An intelligent simulation deduction and dynamic early warning method and system for multi-source state information of electrical equipment

Abstract

The application relates to an intelligent simulation deduction and dynamic early warning method and system for multi-source state information of electrical equipment, and the method comprises the following steps: collecting multi-source state information reflecting the running state of electrical equipment in real time, forming a multi-dimensional time sequence data set, and pre-processing to obtain a standardized feature matrix; based on the standardized feature matrix, a time sequence learning model adopts a dynamic deduction mechanism based on cycle prediction driving to recursively predict and dynamically evolve the running state of the electrical equipment in the time dimension, and a running state prediction sequence of a plurality of future time steps evolving with time is obtained; based on the running state prediction sequence of the plurality of future time steps evolving with time, an adaptive early warning threshold is set, whether an early warning event is automatically triggered and feedback information is generated are judged, and a dynamic early warning process is completed. Compared with the prior art, the application has the advantages of accurate perception, timely early warning and the like.

Description

Technical Field

[0001] This invention relates to the field of novel power systems, and in particular to an intelligent simulation and dynamic early warning method and system for multi-source state information of electrical equipment. Background Technology

[0002] With the deepening of the construction of new power systems and the continuous improvement of equipment intelligence, the operating environment and load characteristics of electrical equipment are becoming increasingly complex. Against the backdrop of high-proportion renewable energy integration, widespread application of power electronic devices, and continuous expansion of distributed resources, equipment operating states exhibit characteristics such as multivariate coupling, nonlinear evolution, and time-varying uncertainty. Traditional equipment monitoring and early warning methods often rely on single-parameter threshold judgments or static model analysis, which are insufficient to meet the demands of modern power grids for real-time performance, accuracy, and adaptability.

[0003] Currently, key electrical equipment such as transformers, switchgear, and reactors are widely equipped with various types of sensors, capable of collecting multi-source status information such as voltage, current, temperature, vibration, partial discharge, and oil gas content. However, different sensors exhibit significant differences in sampling frequency, timing accuracy, and data reliability, making it difficult to achieve unified fusion of multi-source information and limiting the comprehensive characterization and dynamic perception of equipment operating status. Furthermore, traditional methods based on mechanistic modeling or expert experience lack robustness in handling complex coupling and dynamic operating conditions, failing to fully utilize the temporal correlation features inherent in historical data, resulting in a significant decline in predictive performance under multiple operating conditions.

[0004] On the other hand, existing fault early warning methods are generally based on fixed thresholds or rule bases, which have limited ability to respond dynamically to changes in equipment status over time, easily leading to problems such as delayed early warning, high false alarm rates, and difficulties in tracing the source. For example, patent application CN120891243A discloses a transformer state deduction and fault prediction method and system. Especially under multi-physics coupling conditions, there are potential causal relationships between different characteristic variables, making it difficult for single measurement analysis to reveal the evolution law of equipment status. To address these problems, although the industry has attempted to introduce data-driven methods such as deep learning and reinforcement learning for equipment status identification and anomaly detection, limitations remain, including a lack of unified standards for data preprocessing, limited model prediction step size, rigid early warning mechanisms, and a lack of closed-loop optimization.

[0005] Therefore, those skilled in the art are dedicated to researching an intelligent simulation early warning method that can integrate multi-source state information, possess dynamic evolution and deduction capabilities, and achieve self-learning and self-correction, so as to comprehensively improve the accuracy, timeliness, and intelligence level of electrical equipment condition monitoring and provide systematic technical support for equipment health management and risk prevention and control. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent simulation and dynamic early warning method and system for electrical equipment based on multi-source status information, which significantly improves the accuracy of equipment status perception and the timeliness of early warning.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for intelligent simulation and dynamic early warning of multi-source state information of electrical equipment includes the following steps: Multi-source status information reflecting the operating status of electrical equipment is collected in real time to form a multi-dimensional time series dataset, which is then preprocessed to obtain a standardized feature matrix. Based on the standardized feature matrix, the time-series learning model uses a dynamic inference mechanism driven by cyclic prediction to recursively predict and dynamically evolve the operating status of electrical equipment in the time dimension, thereby obtaining a prediction sequence of the operating status for multiple future time steps that evolve over time. Based on the predicted sequence of operating states for multiple future time steps that evolve over time, and using a set adaptive warning threshold, it is determined whether to automatically trigger a warning event and generate feedback information, thus completing the dynamic warning process.

[0008] Furthermore, the multi-source state information includes voltage, current, temperature, vibration, partial discharge, and gas content in the oil.

[0009] Furthermore, the preprocessing steps include: Let the multidimensional time series dataset be... , Indicates time of 3D feature vectors The sampling length; Based on the aforementioned multidimensional time-series dataset, for the sampling timestamp sets of different channels Through time interpolation function To achieve uniform resampling and align all multidimensional time-series data on the same time axis, a standardized multidimensional time-series dataset is obtained. The feature vector for each dimension of this standardized multidimensional time-series dataset is represented as follows: , In the formula, For at any time No. The feature vector after dimensional standardization; Based on the standardized multidimensional time-series dataset, the mean and standard deviation are calculated using the sliding window statistical method, and the range of each channel is constructed based on the mean and standard deviation. When the standardized feature vector in each dimension Deviating from the range If an outlier occurs, it is considered an anomaly and either replaced or removed; otherwise, it is considered normal, resulting in a multidimensional time-series dataset with outliers removed. , The first one in the sliding window Mean and standard deviation in each dimension; Based on the multidimensional time-series dataset with outliers removed, normalization processing is performed to obtain a standardized feature matrix, wherein the normalization expression is: , In the formula, and These represent the functions for finding the maximum and minimum values, respectively. In the first The feature vector after dimensional normalization.

[0010] Furthermore, the preprocessing steps also include: after normalization, using the sliding window method to complete missing data and perform local smoothing, specifically including: Set the sliding window length to , the feature vectors that do not have missing data and are not judged as abnormal within the sliding window are determined as valid feature vectors, and the feature vectors that have missing data or are judged as abnormal and then removed after removing the outliers are determined as invalid feature vectors. Count the number of valid feature vectors within the current sliding window. And calculate the proportion of effective feature vectors. : , when When this happens, no missing data completion is performed on the current sliding window, and the sliding window is moved to the next position to continue processing. The effective feature vector ratio threshold; when When missing data is filled in, linear interpolation or spline interpolation is performed using the effective feature vectors adjacent to the missing position within the current sliding window.

[0011] Furthermore, the step of obtaining the predicted sequence of operating states for multiple future time steps that evolve over time includes: Let the standardized feature matrix be... This is transformed into a sliding window input structure, yielding the input sequence for each sliding window. , For a moment eigenvectors, The length of the sliding window; Based on the input sequence, the time-series learning model extrapolates future time steps through forward propagation. The output of the predicted running status, where The number of steps in advance for predicting an evolution; The future time step The predicted operating state output is used as the input to the time series learning model to continue the extrapolation and obtain the future time steps. The predicted operating state output is used to continuously extrapolate and obtain the predicted operating state sequence for multiple future time steps that evolve over time. subscript This indicates the length of the predicted sequence for the running status.

[0012] Furthermore, within a set period, the weight parameters of the time-series learning model are adjusted in reverse using the time backpropagation algorithm to reduce the overall loss function. Specific steps include: The weight parameter adjustment steps of the time-series learning model include: Within a set period, obtain the predicted output of the running state at future time steps and the corresponding actual observed value of the running state, and calculate the prediction error between the two. The overall loss function of the time-series learning model is constructed based on the prediction error; The time-series learning model is expanded along the time dimension, and the gradient of the overall loss function with respect to each weight parameter in the time-series learning model is calculated using the time backpropagation algorithm. The weight parameters are iteratively updated based on their gradients to reduce the overall loss function, and the updated time-series learning model is used for subsequent continuous deduction.

[0013] Furthermore, it also includes: When a new batch of multi-source state information is collected and preprocessed to obtain newly standardized feature data, the optimal weight parameter set of the previously trained temporal learning model is acquired. Incremental training is performed using a breakpoint-based continuation mechanism, and the specific incremental training includes: Load the optimal weight parameter set saved from the previous training cycle. And use it as the initial weight parameter set for the current training cycle of the time-series learning model; An incremental training input sequence is constructed based on the newly added standardized feature data and then input into the time-series learning model; Based on the incremental training input sequence, the weight parameter set of the time series learning model is adjusted in reverse using the time backpropagation algorithm to reduce the overall loss function of the time series learning model and obtain the updated weight parameter set. The updated set of weight parameters is saved for subsequent runtime state deduction and breakpoint continuation training in the next training cycle.

[0014] Furthermore, the step of determining whether to automatically trigger a warning signal and generate feedback information includes: Based on multi-source state information collected during historical operation, a standardized feature matrix is ​​obtained after preprocessing. A sliding window length is then established for each feature vector in the standardized feature matrix. For the sliding window statistical model, the mean and standard deviation of each feature vector within the sliding window are calculated, and an adaptive warning threshold is set based on the mean and standard deviation, expressed as: , In the formula, For adaptive early warning threshold, and These are the mean and standard deviation of the state variables, respectively. This is the threshold coefficient, used to control the sensitivity of the early warning. Predicting the running state at a future time step Does it meet the requirements? If yes, then the future time step is identified as an abnormal candidate time step; otherwise, the operating state of the future time step is determined to be within the normal fluctuation range. Predicting the running state at a future time step Are all predicted state variables outside the set two-sided threshold range? If yes, then the future time step is identified as an abnormal candidate time step; otherwise, the operating state of the future time step is determined to be within the normal fluctuation range. Introduce stability constraints for anomaly detection: Let the number of future time steps for continuous judgment be . , ,in, Indicates the first judgment in a continuous judgment sequence One future time step; When continuous Predicted operating state values ​​at future time steps All meet If the condition is met, an anomaly is confirmed and an early warning event is automatically triggered; otherwise, it indicates that the continuous condition has not been met. The continuous exceeding of the limit in a future time step will be continuous Anomaly candidate time steps within a given future time step are determined to be transient disturbances and do not trigger warning events; After an automatic warning event is triggered, feedback information and a warning record table are generated, wherein the warning record table includes: the time of the anomaly. Predicted state variable name, deviation magnitude The feedback information includes the feature dimension index corresponding to the abnormal prediction state quantity and the unique identifier of the electrical equipment. This feedback information is used to statistically analyze the prediction deviation within the abnormal period. When the prediction deviation within consecutive abnormal periods exceeds a preset self-correction threshold... If a similar prediction bias pattern appears in multiple warning events, the learning rate or regularization parameter of the time series learning model will be automatically adjusted, and the weight parameters of the time series learning model will be updated based on the adjusted learning rate or regularization parameter.

[0015] Furthermore, it also includes: after the early warning event is triggered, a source tracing analysis is performed to identify the abnormal causality and the self-correction of the time series learning model, wherein the specific steps include: Obtain the abnormal time period after the warning event is triggered, backtrack the input sequence corresponding to the abnormal time period, and apply the feature matrix to the corresponding input sequence. To represent, where, The number of sampling points. For feature dimension, For the first The feature vector at time step; Based on the feature matrix Calculate the Pearson correlation coefficient between the feature vector of each dimension and the corresponding warning quantity, where the warning quantity is the predicted state quantity that triggers the warning event. The formula for calculating the Pearson correlation coefficient is as follows: , In the formula, Indicates the first Real-time warning volume, and All are sample means. The Pearson correlation coefficient has a range of values ​​of [value range missing]. The closer the absolute value is to 1, the stronger the correlation. Based on the absolute value of the Pearson correlation coefficient, the eigenvectors Classification: when It was identified as highly correlated with the anomaly at that time. It was identified as moderately correlated with the abnormality at that time. It was identified as having an abnormally weak correlation, among which , This is an empirical threshold; Filter out those that meet the requirements eigenvectors To form the set of dominant variables ; Introducing comprehensive risk indicators To quantify the combined influence of different eigenvectors in the set of dominant variables on abnormal states, wherein the combined risk index... Represented as: , In the formula, Representing the eigenvector Standardization bias, Let be the weighting coefficient, satisfying and ; When the comprehensive risk index Exceeding the preset risk threshold If all feature vectors in the set of dominant variables jointly drive the abnormal behavior of the electrical equipment, then the abnormal behavior is not driven. The standardized deviation The self-correction process of the time-series learning model is fed back, and the prediction deviation of the time-series learning model during abnormal periods is used to determine whether to trigger the adjustment of weight parameters, wherein the comprehensive risk index Exceeding the preset risk threshold Furthermore, the prediction deviation during the abnormal period exceeds the preset self-correction threshold. If the timing is right, the weight parameters of the time-series learning model will be adjusted; otherwise, the weight parameters of the time-series learning model will not be adjusted.

[0016] The present invention also provides an intelligent simulation and dynamic early warning system for multi-source state information of electrical equipment, including a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention integrates multi-source state information and adopts a dynamic inference mechanism based on cyclic prediction to perform recursive prediction and dynamic evolution in the time dimension, rather than single-point moment prediction. It can realize the dynamic simulation and trend prediction process of the operating state of electrical equipment over time, capture the gradual development trend of electrical equipment state in advance, and adapt to the dynamic changes of equipment state through adaptive early warning threshold. It combines the results of multi-step inference to make early warning judgment. Compared with single-point judgment with fixed threshold, it has achieved a significant improvement in the accuracy of equipment state perception and the timeliness of early warning.

[0018] (2) The present invention establishes an intelligent criterion and dynamic early warning threshold triggering mechanism, which can make forward predictions of potential faults of electrical equipment under multiple working conditions, rather than being limited to alarms after the fault occurs, thereby realizing adaptive risk identification and reducing false alarm and missed alarm rates.

[0019] (3) The present invention adjusts the weight parameters of the time-series learning model according to a set period, and updates the weight parameters by resuming training after breakpoints and incremental training. It can reuse historical training results, eliminates the need to train from scratch, and reduces computing power. It adapts to changes in the operating status of electrical equipment, realizes continuous model iteration, ensures the stability and adaptability of prediction, and ensures the accuracy and timeliness of prediction and inference. It can also avoid data loss caused by training interruption.

[0020] (4) The present invention continuously monitors the prediction deviation during abnormal periods through feedback information. When the prediction deviation meets the self-correction trigger condition, the model weight parameters are automatically adjusted to avoid the prediction deviation from accumulating over time and causing the model performance to deteriorate continuously. At the same time, when similar prediction deviation patterns occur repeatedly, it indicates that the time series data may have a distribution shift. At this time, the model weight parameters are also automatically adjusted to enable the model to quickly adapt to new data features, maintain prediction accuracy, and enhance stability and long-term operational reliability.

[0021] (5) The multidimensional causal analysis and correlation analysis steps of this invention can carry out source tracing analysis to identify abnormal causal relationships after triggering the early warning, and realize the self-correction of the time series learning model. It can locate the root cause of the abnormality and clarify the cause of the fault. It can also correct the model based on the actual abnormal situation of the early warning, reduce the model deviation, improve the subsequent inference and prediction effect, and reduce the problems of false alarms and missed alarms in the subsequent early warning. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0024] Example 1 This embodiment provides an intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment. By setting up a hierarchical structure of a data acquisition and processing layer, a dynamic simulation and analysis layer, and an early warning decision feedback layer, it achieves unified optimization of real-time modeling, trend prediction, and risk identification of the operating state of electrical equipment. This invention utilizes a time-series learning model for state time-series learning and multi-step rolling prediction, not only realizing dynamic evolution simulation and trend extrapolation of equipment states, but also achieving adaptive identification and source tracing of abnormal states by introducing intelligent threshold criteria and multi-dimensional causal analysis mechanisms. While ensuring simulation accuracy, this method possesses parameter self-learning and model closed-loop correction capabilities, continuously optimizing prediction performance based on newly sampled data, thereby improving the accuracy of electrical equipment operating state monitoring, the timeliness of early warning response, and the overall stability and intelligence level of the system.

[0025] Specifically, such as Figure 1 As shown, the method includes the following steps: S1. Electrical equipment data acquisition and processing.

[0026] This step is used to construct the basic dataset of electrical equipment operating status, which is a prerequisite for simulation and dynamic early warning. By deploying various types of intelligent sensors on key electrical equipment, multi-source status information reflecting the health status of the equipment is collected, including parameters such as voltage, current, temperature, vibration, partial discharge, and gas content in oil. The collected signals are stored using timestamps as indexes, forming a multi-dimensional time-series dataset. Due to differences in sampling frequency and response time among different sensors, the raw data often exhibits uneven sampling and temporal misalignment characteristics. Directly using this data for modeling will lead to reduced feature coupling and prediction bias.

[0027] To ensure data consistency, this invention first performs time alignment and anomaly removal on each measurement signal. The multidimensional time-series dataset is defined as follows: (1) in Indicates time of 3D feature vectors This represents the sampling length. It applies to the set of sampling timestamps for different channels. Through time interpolation function To achieve uniform resampling, all signals are aligned on the same time axis, resulting in a standardized multidimensional time-series dataset. The feature vector for each dimension of this standardized multidimensional time-series dataset is represented as follows: (2) In the formula, For at any time No. The feature vector after dimensional standardization; To remove outliers, a sliding window statistical method is used to calculate the mean and standard deviation. When any channel data... Deviation range: (3) Outliers are identified and either replaced or removed; otherwise, they are considered normal, resulting in a multidimensional time-series dataset with outliers removed. , These are the first two numbers in the sliding window. Mean and standard deviation in each dimension.

[0028] After data normalization, to address the differences in the dimensions of different features, normalization is performed to map each feature vector to the [0,1] interval. The normalization expression is: (4) in and These represent the functions for finding the maximum and minimum values, respectively. For the normalized result, that is, the result on the th... The feature vectors are normalized in terms of dimensions. The standardized multi-source time series dataset forms a standardized feature matrix, which serves as the input basis for the subsequent recurrent prediction model.

[0029] To address the issue of intermittent sampling gaps, this embodiment of the invention further employs a sliding window method to complete missing data and perform local smoothing on the feature vectors within the sliding window. Let the length of the sliding window be... Feature vectors within the sliding window that are free of missing data and not identified as outliers by the aforementioned outlier handling process are determined as valid feature vectors. Feature vectors with missing data or empty data points formed after outlier handling within the sliding window are determined as invalid feature vectors. Then, the number of valid feature vectors within the current sliding window is counted. And calculate the proportion of effective feature vectors. : (5) In the formula, The effective feature vector ratio threshold; when If the current sliding window is not filled, the sliding window is moved to the next position to continue processing. when At the same time, linear interpolation or spline completion is performed using adjacent effective feature vectors within the current sliding window to ensure a balance between time series continuity and data density. This process ultimately forms a continuous and reliable standardized input dataset, providing high-quality data support for subsequent dynamic inference and intelligent criterion establishment.

[0030] S2, a dynamic inference mechanism driven by cyclic prediction.

[0031] This step is the core of the invention, used to realize dynamic evolution simulation and multi-step trend prediction of the operating state of electrical equipment. By introducing a time-series learning model, the dependency relationship of multi-source state data in the time dimension is modeled, thereby obtaining the time-series feature representation and rolling deduction results of the operating state of electrical equipment. This time-series learning model can employ recurrent neural networks (RNNs), long short-term memory networks (LSTMs), gated recurrent units (GRUs), etc. In this embodiment, a recurrent neural network is used as an example to perform the following cyclic prediction and dynamic deduction: The normalized feature matrix obtained in step 1 Based on this, it is transformed into a sliding window input structure. Let the time window length be... Then the input sequence for each sliding window is defined as: (6) The output is the predicted state at a future time. ,in, Indicates time eigenvectors, The number of steps in advance for predicting an evolution.

[0032] The core structure of a recurrent neural network consists of an input layer, hidden layers, and an output layer. The network maintains parameter sharing over time, achieving the memorization and updating of historical information through recursive connections of hidden layer states. Its basic forward propagation expression is: (7) in, The hidden layer state vector; The weight matrix is ​​input to the hidden layer; The recursive weight matrix from the hidden layer to itself; This is the weight matrix from the hidden layer to the output layer; , These are the bias terms; For nonlinear activation functions (such as...) Or ReLU).

[0033] During operation, a multi-step rolling prediction strategy is employed to achieve dynamic extrapolation. The time-series learning model extrapolates future time steps through forward propagation. Operating status prediction output This can be used as input for the next step to continue the deduction and obtain the future time step. The predicted operating state output is used to continuously extrapolate and obtain a predicted operating state sequence for multiple future time steps that evolves over time. subscript This indicates the length of the predicted operating state sequence. This process is equivalent to a stepwise simulation of the state evolution trajectory of electrical equipment, which can reflect the current operating trend and quantify potential abnormal changes in advance.

[0034] Furthermore, this embodiment employs the Backpropagation Time (BPTT) algorithm, where the recurrent neural network adjusts its weight parameters in reverse based on the prediction error to minimize the overall loss function. The specific steps for adjusting the weight parameters include: Within a set period, obtain the predicted operating state output for future time steps and its corresponding actual observed operating state values, and calculate the prediction error between the two: (8) In the formula, For prediction error, These are the actual observed values ​​of the operating status and their corresponding predicted operating status outputs, respectively. The overall loss function for constructing a time-series learning model based on prediction error; The temporal learning model is expanded along the time dimension, and the temporal backpropagation algorithm is used to calculate the gradient of the overall loss function with respect to each weight parameter in the temporal learning model; The weight parameters are iteratively updated based on their gradients to reduce the overall loss function, and the updated time-series learning model is then used for further simulations.

[0035] To ensure the stability and adaptability of the predictions, this embodiment of the invention also incorporates a breakpoint-based training continuation mechanism. When a new batch of multi-source state information is collected and preprocessed to obtain newly standardized feature data, the optimal weight parameter set of the previously trained time-series learning model is automatically loaded. Incremental training is then performed based on this, without the need to re-initialize the model. This method ensures model convergence and stability while adapting to changes in the operating state of electrical equipment in real time, significantly improving the timeliness of predictions. Specifically, the incremental training steps include: Load the optimal weight parameter set saved from the previous training cycle. And use it as the initial weight parameter set for the current training cycle of the time-series learning model; Incremental training input sequences are constructed based on newly added standardized feature data and then input into the time-series learning model; Based on the incremental training input sequence, the weight parameter set of the time series learning model is adjusted in reverse using the weight parameter adjustment method of the time backpropagation algorithm mentioned above, so as to reduce the overall loss function of the time series learning model and obtain the updated weight parameter set. Save the updated set of weight parameters for subsequent runtime state simulation and breakpoint continuation training in the next training cycle.

[0036] S3, intelligent criterion triggering and dynamic early warning feedback.

[0037] This step is used to dynamically identify and provide early warnings of potential anomalies in electrical equipment during the prediction phase. By establishing an adaptive threshold criterion model, the cyclic prediction results are monitored in real time. When the predicted state deviates significantly from historical statistical characteristics, an early warning signal is automatically triggered, and feedback information is generated for subsequent causal analysis and model correction.

[0038] First, based on the standardized feature matrix obtained from multi-source state information collected during historical operation and preprocessed, a sliding window statistical model is established for its key feature vectors (such as temperature, current, and partial discharge amplitude). Let the sliding window length be... Within the interval, the mean and standard deviation of the eigenvectors are respectively and The adaptive warning threshold is defined as follows: (9) in, This is a threshold coefficient used to control the sensitivity of the early warning system, typically ranging from [1.5, 3.0]. It represents the predicted operating status at a future time step. satisfy If the condition is met, then that future time step is identified as an abnormal candidate time step; otherwise, the operating state of that future time step is determined to be within the normal fluctuation range.

[0039] If the feature vector has both vertical and horizontal offset features, symmetrical threshold intervals can be set on both sides. When the prediction result If all predicted state quantities exceed this range, then a future time step is also identified as an abnormal candidate time step; otherwise, the operating state of that future time step is also determined to be within the normal fluctuation range.

[0040] To avoid false alarms caused by short-term fluctuations, this embodiment of the invention introduces a stability constraint into the threshold criterion: Let the number of future time steps for continuous judgment be . , ,in, Indicates the first judgment in a continuous judgment sequence One future time step; When continuous Predicted operating state values ​​at future time steps All meet Only when the threshold is exceeded is the exception confirmed and an early warning event is automatically triggered; otherwise, it indicates that the threshold exceeding the limit has not been met, meaning that only some future time steps have not met the threshold or that the threshold is not met continuously. If none of the future time steps are satisfied, then when only some of the future time steps are not satisfied, it will be continuous. Anomaly candidate time steps within a given future time step are treated as transient disturbances and not processed. This design can significantly reduce the false alarm rate and improve the reliability and accuracy of early warnings and responses.

[0041] Upon triggering of an alert event, an alert record table containing multiple key pieces of information is immediately generated, including: the time of the anomaly. Predicted state variable name, deviation magnitude The system includes the feature dimension index corresponding to the abnormal predicted state quantity and the unique identifier of the electrical equipment. All early warning records are stored in chronological order and simultaneously pushed to the visualization interface or alarm management system for on-site personnel decision-making reference. Feedback information is also generated to statistically analyze the prediction deviation within abnormal periods. Prediction deviations within consecutive abnormal periods that meet the set self-correction trigger conditions (i.e., prediction deviations exceeding a preset self-correction threshold) are considered valid. When similar prediction bias patterns appear in multiple warning events, the learning rate or regularization parameter of the time series learning model is automatically adjusted, and the model weight parameters are updated based on the adjusted learning rate or regularization parameter, so that the time series learning model can gradually adapt to the new operating conditions, realize self-learning and dynamic convergence, and thus maintain the accuracy and robustness of the time series learning model in long-term operation.

[0042] Through the above method, this step realizes an automated closed-loop response from prediction results to early warning signals. It can not only identify potential risks of electrical equipment in advance, but also continuously optimize the accuracy of early warning during long-term operation, providing dynamic safety assurance for intelligent operation and maintenance of electrical equipment.

[0043] S4. Multidimensional causal and correlation analysis.

[0044] This step is used to perform multi-dimensional causal analysis and model self-correction for abnormal states after an early warning is triggered. The system identifies key influencing factors that lead to abnormal electrical equipment conditions by establishing an analytical model based on statistical correlation and feature contribution, achieving intelligent mapping and closed-loop optimization from "early warning result" to "abnormal source".

[0045] After the warning event is triggered in step 3, the input sequence corresponding to the abnormal period is first backtracked. Let the backtracking window contain... There are sampling points, and the feature dimension is... The corresponding feature matrix is ​​expressed as ,in For the first The feature vector at each time point. To analyze the correlation between each feature vector and the prediction target, the Pearson correlation coefficient between each feature channel and the warning quantity is calculated. The formula is as follows: (10) in, Indicates the first The feature vector of each sample This represents the warning quantity at the corresponding moment, i.e., the predicted state quantity that triggers the warning event. and All are sample means. The Pearson correlation coefficient has a range of values ​​of [value range missing]. The closer its absolute value is to The stronger the correlation, the better.

[0046] To enhance the readability and robustness of the analysis results, this embodiment of the invention classifies the feature vectors according to the absolute value of the correlation coefficient: when It was considered highly correlated at that time. It was considered moderately relevant at that time. The correlation was initially determined to be weak. Based on this correlation level, a set of possible dominant variables was selected. ,in , For empirical thresholds, preferably, Set it to 0.4. Set it to 0.2.

[0047] After identifying the set of dominant variables, a comprehensive risk indicator is further introduced. This is used to quantify the combined influence of different eigenvectors in the set of dominant variables on anomaly states. It is defined as follows: (11) in, Representing the eigenvector Standardization bias, For its weighting coefficients, satisfying and Weight It can be adaptively updated based on the sensitivity of features during the training phase or their relevance during the anomaly phase. When Exceeding the preset risk threshold When the condition is met, it is determined that all eigenvectors in the set of dominant variables jointly drive the abnormal behavior of the electrical equipment; otherwise, it indicates that no action is taken.

[0048] In addition, the standardized deviation will also be The feedback to the time-series learning model's self-correction process determines whether to trigger adjustments to the model's weight parameters. Specifically, this involves adjusting the weight parameters based on the comprehensive risk index. Exceeding the preset risk threshold Furthermore, the prediction deviation during the abnormal period exceeds the preset self-correction threshold. When the timing is right, the weight parameters of the time-series learning model are adjusted; otherwise, they are not. Through this feedback mechanism, the time-series learning model can more accurately capture the dynamic responses of key variables in subsequent runs, achieving self-learning and closed-loop optimization.

[0049] Based on the above results, a causal analysis report is automatically generated, which includes: a list of key influencing factors, correlation ranking, risk score, and recommended maintenance items.

[0050] Through this step, multidimensional causal and correlation analysis can not only reveal the root cause of electrical equipment anomalies, but also guide the continuous correction of model structure and parameters, making the entire simulation system interpretable, adaptive and stable in the long term.

[0051] Example 2 This embodiment provides an intelligent simulation and dynamic early warning system for multi-source state information of electrical equipment, including a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment as described in Embodiment 1 above.

[0052] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0057] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent simulation and dynamic early warning of multi-source state information of electrical equipment, characterized in that, Includes the following steps: Multi-source status information reflecting the operating status of electrical equipment is collected in real time to form a multi-dimensional time series dataset, which is then preprocessed to obtain a standardized feature matrix. Based on the standardized feature matrix, the time-series learning model uses a dynamic extrapolation mechanism driven by cyclic prediction to recursively predict and dynamically extrapolate the operating status of electrical equipment in the time dimension, thereby obtaining a prediction sequence of the operating status for multiple future time steps that evolve over time. Based on the predicted sequence of operating states for multiple future time steps that evolve over time, and using a set adaptive warning threshold, it is determined whether to automatically trigger a warning event and generate feedback information, thus completing the dynamic warning process.

2. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, The multi-source state information includes voltage, current, temperature, vibration, partial discharge, and gas content in the oil.

3. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, The preprocessing steps include: Let the multidimensional time series dataset be... , Indicates time of 3D feature vectors The sampling length; Based on the aforementioned multidimensional time-series dataset, for the sampling timestamp sets of different channels... Through time interpolation function To achieve uniform resampling and align all multidimensional time-series data on the same time axis, a standardized multidimensional time-series dataset is obtained. The feature vector for each dimension of this standardized multidimensional time-series dataset is represented as follows: ， In the formula, For at any time No. The feature vector after dimensional standardization; Based on the standardized multidimensional time-series dataset, the mean and standard deviation are calculated using the sliding window statistical method, and the range of each channel is constructed based on the mean and standard deviation. When the standardized feature vector in each dimension Deviating from the range If an outlier occurs, it is considered an anomaly and either replaced or removed; otherwise, it is considered normal, resulting in a multidimensional time-series dataset with outliers removed. , The first one in the sliding window Mean and standard deviation in each dimension; Based on the multidimensional time-series dataset with outliers removed, normalization processing is performed to obtain a standardized feature matrix, wherein the normalization expression is: ， In the formula, and These represent the functions for finding the maximum and minimum values, respectively. In the first The feature vector after dimensional normalization.

4. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 3, characterized in that, The preprocessing steps also include: after normalization, using the sliding window method to complete missing data and perform local smoothing, specifically including: Set the sliding window length to , the feature vectors that do not have missing data and are not judged as abnormal within the sliding window are determined as valid feature vectors, and the feature vectors that have missing data or are judged as abnormal and then removed after removing the outliers are determined as invalid feature vectors. Count the number of valid feature vectors within the current sliding window. And calculate the proportion of effective feature vectors. : ， when If missing data is not filled in at the current sliding window, the sliding window is moved to the next position to continue processing. The effective feature vector ratio threshold; when When missing data is filled in, linear interpolation or spline interpolation is performed using the effective feature vectors adjacent to the missing position within the current sliding window.

5. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, The step of obtaining the predicted sequence of operating states for multiple future time steps that evolve over time includes: Let the standardized feature matrix be... This is transformed into a sliding window input structure, yielding the input sequence for each sliding window. , For a moment eigenvectors, The length of the sliding window; Based on the input sequence, the time-series learning model extrapolates future time steps through forward propagation. The output of the predicted running status, where The number of steps in advance for predicting an evolution; The future time step The predicted operating state output is used as the input to the time series learning model to continue the extrapolation and obtain the future time steps. The predicted operating state output is used to continuously extrapolate and obtain the predicted operating state sequence for multiple future time steps that evolve over time. subscript This indicates the length of the predicted sequence for the running status.

6. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, Within a set period, the weight parameters of the time-series learning model are adjusted backward using the time backpropagation algorithm to reduce the overall loss function. Specific steps include: The weight parameter adjustment steps of the time-series learning model include: Within a set period, obtain the predicted output of the running state at future time steps and the corresponding actual observed value of the running state, and calculate the prediction error between the two. The overall loss function of the time-series learning model is constructed based on the prediction error; The time-series learning model is expanded along the time dimension, and the gradient of the overall loss function with respect to each weight parameter in the time-series learning model is calculated using the time backpropagation algorithm. The weight parameters are iteratively updated based on their gradients to reduce the overall loss function, and the updated time-series learning model is used for subsequent continuous deduction.

7. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, Also includes: When a new batch of multi-source state information is collected and preprocessed to obtain newly standardized feature data, the optimal weight parameter set of the previously trained temporal learning model is acquired. Incremental training is performed using a breakpoint-based continuation mechanism, and the specific incremental training includes: Load the optimal weight parameter set saved from the previous training cycle. And use it as the initial weight parameter set for the current training cycle of the time-series learning model; An incremental training input sequence is constructed based on the newly added standardized feature data and then input into the time-series learning model; Based on the incremental training input sequence, the weight parameter set of the time series learning model is adjusted in reverse using the time backpropagation algorithm to reduce the overall loss function of the time series learning model and obtain the updated weight parameter set. The updated set of weight parameters is saved for subsequent runtime state deduction and breakpoint continuation training in the next training cycle.

8. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, The step of determining whether to automatically trigger a warning signal and generate feedback information includes: Based on multi-source state information collected during historical operation, a standardized feature matrix is ​​obtained after preprocessing. A sliding window length is then established for each feature vector in the standardized feature matrix. For the sliding window statistical model, the mean and standard deviation of each feature vector within the sliding window are calculated, and an adaptive warning threshold is set based on the mean and standard deviation, expressed as: ， In the formula, For adaptive early warning thresholds, and These are the mean and standard deviation of the state variables, respectively. This is the threshold coefficient, used to control the sensitivity of the early warning. Predicting the running state at a future time step Does it meet the requirements? If yes, then the future time step is identified as an abnormal candidate time step; otherwise, the operating state of the future time step is determined to be within the normal fluctuation range. Predicting the running state at a future time step Are all predicted state variables outside the set two-sided threshold range? If yes, then the future time step is identified as an abnormal candidate time step; otherwise, the operating state of the future time step is determined to be within the normal fluctuation range. Introduce stability constraints for anomaly detection: Let the number of future time steps for continuous judgment be . , ,in, Indicates the first judgment in the continuous judgment sequence One future time step; When continuous Predicted operating state values ​​at future time steps All meet If the condition is met, an anomaly is confirmed and an early warning event is automatically triggered; otherwise, it indicates that the continuous condition has not been met. The continuous violation of the limit condition at each future time step will be continuous Anomaly candidate time steps within a given future time step are determined to be transient disturbances and do not trigger warning events; After an automatic warning event is triggered, feedback information and a warning record table are generated, wherein the warning record table includes: the time of the anomaly. Predicted state quantity name, deviation magnitude The feedback information includes the feature dimension index corresponding to the abnormal prediction state quantity and the unique identifier of the electrical equipment. This feedback information is used to statistically analyze the prediction deviation within the abnormal period. When the prediction deviation within consecutive abnormal periods exceeds a preset self-correction threshold... If a similar prediction bias pattern appears in multiple warning events, the learning rate or regularization parameter of the time series learning model will be automatically adjusted, and the weight parameters of the time series learning model will be updated based on the adjusted learning rate or regularization parameter.

9. The intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment according to claim 1, characterized in that, Also includes: After an early warning event is triggered, a source tracing analysis is performed to identify the abnormal causal relationship and the self-correction of the time-series learning model. Specific steps include: Obtain the abnormal time period after the warning event is triggered, backtrack the input sequence corresponding to the abnormal time period, and apply the feature matrix to the corresponding input sequence. To represent, where, The number of sampling points. For feature dimension, For the first The feature vector at time step; Based on the feature matrix Calculate the Pearson correlation coefficient between the feature vector of each dimension and the corresponding warning quantity, where the warning quantity is the predicted state quantity that triggers the warning event. The formula for calculating the Pearson correlation coefficient is as follows: ， In the formula, Indicates the first Real-time warning volume, and All are sample means. The Pearson correlation coefficient has a range of values ​​of [value range missing]. The closer the absolute value is to 1, the stronger the correlation. Based on the absolute value of the Pearson correlation coefficient, the eigenvectors Classification: when It was identified as highly correlated with the anomaly at that time. It was identified as moderately correlated with the abnormality at that time. It was identified as having an abnormally weak correlation, among which , This is an empirical threshold; Filter out those that meet the requirements eigenvectors To form the set of dominant variables ; Introducing comprehensive risk indicators To quantify the combined influence of different eigenvectors in the set of dominant variables on abnormal states, wherein the combined risk index... Represented as: ， In the formula, Representing the eigenvector Standardization bias, Let be the weighting coefficient, satisfying and ; When the comprehensive risk index Exceeding the preset risk threshold If all feature vectors in the set of dominant variables jointly drive the abnormal behavior of the electrical equipment, then the abnormal behavior is not driven. The standardized deviation The self-correction process of the time-series learning model is fed back, and the prediction deviation of the time-series learning model during abnormal periods is used to determine whether to trigger the adjustment of weight parameters, wherein the comprehensive risk index Exceeding the preset risk threshold Furthermore, the prediction deviation during the abnormal period exceeds the preset self-correction threshold. If the timing is right, the weight parameters of the time-series learning model will be adjusted; otherwise, the weight parameters of the time-series learning model will not be adjusted.

10. An intelligent simulation and dynamic early warning system for multi-source state information of electrical equipment, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the intelligent simulation and dynamic early warning method for multi-source state information of electrical equipment as described in any one of claims 1-9.

Images