate the ratio of the sample enhancement judgment index to the enhancement judgment threshold, round up to obtain the number of supplementary samples; set a maximum number of supplementary samples; if the number of supplementary samples is less than or equal to the maximum number of supplementary samples, randomly select fault samples corresponding to the number of supplementary samples from the synthetic fault sample set and randomly insert them into the historical samples in the historical sub-period; if the number of supplementary samples is greater than the maximum number of supplementary samples, randomly select fault samples corresponding to the maximum number of supplementary samples from the synthetic fault sample set and randomly insert them into the historical samples in the historical sub-period.

[0014] Preferably, the construction steps of the actual power fault prediction model are as follows: Step 5.1: Divide the historical data of the supplementary historical sample sequence into a training sample set and a validation sample set, and use the training sample set to train the fault prediction model to obtain an initial power fault prediction model; Step 5.2: Obtain the performance index of the initial power fault prediction model based on the validation sample set, and calculate the model qualification index based on the performance index; compare the model qualification index with the qualification threshold. If the model qualification index is greater than or equal to the qualification threshold, the current power fault prediction model is determined to be qualified and can perform fault prediction. The initial power fault prediction model is then recorded as the actual power fault prediction model; Step 5.3: If the model qualification index is less than the qualification threshold, the current power fault prediction model is determined to be unqualified and cannot perform fault prediction. The process returns to Step 5.1 to continue model training until the current power fault prediction model is determined to be able to perform fault prediction. The current power fault prediction model is then recorded as the actual power fault prediction model.

[0015] Preferably, a power fault prediction method includes the following steps: Step 1: Collect and store historical data of power equipment within a historical operating cycle to construct an initial historical sample sequence; Step 2: Determine whether fault samples need to be supplemented based on the initial historical sample sequence. If it is determined that fault samples need to be supplemented, a synthetic fault sample set is generated through perturbation and SMOTE; Step 3: Divide the historical operating cycle into several historical sub-cycles, obtain the sub-cycle sample enhancement evaluation parameters for each historical sub-cycle, evaluate the sample enhancement judgment index based on the sub-cycle sample enhancement evaluation parameters, and determine whether a fault occurs in the historical sub-cycle based on the sample enhancement judgment index. Step 4: If it is determined that fault samples need to be supplemented in the historical sub-cycle, then fault samples are supplemented in the historical sub-cycle according to the sample enhancement judgment index. All historical sub-cycles are traversed to obtain the supplemented historical sample sequence; Step 5: Construct an actual power fault prediction model based on the supplemented historical sample sequence; Step 6: Collect real-time operating data within the current power equipment operating cycle; Step 7: Input the real-time operating data into the actual power fault prediction model, output the prediction probability value at the current moment, compare the prediction probability value at the current moment with the set risk threshold, and if it exceeds the threshold, generate the corresponding power equipment fault early warning information.

[0016] The technical effects and advantages of this invention are as follows: An initial historical sample sequence is obtained. Based on the initial historical sample sequence, it is determined whether fault samples need to be supplemented. If it is determined that fault samples need to be supplemented, a synthetic fault sample set is generated. The historical operating cycle is divided into several historical sub-cycles. The sub-cycle sample enhancement evaluation parameters of each historical sub-cycle are obtained, and the sample enhancement judgment index is evaluated to determine whether fault samples need to be supplemented in the historical sub-cycle. If it is determined that fault samples need to be supplemented in the historical sub-cycle, fault samples are supplemented in the historical sub-cycle. All historical sub-cycles are traversed to obtain the supplemented historical sample sequence. Based on the supplemented historical sample sequence, an actual power fault prediction model is constructed, which effectively improves the accuracy of power fault prediction. Attached Figure Description

[0017] Figure 1 A structural diagram of a power fault prediction system provided in this application embodiment; Figure 2 A flowchart of a power fault prediction method provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The power fault prediction system and method involved in the present invention are not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a power fault prediction system, such as Figure 1 As shown, the system includes: The historical data acquisition module is used to collect and store historical data of power equipment during historical operating cycles, and construct an initial historical sample sequence. The historical data includes operating data, environmental parameters, and marked historical fault records. The operating data includes indicators such as current, voltage, frequency, active power, and temperature. It should be noted that the operating cycle can be varied according to the actual situation; for example, the operating cycle can be 1 month or 3 months.

[0020] The sample supplementation determination module is used to determine whether fault samples need to be supplemented based on the initial historical sample sequence. If it is determined that fault samples need to be supplemented, a synthetic fault sample set is generated through perturbation and SMOTE, and the synthetic fault sample set is transmitted to the sample supplementation module. If it is determined that fault samples do not need to be supplemented, the historical sample sequence is transmitted to the supplementary historical sample sequence construction module. Perturbation and SMOTE are two commonly used sample augmentation methods, mainly used to generate synthetic samples to alleviate class imbalance when there are insufficient minority class samples in the training samples. The perturbation method generates new samples that are close to but not exactly the same as the original samples by applying a small perturbation (such as adding Gaussian noise or scrambling some feature values) to the feature vector of the original samples. The SMOTE method generates new samples within the distribution range of the original samples by performing linear interpolation between minority class samples and their neighboring samples in the feature space.

[0021] It should be noted that generating a synthetic fault sample set through perturbation and SMOTE is an existing technology, and this embodiment will not describe its specific steps in detail.

[0022] In this embodiment, it should be specifically explained that the step of determining whether to supplement fault samples based on the initial historical sample sequence is as follows: Obtain the total number of samples in the initial historical sample sequence, count the number of fault samples in the initial historical sample sequence, and calculate the ratio of the number of fault samples to the total number of samples to obtain the historical fault percentage. The historical sample sequence is divided into several sub-sample sequences on an average basis. Each sub-sample sequence contains a corresponding sample subset. The number of fault samples in each sub-sample sequence is obtained and the ratio is calculated with the total number of samples in the sub-sample sequence to obtain the fault percentage of the sub-sequence. The variance of the fault proportion of all sub-sequences is calculated to obtain the fault distribution dispersion, which is used to measure the time balance of fault samples. The larger the fault sample distribution dispersion, the more concentrated the fault sample distribution is and the worse the coverage. The sample enhancement start-up judgment index is calculated based on the historical failure rate and failure distribution dispersion. The specific steps for obtaining the index are as follows: ; In the formula, This is represented as the sample augmentation priming index. This is expressed as the percentage of historical failures. This is expressed as the fault distribution dispersion, which comprehensively considers both the global proportion of fault samples and the temporal distribution dispersion, jointly evaluating the scarcity and coverage balance of the historical sample structure. The scarcity of the fault sample proportion serves as a measure of insufficient sample quantity, while the distribution dispersion reflects the concentration of samples over time. Multiplying these two values ​​creates an exponential value, ensuring that the judgment value is significantly improved only when the sample size is both small and unevenly distributed. This effectively avoids false triggering of enhancement operations on already representative datasets, improving the accuracy of fault sample supplementation strategies and the efficiency of system resource utilization. The sample enhancement start-up determination index is compared with the start-up threshold. If the sample enhancement start-up determination index is greater than or equal to the start-up threshold, it is determined that faulty samples need to be added; if the sample enhancement start-up determination index is less than the start-up threshold, it is determined that faulty samples do not need to be added. The start-up threshold is obtained by the adaptive threshold method.

[0023] The sample supplementation module is used to divide the historical operating cycle into several historical sub-cycles on average, obtain the sub-cycle sample enhancement evaluation parameters for each historical sub-cycle, including the number of fault samples, the labels of historical samples, and the output structure of the fault prediction model. The sample enhancement judgment index is obtained based on the sub-cycle sample enhancement evaluation parameters, and it is determined whether fault samples need to be supplemented in the historical sub-cycle based on the sample enhancement judgment index. In this embodiment, it should be specifically explained that the steps for obtaining the sample augmentation determination index are as follows: Obtain the total number of sub-cycle samples in each historical sub-cycle, obtain the number of sub-cycle fault samples in each historical sub-cycle, and calculate the ratio of the number of sub-cycle fault samples to the total number of sub-cycle samples to obtain the sub-cycle fault percentage. Obtain the label of each historical sample in the model prediction log within each historical sub-period, and evaluate the historical prediction misjudgment impact coefficient based on the label of each historical sample. Obtain the output structure of the fault prediction model within each historical sub-cycle, and evaluate the prediction gradient sensitivity coefficient based on the output structure of the fault prediction model. The sub-cycle fault ratio, historical prediction misjudgment impact coefficient, and prediction gradient sensitivity impact coefficient are normalized. The sample augmentation judgment index is then evaluated based on the normalized local sample ratio, historical prediction misjudgment impact coefficient, and prediction gradient sensitivity impact coefficient. The specific steps for obtaining this index are as follows: ; In the formula, This is represented as the sample augmentation determination index. This represents the proportion of fault samples in a sub-cycle after normalization. A high proportion of fault samples indicates sufficient fault information to support model learning, eliminating the need for further sample supplementation; therefore, the sample augmentation judgment index decreases. Conversely, a low proportion of fault samples suggests insufficient fault information in that sub-cycle, potentially causing the model to ignore minority class features during training. In this case, the sample augmentation judgment index increases to drive the synthesis and supplementation of fault samples, thereby enhancing the model's ability to learn and identify low-frequency fault patterns. This represents the normalized historical prediction misjudgment impact coefficient. When there is a high rate of misjudgment in the historical predictions of fault states (e.g., high false positive or false negative rates), the prediction reliability of that sub-cycle is poor, indicating that the current training samples have not effectively supported the model's ability to identify this type of fault, and further fault samples need to be augmented to improve model performance. Therefore, the higher the misjudgment impact coefficient, the more the system tends to increase the sample augmentation judgment index, thereby triggering fault sample supplementation operations and improving the model's accuracy and robustness in this type of fault scenario. This is represented by the normalized prediction gradient sensitivity coefficient. When the predicted output of a fault sample in the model exhibits strong gradient variability to input perturbations, it indicates that the model has not yet fully learned the characteristics of this type of fault, or that the driving effect of fault samples on model learning is unstable, resulting in high training sensitivity and weak generalization ability. In this case, the sample augmentation judgment index should be increased to determine whether there is still a strong need for sample supplementation in this sub-cycle, thereby guiding fault sample augmentation operations to enhance the model's learning stability and discrimination ability for this type of sample. These represent the weighting coefficients for the proportion of sub-cycle faults after normalization, the weighting coefficients for the impact of historical prediction misjudgments after normalization, and the weighting coefficients for the impact of prediction gradient sensitivity after normalization. Obtained through the Analytic Hierarchy Process (AHP), and ,For example The weights can be 0.3, 0.5, or 0.2. The Analytic Hierarchy Process (AHP) is a method for determining the weights of influencing factors in multi-factor decision-making problems to assess their relative importance. This method constructs a judgment matrix, compares multiple evaluation indicators pairwise, extracts the decision-maker's subjective judgment of the relative importance of each indicator, performs a consistency test, and finally obtains a set of normalized weight coefficients. This method has the advantages of clear structure and strong adaptability, and is suitable for determining the weight ratio of each influencing factor in a sample augmentation judgment index to enhance the rationality and interpretability of the decision-making logic.

[0024] In this embodiment, it should be specifically explained that the steps for obtaining the historical prediction misjudgment impact coefficient are as follows: Obtain model prediction logs for each historical sub-cycle within the historical operating cycle. The prediction logs include the prediction labels and corresponding actual labels of the prediction model for historical samples in each historical sub-cycle. The actual labels are obtained from historical fault records and include two categories: "fault" and "non-fault," which are used for subsequent misjudgment statistical analysis. For each historical sub-cycle, iterate through all historical samples within that sub-cycle, and count the number of false alarms and the number of false alarms. The number of false alarms is the number of samples that are actually non-faulty but predicted to be faulty, and the number of false alarms is the number of samples that are actually faulty but predicted to be non-faulty. Sum the number of false alarms and the number of false alarms to obtain the number of misjudged samples. For each historical sub-period, the total number of samples within the historical sub-period is obtained, and the ratio of the number of misjudged samples to the total number of samples is calculated to obtain the historical sub-period misjudgment rate. Obtain the predicted probability of all samples within the historical sub-period, i.e. the failure probability value output by the model, and calculate the standard deviation of the predicted probability of all samples to obtain the prediction confidence volatility. The historical prediction misjudgment impact coefficient is calculated based on the historical sub-cycle misjudgment rate and the prediction confidence volatility. The specific steps for obtaining this coefficient are as follows: ; In the formula, This is represented as the historical prediction misjudgment impact coefficient. This is expressed as the historical sub-cycle misjudgment rate. This represents the predicted confidence volatility, where +1 is used to avoid the case where the predicted confidence volatility is 0.

[0025] In this embodiment, it should be specifically explained that the steps for obtaining the predicted gradient sensitivity influence coefficient are as follows: Obtain the output layer weight matrix and activation function form of the fault prediction model structure deployed in the historical operating cycle. The output structure is represented as follows: ; In the formula, Indicates the probability of failure. The feature vector representing a historical sub-cycle sample. This is the output layer weight vector. For bias terms, The Sigmoid activation function converts the model's output into a probability value (between 0 and 1). This indicates that the output layer weight vector is transposed. A bias term is a trainable parameter, independent of the input features, introduced in a neural network or machine learning model to improve the model's fitting ability. In the output structure of the fault prediction model in this embodiment, the bias term participates in the linear transformation process along with the weight vector and the input feature vector. The role of the bias term is to adjust the input position of the activation function, so that the model can produce a non-zero output even when all input features are zero, thereby improving the model's ability to represent data and its fitting flexibility.

[0026] The output layer weight matrix refers to the set of weight vectors connecting the last layer of the fault prediction model deployed throughout the historical operating cycle to the output node. In this embodiment, a single-output structure is adopted, meaning that fault state prediction is a binary classification problem, and the output layer contains only one node. Therefore, the weight matrix can be simplified to a weight vector, denoted as […]. .

[0027] The activation function form refers to the type of nonlinear function used in the output layer of the fault prediction model, which maps the linear output result to a probability value. In this embodiment, the standard Sigmoid function is selected as the activation function, which can compress and map any real number to the (0,1) interval, making it suitable for representing the predicted probability of fault occurrence.

[0028] Obtain the historical sample sequence of each historical sub-cycle in the historical operating cycle, and construct the input feature matrix of the historical sub-cycle. The input feature matrix includes several input feature vectors. For each input feature vector within a historical sub-period, the gradient of the output probability with respect to the input feature is calculated. The specific steps are as follows: ; In the formula, This is expressed as the output result (i.e., the failure probability) when each dimension of the input feature vector undergoes a small change. How it will change is a gradient vector. Represented as gradient, Indicates the probability of failure. The derivative of the Sigmoid activation function is given by the regular differential rule. This is the output layer weight vector; Arrange all gradients in chronological order to obtain a gradient vector sequence. Calculate the Euclidean distance between any two adjacent gradient vectors in sequence to obtain a gradient rate of change sequence. The arithmetic mean of the gradient change rate sequence is used to obtain the predicted gradient sensitivity coefficient. The larger the value, the more severe the model's response to input disturbances, and the more sensitive the prediction.

[0029] In this embodiment, it should be specifically explained that the step of determining whether faulty samples need to be supplemented in the historical sub-period based on the sample enhancement judgment index is as follows: The sample augmentation determination index is compared with the augmentation determination threshold. If the sample augmentation determination index is greater than or equal to the augmentation determination threshold, it is determined that faulty samples need to be supplemented within the historical sub-cycle; if the sample augmentation determination index is less than the augmentation determination threshold, it is determined that faulty samples do not need to be supplemented within the historical sub-cycle. The augmentation determination threshold is obtained through an adaptive threshold method. The adaptive threshold method refers to dynamically setting an augmentation determination threshold suitable for the current data structure and model state by analyzing the distribution characteristics and changing trends of the sample augmentation determination index in historical operating cycles or sub-cycles, combined with the historical records of whether sample supplementation operations have been performed in different cycles. Based on a data-driven approach, the differences in sample distribution and model response status in different cycles are automatically perceived, thereby adjusting the determination threshold to improve the adaptability, accuracy, and robustness of the sample augmentation triggering mechanism under multiple operating conditions.

[0030] The supplementary historical sample sequence construction module, if it is determined that faulty samples need to be supplemented in the historical sub-cycle, then the faulty samples are supplemented in the historical sub-cycle according to the sample enhancement judgment index, and all historical sub-cycles are traversed to obtain the supplementary historical sample sequence. In this embodiment, it should be specifically explained that the step of supplementing faulty samples in the historical sub-periods according to the sample enhancement judgment index is as follows: The ratio of the sample augmentation judgment index to the augmentation judgment threshold is calculated and then rounded up to obtain the number of supplementary samples. Set a maximum number of supplementary samples. If the number of supplementary samples is less than or equal to the maximum number of supplementary samples, then randomly select the corresponding number of supplementary samples from the synthetic fault sample set and randomly insert them into the historical samples within the historical sub-cycle. If the number of supplementary samples is greater than the maximum number of supplementary samples, then a fault sample corresponding to the maximum number of supplementary samples is randomly selected from the synthetic fault sample set and randomly inserted into the historical samples within the historical sub-cycle.

[0031] The ratio of the sample enhancement judgment index to the enhancement judgment threshold is calculated and rounded up to dynamically determine the number of fault samples to be added in each historical sub-cycle, enabling differentiated responses to the degree of sample structural anomalies in each sub-cycle. Furthermore, setting a maximum addition number as an upper limit constraint effectively prevents training data imbalance or overfitting caused by excessive synthetic samples. The random insertion strategy improves sample diversity and generalization ability, avoiding model bias caused by concentrated addition locations. This approach ensures the effectiveness of addition while balancing training stability and model robustness, improving the reliability and practicality of the fault prediction model in scenarios with insufficient minority fault samples.

[0032] The fault prediction model building module is used to build a real power fault prediction model based on supplementary historical sample sequences; In this embodiment, it should be specifically explained that the construction steps of the actual power fault prediction model are as follows: Step 5.1: Divide the historical data of the supplementary historical sample sequence into a training sample set and a validation sample set. The training sample set includes the sequence of operating indicators and the corresponding fault labels. The validation sample set is used for subsequent model performance evaluation. The fault prediction model is trained using the training sample set to obtain the initial power fault prediction model. Step 5.2: Obtain the performance metrics of the initial power fault prediction model based on the validation sample set. These metrics include accuracy, precision, recall, and F1 score. Calculate the model qualification index based on these performance metrics. The specific steps for obtaining this index are as follows: ; In the formula, This is represented as the model qualification index. as well as These are represented as accuracy, precision, recall, and F1 score, respectively. The model qualification index is compared with the qualification threshold. If the model qualification index is greater than or equal to the qualification threshold, the current power fault prediction model is deemed qualified and can be used for fault prediction. The initial power fault prediction model is then recorded as the actual power fault prediction model. The actual power fault prediction model is a supervised classification model that supports probability output. The actual power fault prediction model is trained based on supplementary historical sample sequences and can adopt structural forms such as logistic regression, gradient boosting tree, or neural network with Sigmoid function in the output layer. The model outputs the predicted probability of the equipment being in a fault state at the current moment to the input real-time operating data. The value range is [0,1], which can be compared with the set risk threshold to determine whether to generate fault warning information. Step 5.3: If the model qualification index is less than the qualification threshold, the current power fault prediction model is deemed unqualified and cannot perform fault prediction. Then, return to step 5.1 to continue model training until the current power fault prediction model is deemed capable of performing fault prediction. Record the current power fault prediction model as the actual power fault prediction model. The qualification threshold is obtained through the adaptive threshold method.

[0033] The real-time data acquisition module is used to collect real-time operating data within the current operating cycle of the power equipment. The real-time operating data includes electrical operating parameters and environmental parameters that are consistent with the historical data structure. The fault risk warning and assessment module is used to input real-time operating data into the actual power fault prediction model, output the predicted probability value at the current moment, compare the predicted probability value at the current moment with the set risk threshold, and if the threshold is exceeded, generate the corresponding power equipment fault warning information and push it to the operation and maintenance system or personnel terminal.

[0034] In this embodiment, it is necessary to specifically explain a power fault prediction method, such as... Figure 2 As shown, it includes the following steps: Step 1: Collect and store historical data of power equipment during its historical operating cycle to construct an initial historical sample sequence; Step 2: Determine whether it is necessary to supplement fault samples based on the initial historical sample sequence. If it is determined that fault samples need to be supplemented, generate a synthetic fault sample set through perturbation and SMOTE. Step 3: Divide the historical operating cycle into several historical sub-cycles, obtain the sub-cycle sample enhancement evaluation parameters for each historical sub-cycle, evaluate the sample enhancement judgment index based on the sub-cycle sample enhancement evaluation parameters, and determine whether to supplement fault samples in the historical sub-cycle based on the sample enhancement judgment index. Step 4: If it is determined that fault samples need to be added in the historical sub-cycle, then fault samples are added in the historical sub-cycle according to the sample enhancement determination index. This process is repeated for all historical sub-cycles to obtain the supplementary historical sample sequence. Step 5: Construct an actual power fault prediction model based on supplementary historical sample sequences; Step 6: Collect real-time operating data within the current power equipment operating cycle; Step 7: Input the real-time operating data into the actual power fault prediction model, output the predicted probability value at the current moment, compare the predicted probability value at the current moment with the set risk threshold, and if the threshold is exceeded, generate the corresponding power equipment fault warning information.

[0035] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0036] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A power fault prediction system, characterized in that, The system includes: The historical data acquisition module is used to collect and store historical data of power equipment during historical operating cycles, and to construct an initial historical sample sequence. The sample supplementation determination module is used to determine whether fault samples need to be supplemented based on the initial historical sample sequence. If it is determined that fault samples need to be supplemented, a synthetic fault sample set is generated by perturbation and SMOTE, and the synthetic fault sample set is transmitted to the sample supplementation module. If it is determined that no additional fault samples are needed, the historical sample sequence will be transferred to the supplementary historical sample sequence construction module. The sample supplementation module is used to divide the historical operating cycle into several historical sub-cycles on average, obtain the sub-cycle sample enhancement evaluation parameters for each historical sub-cycle, including the number of fault samples, the labels of historical samples, and the output structure of the fault prediction model. The sample enhancement judgment index is obtained based on the sub-cycle sample enhancement evaluation parameters, and it is determined whether fault samples need to be supplemented in the historical sub-cycle based on the sample enhancement judgment index. The supplementary historical sample sequence construction module, if it is determined that faulty samples need to be supplemented in the historical sub-cycle, then the faulty samples are supplemented in the historical sub-cycle according to the sample enhancement judgment index, and all historical sub-cycles are traversed to obtain the supplementary historical sample sequence. The fault prediction model building module is used to build a real power fault prediction model based on supplementary historical sample sequences; The real-time data acquisition module is used to collect real-time operating data within the current operating cycle of the power equipment. The real-time operating data includes electrical operating parameters and environmental parameters that are consistent with the historical data structure. The fault risk warning and assessment module is used to input real-time operating data into the actual power fault prediction model, output the predicted probability value at the current moment, compare the predicted probability value at the current moment with the set risk threshold, and generate corresponding power equipment fault warning information if the threshold is exceeded.

2. The power fault prediction system according to claim 1, characterized in that: The step of determining whether to supplement fault samples based on the initial historical sample sequence is as follows: Obtain the total number of samples in the initial historical sample sequence, count the number of fault samples in the initial historical sample sequence, and calculate the ratio of the number of fault samples to the total number of samples to obtain the historical fault percentage. The historical sample sequence is divided into several sub-sample sequences on an average basis. Each sub-sample sequence contains a corresponding sample subset. The number of fault samples in each sub-sample sequence is obtained and the ratio is calculated with the total number of samples in the sub-sample sequence to obtain the fault ratio of the sub-sequence. The variance of the fault distribution dispersion is obtained by calculating the proportion of faults in all sub-sample sequences. The sample-enhanced startup judgment index is calculated based on the historical failure rate and the failure distribution dispersion. The need to supplement fault samples is determined based on the sample enhancement start-up judgment index.

3. The power fault prediction system according to claim 2, characterized in that, The step of determining whether to supplement fault samples based on the sample enhancement startup judgment index is as follows: The sample enhancement start-up determination index is compared with the start-up threshold. If the sample enhancement start-up determination index is greater than or equal to the start-up threshold, it is determined that faulty samples need to be added; if the sample enhancement start-up determination index is less than the start-up threshold, it is determined that faulty samples do not need to be added.

4. The power fault prediction system according to claim 1, characterized in that, The steps for obtaining the sample augmentation determination index are as follows: Obtain the total number of sub-cycle samples in each historical sub-cycle, obtain the number of sub-cycle fault samples in each historical sub-cycle, and calculate the ratio of the number of sub-cycle fault samples to the total number of sub-cycle samples to obtain the sub-cycle fault ratio. Obtain the label of each historical sample in the model prediction log within each historical sub-period, and evaluate the historical prediction misjudgment impact coefficient based on the label of each historical sample. Obtain the output structure of the fault prediction model within each historical sub-cycle, and evaluate the prediction gradient sensitivity coefficient based on the output structure of the fault prediction model. The sub-cycle fault ratio, historical prediction misjudgment impact coefficient, and prediction gradient sensitivity impact coefficient are normalized. The sample augmentation judgment index is then evaluated based on the normalized local sample ratio, historical prediction misjudgment impact coefficient, and prediction gradient sensitivity impact coefficient. The specific steps for obtaining this index are as follows: ; In the formula, This is represented as the sample augmentation determination index. This represents the percentage of sub-cycle failures after normalization. This is represented by the normalized historical prediction misjudgment impact coefficient. This is expressed as the normalized prediction gradient sensitivity coefficient. These are represented as the weighting coefficients for the proportion of sub-cycle faults after normalization, the weighting coefficients for the impact of historical prediction misjudgments after normalization, and the weighting coefficients for the impact of prediction gradient sensitivity after normalization.

5. A power fault prediction system according to claim 4, characterized in that: The steps for obtaining the historical prediction misjudgment impact coefficient are as follows: Obtain the model prediction logs for each historical sub-cycle within the historical operating cycle. The prediction logs include the prediction labels and corresponding actual labels of the prediction model for historical samples in each historical sub-cycle. The labels include two categories: "fault" and "non-fault". For each historical sub-cycle, iterate through all historical samples within that sub-cycle, count the number of false alarms and the number of missed alarms, and sum the number of false alarms and the number of missed alarms to obtain the number of misjudged samples. For each historical sub-period, the total number of samples within the historical sub-period is obtained, and the ratio of the number of misjudged samples to the total number of samples is calculated to obtain the historical sub-period misjudgment rate. Obtain the predicted probability of all samples within the historical sub-period, calculate the standard deviation of the predicted probability of all samples, and obtain the predicted confidence volatility. The historical prediction misjudgment impact coefficient is calculated based on the historical sub-cycle misjudgment rate and the prediction confidence volatility.

6. A power fault prediction system according to claim 4, characterized in that: The steps for obtaining the predicted gradient sensitivity influence coefficient are as follows: Obtain the output layer weight matrix and activation function form of the fault prediction model structure deployed in the historical operation cycle; Obtain the historical sample sequence of each historical sub-cycle in the historical operating cycle, and construct the input feature matrix of the historical sub-cycle. The input feature matrix includes several input feature vectors. For each input feature vector within a historical sub-period, calculate the gradient of the output probability with respect to the input feature. Arrange all gradients in chronological order to obtain a gradient vector sequence. Calculate the Euclidean distance between any two adjacent gradient vectors in sequence to obtain a gradient rate of change sequence. The gradient change rate sequence is arithmetically averaged to obtain the predicted gradient sensitivity coefficient.

7. The power fault prediction system according to claim 1, characterized in that: The step of determining whether faulty samples need to be supplemented in the historical sub-period based on the sample enhancement judgment index is as follows: The sample enhancement judgment index is compared with the enhancement judgment threshold. If the sample enhancement judgment index is greater than or equal to the enhancement judgment threshold, it is determined that faulty samples need to be added in the historical sub-cycle. If the sample enhancement judgment index is less than the enhancement judgment threshold, it is determined that no faulty samples need to be added in the historical sub-cycle.

8. A power fault prediction system according to claim 7, characterized in that: The step of supplementing faulty samples within the historical sub-period based on the sample enhancement determination index is as follows: The ratio of the sample enhancement judgment index to the enhancement judgment threshold is calculated and then rounded up to obtain the number of supplementary samples. Set a maximum number of supplementary samples. If the number of supplementary samples is less than or equal to the maximum number of supplementary samples, then randomly select the corresponding number of supplementary samples from the synthetic fault sample set and randomly insert them into the historical samples within the historical sub-cycle. If the number of supplementary samples is greater than the maximum number of supplementary samples, then a fault sample corresponding to the maximum number of supplementary samples is randomly selected from the synthetic fault sample set and randomly inserted into the historical samples within the historical sub-cycle.

9. A power fault prediction system according to claim 1, characterized in that: The steps for constructing the actual power fault prediction model are as follows: Step 5.1: Divide the historical data of the supplementary historical sample sequence into a training sample set and a validation sample set. Use the training sample set to train the fault prediction model to obtain the initial power fault prediction model. Step 5.2: Obtain the performance index of the initial power fault prediction model based on the validation sample set, and calculate the model qualification index based on the performance index; The model qualification index is compared with the qualification threshold. If the model qualification index is greater than or equal to the qualification threshold, the current power fault prediction model is deemed qualified and can be used for fault prediction. The initial power fault prediction model is then recorded as the actual power fault prediction model. Step 5.3: If the model qualification index is less than the qualification threshold, the current power fault prediction model is deemed unqualified and cannot perform fault prediction. Then, return to step 5.1 to continue model training until the current power fault prediction model is deemed capable of performing fault prediction. Record the current power fault prediction model as the actual power fault prediction model.

10. A power fault prediction method according to claim 1, used to implement a power fault prediction system according to any one of claims 1-9, characterized in that: Includes the following steps: Step 1: Collect and store historical data of power equipment during its historical operating cycle to construct an initial historical sample sequence; Step 2: Determine whether it is necessary to supplement fault samples based on the initial historical sample sequence. If it is determined that fault samples need to be supplemented, generate a synthetic fault sample set through perturbation and SMOTE. Step 3: Divide the historical operating cycle into several historical sub-cycles, obtain the sub-cycle sample enhancement evaluation parameters for each historical sub-cycle, evaluate the sample enhancement judgment index based on the sub-cycle sample enhancement evaluation parameters, and determine whether to supplement fault samples in the historical sub-cycle based on the sample enhancement judgment index. Step 4: If it is determined that fault samples need to be added in the historical sub-cycle, then fault samples are added in the historical sub-cycle according to the sample enhancement determination index. This process is repeated for all historical sub-cycles to obtain the supplementary historical sample sequence. Step 5: Construct an actual power fault prediction model based on supplementary historical sample sequences; Step 6: Collect real-time operating data within the current power equipment operating cycle; Step 7: Input the real-time operating data into the actual power fault prediction model, output the predicted probability value at the current moment, compare the predicted probability value at the current moment with the set risk threshold, and if the threshold is exceeded, generate the corresponding power equipment fault warning information.

Images