Early warning method and system for abnormity of power supply system

Abstract

The invention relates to the field of intelligent early warning of guaranteed power supply, and provides an abnormal early warning method and system for a guaranteed power supply system, and the method comprises the steps: obtaining the topological structure, real-time operation data and equipment state data of each node in a guaranteed power supply network, employing the conditional entropy and instantaneous causal entropy calculation technology, precisely evaluating the causal intensity between variables, and achieving the early warning of the abnormal state of the guaranteed power supply system. And a causal adjacency matrix dynamically changing along with time is constructed to reflect a complex causal relationship in the network. Then, in combination with an equipment topological structure and dynamic causal information, multi-scale time sequence features are extracted through feature updating and expansion causal convolution, and a comprehensive feature vector is formed; after the vector is input into the Bayesian neural network, a plurality of prediction value samples are generated through sampling weight distribution, a mean value is used as a prediction result, and the prediction uncertainty is measured through variance. Accurate prediction and timely early warning of potential anomalies of the power supply system are realized, and the stability and safety of the system are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent early warning for power supply guarantee, and in particular to an early warning method and system for abnormal power supply guarantee system. Background Art

[0002] The contents in this section merely provide background information related to the present invention and may not constitute prior art.

[0003] In the power supply system, the stable operation of the power supply system is crucial, and its abnormal warning technology has always been a hot topic of research. The existing abnormal warning methods of the power supply system mainly rely on real-time monitoring of each node of the power supply network, and evaluate the operating status of the system by collecting the topological structure, real-time operation data and equipment status data of the nodes. However, these methods have some obvious defects when processing complex and changeable power system data.

[0004] For example, a Chinese patent with authorization announcement number CN117172522A discloses an intelligent early warning method and system for power grid risk, including: marking the abnormal site according to the determined abnormal site, and after marking, analyzing whether the adjacent sites are abnormal. If there are several abnormal sites connected to each other, the digital marks of the corresponding sites are changed; subsequently, a layout diagram generation unit is used to generate a layout diagram for display, and then the area analysis unit is used to analyze the monitored area to see whether it belongs to a risk area, and the risk area data packet of the risk area is transmitted to the display terminal for display, so that the operator can quickly understand the specific situation without looking up the corresponding parameter data, thereby further improving the management efficiency of the operator and avoiding power supply risks in the power grid.

[0005] However, the warning model of the above-mentioned warning method is relatively simple, mainly based on basic machine learning algorithms, and lacks adaptability to complex network structures and dynamic changes. It ignores the time delay and causal relationship between data, and only relies on simple data comparison or statistical analysis to make abnormal judgments. When processing power system data with strong time correlation and causal relationship, this method often cannot accurately capture the inherent connection and potential laws between data, resulting in insufficient accuracy and reliability of the warning results. In addition, the existing methods have low computational efficiency when processing large-scale data, making it difficult to achieve real-time and efficient warning of power systems. Summary of the invention

[0006] In order to solve the above technical problems, the purpose of the present invention is to provide an early warning method and system for abnormal power supply system, which calculates the instantaneous causal entropy value and uses a Bayesian neural network to effectively predict abnormal conditions of the power supply system and issue early warnings in time, thereby helping to ensure the stable operation of the power supply system.

[0007] The object of the present invention is achieved by the following technical solutions:

[0008] In a first aspect, the present invention provides a method for warning of abnormal power supply guarantee system, including:

[0009] Obtain the topological structure, real-time operation data and equipment status data of each node in the power supply guarantee network;

[0010] Based on the topological structure, real-time operation data and equipment status data, after determining the time delay window and causal lag coefficient, calculate the conditional entropy between each pair of variables; by performing a time difference process on all conditional entropies, calculate the instantaneous causal entropy value between each pair of variables;

[0011] According to the instantaneous causal entropy value, calculate the causal strength between variable pairs; calculate the instantaneous causal strength between each node through a sliding window and construct a dynamic causal adjacency matrix to reflect the causal relationship of the network changing over time;

[0012] Combined with the equipment topological structure and the dynamic causal adjacency matrix, update and represent the node features by layer-by-layer transmission and transformation of node features; use dilated causal convolution to extract multi-scale time series features; fuse the updated node features and multi-scale time series features through weighted summation to obtain a comprehensive feature vector;

[0013] Take the comprehensive feature vector as the input of the Bayesian neural network. The Bayesian neural network obtains multiple weight samples by sampling the weight distribution; perform forward propagation calculation on each sample to obtain multiple prediction value samples; the mean of the multiple prediction value samples can be used as the final prediction result, and the variance of the multiple prediction value samples measures the uncertainty of the prediction;

[0014] Compare the mean and variance of the prediction value samples with the corresponding preset thresholds respectively. If the comparison result does not exceed the error range, it is determined that the prediction is accurate; if the comparison result exceeds the error range, an abnormal warning is triggered.

[0015] Further, after obtaining the topological structure, real-time operation data and equipment status data of each node in the power supply guarantee network, it further includes:

[0016] Perform standardization processing on the topological structure, and the standardization processing includes unifying the node identification format, correcting the error information in the topological connection relationship, and removing redundant topological connection information.

[0017] Further, after obtaining the topological structure, real-time operation data and equipment status data of each node in the power supply guarantee network, it further includes:

[0018] Use mean filtering to remove abnormal mutation data points in the real-time operation data, and fill in the missing data values in the real-time operation data by the median method.

[0019] Further, after obtaining the topological structure, real-time operation data, and device status data of each node in the power supply guarantee network, it further includes:

[0020] Classify and sort out the device status data, divide the devices into different status levels according to factors such as device type, operation years, maintenance records, etc., assign corresponding weight coefficients to each level, and establish an association index between the device status data and the topological structure and real-time operation data.

[0021] Further, the steps of establishing an association index between the device status data and the topological structure and real-time operation data specifically include:

[0022] Preset a composite index structure, which includes three basic dimensions: device ID, node ID, and timestamp, as well as additional dimensions of device status level and real-time operation data category to achieve multi-dimensional fast retrieval;

[0023] For each device, according to its status level and category, assign a unique status code to it in the composite index structure, and the status code can reflect information such as the health status and maintenance priority of the device;

[0024] When the real-time operation data is collected, automatically record the corresponding device ID, node ID, and timestamp, and associate them with the status code to form data entries and store them in the index database;

[0025] When it is necessary to analyze the status changes of a specific device or node, quickly retrieve the relevant real-time operation data and device status information through conditions such as device ID, node ID, or time range.

[0026] Further, the steps of calculating the instantaneous causal strength between each node through a sliding window and constructing a dynamic causal adjacency matrix specifically include:

[0027] Define the size and step length of the sliding window;

[0028] Within each sliding window, based on the calculated instantaneous causal entropy value, use a weighted method to calculate the instantaneous causal strength between variable pairs;

[0029] When constructing the dynamic causal adjacency matrix, each element of the matrix represents the instantaneous causal strength of the corresponding node pair within the current window. As the window slides, the matrix is updated in real time to dynamically reflect the changes in the causal relationship in the network;

[0030] Introduce a time decay factor to perform a weighted sum of the causal strengths in the historical window to reduce the impact of old data on the current state assessment and enable the dynamic causal adjacency matrix to reflect the recent network state changes.

[0031] Further, after triggering the anomaly warning, it further includes:

[0032] According to the degree of deviation of the mean value of the predicted value sample from the corresponding threshold and the magnitude of the variance, the anomalies are divided into three levels: minor anomalies, general anomalies, and severe anomalies;

[0033] If the anomaly is a minor anomaly, a warning message is sent to the operation and maintenance personnel to prompt them to pay attention to the running status of the corresponding nodes and devices;

[0034] For general anomalies, in addition to sending warning messages, detailed data collection and diagnostic procedures for the relevant nodes and their surrounding associated nodes are automatically started. At the same time, some parameter settings of the Bayesian neural network are adjusted to improve the recognition accuracy of abnormal situations, and the diagnostic results and the adjusted network parameters are fed back to the operation and maintenance personnel to assist them in troubleshooting and decision-making;

[0035] For severe anomalies, an emergency warning signal is immediately triggered, the power supply of the relevant nodes is suspended, and the abnormal data and relevant information are transmitted to the remote expert diagnosis system in real time.

[0036] In a second aspect, the present invention provides an early warning system for power supply protection system anomalies, including:

[0037] A data acquisition module for acquiring the topological structure, real-time operation data, and device status data of each node of the power supply protection network;

[0038] An instantaneous causal entropy calculation module for calculating the conditional entropy between each pair of variables based on the topological structure, real-time operation data, and device status data after determining the time delay window and causal lag coefficient; by performing temporal difference processing on all conditional entropies, calculating the instantaneous causal entropy value between each pair of variables;

[0039] A dynamic causal adjacency matrix construction module for calculating the causal strength between variable pairs according to the instantaneous causal entropy value; calculating the instantaneous causal strength between each node through a sliding window and constructing a dynamic causal adjacency matrix to reflect the causal relationship of the network changing over time;

[0040] A comprehensive feature vector module for updating and representing node features by layer-by-layer passing and transforming node features in combination with the device topological structure and the dynamic causal adjacency matrix; extracting multi-scale time series features using dilated causal convolution; fusing the updated node features and multi-scale time series features through weighted summation to obtain a comprehensive feature vector;

[0041] The predicted value sample module is used to take the comprehensive feature vector as the input of the Bayesian neural network. The Bayesian neural network obtains multiple weight samples by sampling the weight distribution, performs forward propagation calculations on each sample to obtain multiple predicted value samples, takes the mean of the multiple predicted value samples as the final prediction result, and measures the uncertainty of the prediction with the variance of the multiple predicted value samples.

[0042] The determination module is used to compare the mean and variance of the predicted value samples with the corresponding preset thresholds respectively. If the comparison results are within the error range, it is determined that the prediction is accurate; if the comparison results exceed the error range, an abnormal warning is triggered.

[0043] In a third aspect, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps corresponding to the method in the first aspect are implemented.

[0044] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps corresponding to the method in the first aspect are implemented.

[0045] In summary, the technical solutions of the embodiments of the present invention have at least the following advantages and beneficial effects:

[0046] Based on the topological structure, real-time operation data, and device status data of each node in the power supply guarantee network, after determining the time delay window and causal lag coefficient, the conditional entropy between variable pairs is calculated, and then the instantaneous causal entropy value is obtained through time difference processing. According to this, the causal strength is calculated, and the instantaneous causal strength between nodes is calculated through a sliding window and a dynamic causal adjacency matrix is constructed to reflect the dynamic changes of the network causal relationship. Subsequently, in combination with the device topological structure and the dynamic causal adjacency matrix, the node features are updated by layer-by-layer transmission and transformation, and the dilated causal convolution is used to extract multi-scale time series features. The updated node features and multi-scale time series features are weighted and summed and fused to obtain a comprehensive feature vector. This vector is used as the input of the Bayesian neural network. The network samples the weight distribution to obtain multiple predicted value samples, takes the mean of them as the final prediction result, and uses the variance to measure the prediction uncertainty. Finally, the mean and variance of the predicted value samples are respectively compared with the preset thresholds, and according to the comparison results, it is determined whether the prediction is accurate and the corresponding abnormal warning is triggered. It realizes the accurate analysis of the network causal relationship and node features, effectively predicts the abnormal situation of the power supply guarantee system with the help of the Bayesian neural network, issues warnings in a timely manner, and helps to ensure the stable operation of the power supply system. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of a method for warning of abnormal situations in a power supply guarantee system provided by the present invention;

[0048] Figure 2 Schematic diagram of an early warning system for abnormalities in the power supply guarantee system provided by the present invention;

[0049] Figure 3 Schematic diagram of an electronic device provided by the present invention. Detailed implementation manners

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0051] As Figure 1 shown, an early warning method for abnormalities in the power supply guarantee system proposed in the embodiments of the present invention includes:

[0052] S1. Obtain the topological structure, real-time operation data, and device status data of each node in the power supply guarantee network.

[0053] Among them, the topological structure refers to the connection method and layout form between each node in the power supply guarantee network, which can clearly show the architecture framework of the entire power supply guarantee network and provide a basic basis for subsequent analysis of the mutual relationship between each node. The real-time operation data covers the specific values of key electrical parameters such as voltage, current, and power at the current moment. These data can reflect the operation status of the network in real time and are important indicators for judging whether the system is operating normally. The device status data includes information such as the operating temperature, pressure, and vibration of the device. By collecting these data, the health status of the device can be understood, and potential device failure hazards can be discovered in a timely manner.

[0054] The principle of obtaining these data is that ensuring the stable operation of the power supply network depends on the coordinated cooperation between each node and the normal operation of the device, and the topological structure, real-time operation data, and device status data comprehensively describe the operation of the network from different perspectives. Accurately obtaining these data can provide a reliable basis for subsequent analysis and early warning, enabling the entire early warning method to be based on real and comprehensive network information, thereby improving the accuracy and effectiveness of the early warning. For example, when the voltage of a certain node shows abnormal fluctuations, this change can be captured in a timely manner through the real-time operation data. Combining the topological structure and device status data, the impact of this voltage fluctuation on the entire network and the possible device failures can be further analyzed, laying a solid foundation for the subsequent early warning steps and ensuring the safe and stable operation of the power supply guarantee system.

[0055] In addition, after obtaining the topological structure, real-time operation data, and device status data of each node in the power supply guarantee network, the topological structure is standardized. The standardization process includes unifying the node identification format, correcting the error information in the topological connection relationship, and removing redundant topological connection information. Unifying the node identification format can ensure that each node has a unique identifier in the system, avoiding data processing errors caused by identifier confusion. Correcting the error information in the topological connection relationship is to ensure the accuracy of network connections, because any connection error may affect the judgment of the network operation status. Removing redundant topological connection information can simplify the network structure and improve the efficiency of data processing.

[0056] In addition, after obtaining the topological structure, real-time operation data, and device status data of each node in the power supply guarantee network, abnormal mutation data points in the real-time operation data are removed using mean filtering, and missing data values in the real-time operation data are filled by the median method. The device status data is classified and sorted. Devices are divided into different status levels according to factors such as device type, operation years, and maintenance records, and corresponding weight coefficients are assigned to each level. At the same time, an association index is established between the device status data, the topological structure, and the real-time operation data.

[0057] Among them, the steps of establishing the association index between the device status data, the topological structure, and the real-time operation data specifically include:

[0058] Preset a composite index structure. The composite index structure includes three basic dimensions: device ID, node ID, and timestamp, as well as additional dimensions such as device status level and real-time operation data category to achieve multi-dimensional fast retrieval;

[0059] For each device, according to its status level and category, assign a unique status code to it in the composite index structure. The status code can reflect information such as the health status and maintenance priority of the device;

[0060] When the real-time operation data is collected, the corresponding device ID, node ID, and timestamp are automatically recorded and associated with the status code, and data entries are formed and stored in the index database;

[0061] When analyzing the status changes of specific devices or nodes, relevant real-time operation data and device status information can be quickly retrieved through conditions such as device ID, node ID, or time range.

[0062] S2. Based on the topological structure, real-time operation data, and device status data, after determining the time delay window and causal lag coefficient, calculate the conditional entropy between each pair of variables; by performing temporal difference processing on all conditional entropies, calculate the instantaneous causal entropy value between each pair of variables.

[0063] Specifically, based on the topological structure, real-time operation data, and device status data obtained in the early stage, after determining the time delay window and causal lag coefficient, the conditional entropy is calculated to measure the dependence degree between variables, and further processed to obtain the instantaneous causal entropy value. The setting of the time delay window takes into account the dynamic characteristics of the power supply network data. Determining an appropriate window range can capture the relationship changes between variables in a short time, rather than resulting in lagged results due to too long a time. The causal lag coefficient is preset based on the actual operation experience and data analysis of the power supply system. This coefficient reflects the delay time required for one variable to affect another variable.

[0064] In specific operations, for each pair of variables, the conditional entropy is used to quantify their mutual dependence. Conditional entropy is a concept in information theory. Intuitively, it measures the degree to which the uncertainty of another variable is reduced given a certain variable. After calculating the conditional entropy for all variable pairs, a time difference operation is performed. This process is similar to analyzing the instantaneous change rate of variables over time, thereby obtaining the instantaneous causal entropy value of each variable pair in the time dimension. This instantaneous causal entropy value can better reflect the dynamic change characteristics of the causal relationship between variables at different moments, rather than the static dependence degree, providing more refined time dimension information for subsequent analysis. Thus, the instantaneous causal relationships between different variables (such as electrical parameters like voltage, current, power, and device status data) in the power supply system can be captured. For example, when the operating temperature of a certain device is abnormal (a dynamic variable), the instantaneous causal relationship strength between this temperature change and relevant electrical parameters (such as current and voltage) can be quickly quantified through conditional entropy and difference processing, providing an accurate basis for calculating the causal strength for subsequent steps. This enables the entire early warning model to accurately judge the system status based on the dynamically changing causal relationship, early warning of potential risks in the power supply system, and enhancing the real-time response ability and accuracy of the early warning system.

[0065] Among them, the calculation of conditional entropy is as follows:

[0066] Assuming the variable pair is (Y, X), the conditional entropy H(Y∣X) can be expressed as:

[0067] H(Y∣X)=-∑ x∈X ∑ y∈Y p(y∣x)logp(y∣x) (1)

[0068] Among them, the variable pair Y and X respectively represent different monitoring variables in the power supply network; p(y∣x) is the conditional probability that Y = y given X = x.

[0069] The calculation of the instantaneous causal entropy value is as follows:

[0070] Assume that the time delay window is Δt and the causal lag coefficient is τ, then the instantaneous causal entropy value I(X→Y) can be expressed as:

[0071] I(X→Y) = H(Y t ∣Y t-τ ) - H(Y t ∣Y t-τ , X t-τ ) (2)

[0072] Among them, H(Y t ∣Y t-τ ) is the conditional entropy of Y at time t, and H(Y t ∣Y t-τ , X t-τ ) is the conditional entropy of Y considering X.

[0073] S3. Calculate the causal strength between variable pairs according to the instantaneous causal entropy value; calculate the instantaneous causal strength between each node through a sliding window and construct a dynamic causal adjacency matrix to reflect the causal relationship of the network changing over time.

[0074] Specifically, first define the size and step length of the sliding window, so as to effectively divide the data in the time dimension, so that the data characteristics within a certain time range can be captured within each window. At the same time, through the setting of the step length, ensure the connection between windows and the comprehensive coverage of the data.

[0075] Next, within each sliding window, based on the previously calculated instantaneous causal entropy value, use a weighted method to calculate the instantaneous causal strength between variable pairs. The weighted method can fully consider the differences in the importance of different factors in the causal relationship, making the calculated causal strength more accurately reflect the actual influence between variables.

[0076] Subsequently, construct a dynamic causal adjacency matrix. Each element of the matrix represents the instantaneous causal strength of the corresponding node pair within the current window. As the window slides continuously, the matrix can be updated in real time, thus dynamically reflecting the change of the causal relationship in the network. In addition, a time decay factor is introduced to perform a weighted sum of the causal strengths in the historical windows. The purpose of this is to reduce the influence of old data on the current state evaluation. Because as time goes by, the influence of past data on the current network state will gradually weaken. Through the weighted processing of the time decay factor, the dynamic causal adjacency matrix can more accurately reflect the recent changes in the network state, providing causal relationship information that is more in line with the actual situation for subsequent analysis.

[0077] Among them, the causal strength C calculated by the sliding window ij is specifically as follows:

[0078]

[0079] where \(w\) is the sliding window size, and \(\alpha\) t is the time decay factor, is the instantaneous causal entropy value at time \(t\)

[0080] The expression of the dynamic causal adjacency matrix is as follows:

[0081]

[0082] S4. Combine the device topology structure and the dynamic causal adjacency matrix, and update and represent the node features by layer-by-layer passing and transforming the node features; use dilated causal convolution to extract multi-scale temporal features; fuse the updated node features and the multi-scale temporal features through weighted summation to obtain a comprehensive feature vector.

[0083] First, by combining the device topology structure and the dynamic causal adjacency matrix, a deep exploration of the updated representation of node features is initiated. Specifically, the process of layer-by-layer passing and transforming node features is like constructing a precise information transmission chain for node features. In this process, the initial feature information of the node itself continuously incorporates relevant information from adjacent nodes and the overall network structure during each layer of passing and transformation, thereby realizing the update and enrichment of features. This layer-by-layer updated representation enables node features to not only be limited to their single attributes but also comprehensively absorb the associated information in the network, making node features more representative and comprehensive, laying a solid foundation for subsequent analysis.

[0084] The application of dilated causal convolution provides support for extracting multi-scale temporal features. Dilated causal convolution performs flexible dilation operations on time series, thereby capturing feature information at different time scales. From short-term instantaneous changes to long-term trend evolutions, dilated causal convolution can accurately extract them and transform them into meaningful temporal features.

[0085] Finally, fuse the updated node features and the multi-scale temporal features through weighted summation to obtain a comprehensive feature vector. This fusion process is like a carefully planned resource integration. By means of a reasonable weighting method, considering the importance of different features and their influence on the system state, the node features and temporal features are integrated. The obtained comprehensive feature vector not only contains information about the node's spatial structure but also incorporates its dynamic change features in the time dimension, thus being able to more comprehensively and accurately depict the operation state of the node and the entire power supply protection network. This comprehensive feature vector, as the input of the Bayesian neural network, provides a richer and more accurate information basis for subsequent prediction and early warning, enabling the entire early warning method to more effectively identify potential anomalies and improve the accuracy and reliability of the power supply protection system's anomaly early warning.

[0086] Among them, the specific calculation of layer-by-layer transmission and transformation of node features is as follows:

[0087] Assume that the node feature is h i , the topological structure is A, and the dynamic causal adjacency matrix is A'. Then the updated node feature h' i can be expressed as:

[0088] h' i = σ(A·h i + A'·h i ) (5)

[0089] Among them, σ is the activation function, such as ReLU.

[0090] The specific calculation of dilated causal convolution is as follows:

[0091] Assume that the input sequence is x t , the dilation factor is d, and the convolution kernel is θ. Then the dilated causal convolution is expressed as:

[0092]

[0093] Among them, K is the size of the convolution kernel.

[0094] The specific calculation of the comprehensive feature vector is as follows:

[0095] The updated node feature is h' i , the multi-scale temporal feature is y i , then the comprehensive feature vector z i is:

[0096] z i = λ·h' i + (1 - λ)·y i (7)

[0097] Among them, λ is the weighting coefficient.

[0098] S5. Use the comprehensive feature vector as the input of the Bayesian neural network. The Bayesian neural network obtains multiple weight samples by sampling the weight distribution; performs forward propagation calculations for each sample to obtain multiple prediction value samples; the mean of the multiple prediction value samples can be used as the final prediction result, and the variance of the multiple prediction value samples measures the uncertainty of the prediction.

[0099] Specifically, first use the comprehensive feature vector generated in the previous steps as the input of the Bayesian neural network. Among them, the comprehensive feature vector not only contains the information of the node in the spatial structure, but also integrates its dynamic change characteristics in the time dimension, so as to comprehensively and accurately characterize the operation state of the node and the entire power supply guarantee network.

[0100] By sampling the weight distribution, a Bayesian neural network obtains multiple weight samples, thereby quantifying the uncertainty of model parameters. Specifically, a Bayesian neural network represents the weight parameters of the network as a probability distribution rather than a deterministic point value, enabling it to better handle noise and outliers when processing data and achieve higher robustness.

[0101] Next, forward propagation calculations are performed on each weight sample to obtain multiple prediction value samples. This process is similar to the inference process of a traditional neural network, but the difference is that a Bayesian neural network can generate multiple prediction value samples through the forward propagation of multiple weight samples, thereby capturing the uncertainty of model predictions.

[0102] Finally, the mean of multiple prediction value samples is used as the final prediction result, and the variance of multiple prediction value samples is used to measure the uncertainty of the prediction. This prediction method based on probability distribution can not only provide a point estimate of the prediction result but also quantify the uncertainty of the prediction result, providing a more reliable basis for subsequent anomaly warnings.

[0103] That is to say, the mean of the prediction value samples can reflect the expected operating state of the power supply guarantee system at a certain future moment, while the variance of the prediction value samples can measure the uncertainty of this expectation. If the variance is large, it indicates that there is a large uncertainty in the model's prediction of the future, which may be caused by noise in the data, imperfections in the model, or the complexity of the system itself.

[0104] Among them, the mean μ of multiple prediction value samples pred has the following expression:

[0105]

[0106] where z * represents the new input data, that is, the input feature vector to be predicted in the prediction stage; f(z * ; θ (m) ) represents the prediction output obtained by substituting z * into the neural network and performing forward propagation using the m-th parameter sample θ (m) , and M represents that there are M different parameter configurations θ (m) .

[0107] The expression for the variance of multiple prediction value samples is

[0108]

[0109] where (f(z * ; θ (m) ) - μ pred) is the difference between the predicted output and the predicted mean under the m-th parameter sample; (f(z * ; θ (m) ) - μ pred ) T is the transpose of the difference between the predicted output and the predicted mean; is the noise variance in the likelihood function; is the additional covariance caused by the observation noise variance in the likelihood function. I is the identity matrix, ensuring that the noise is independent in each prediction dimension.

[0110] S6. Compare the mean and variance of the predicted value samples with the corresponding preset thresholds respectively. If the comparison result does not exceed the error range, it is determined that the prediction is accurate; if the comparison result exceeds the error range, an anomaly warning is triggered.

[0111] Specifically, the comparison process is to detect whether the current prediction result exceeds the boundary of the normal operation of the system. If the comparison result does not exceed the error range, it means that the predicted value of the system is within the normal fluctuation range, and the system operation is relatively stable. At this time, it is determined that the prediction is accurate, and no special treatment measures are required to ensure that the power supply system can continue to operate according to the established mode.

[0112] If the comparison result shows that the mean or variance exceeds the error range, an anomaly warning will be triggered. This step can timely capture potential abnormal situations that may occur in the system. Due to the complexity of the power supply protection system, factors such as noise in the data, imperfections in the model, or changes in the system itself may lead to prediction uncertainties. Therefore, by comparing with the preset thresholds, abnormal situations that need further attention can be effectively screened out, enabling operation and maintenance personnel to focus their energy on nodes and devices that may have problems, improving the pertinence of early warning and operation and maintenance efficiency.

[0113] After the anomaly warning is triggered, it is also necessary to further refine the anomaly into three levels: minor anomaly, general anomaly, and serious anomaly according to the degree of deviation of the mean of the predicted value sample from the corresponding threshold and the magnitude of the variance. This hierarchical processing method is based on the quantitative assessment of the severity of the abnormal situation and can provide a basis for subsequent different countermeasures.

[0114] When the anomaly is a minor anomaly, the system will send a warning message to the operation and maintenance personnel. This measure is to let the operation and maintenance personnel know the preliminary signs of the problem in advance, and they can timely arrange low-level intervention means such as inspections and monitoring to pay attention to and preliminarily investigate the operating status of relevant nodes and devices, such as viewing the subtle change trend of device parameters through remote monitoring and arranging on-site personnel for simple inspections, etc., to avoid the further development of the problem, play a role of prevention first, and at the same time reduce the interference to the normal operation of the system.

[0115] For general anomalies, in addition to sending early warning messages, detailed data collection and diagnostic procedures for relevant nodes and their surrounding associated nodes will be automatically initiated. Because general anomalies indicate that problems may have begun to affect the local stability of the system and more in-depth data support is needed for analysis and diagnosis. At the same time, some parameter settings of the Bayesian neural network are adjusted to improve the model's recognition accuracy for abnormal situations, making subsequent predictions and analyses more accurate and reliable. The adjusted network parameter settings and detailed diagnostic results will be fed back to the operation and maintenance personnel, who can quickly make decisions based on this information, such as formulating targeted maintenance plans, optimizing equipment operation parameters, etc., to ensure that the system can return to normal in a short time, effectively reducing the system risks that may be caused by general anomalies, providing a scientific basis for operation and maintenance decisions, and improving the accuracy and efficiency of operation and maintenance.

[0116] When the anomaly level reaches a severe anomaly, the system will immediately trigger an emergency warning signal and suspend the power supply of relevant nodes. This is because in the case of severe anomalies, the system may face major risks, such as equipment damage, power supply interruption, etc., and the most urgent measures need to be taken to protect the system and equipment. At the same time, the abnormal data and related information are transmitted to the remote expert diagnosis system in real time to quickly obtain professional opinions and solutions. Ensure that the safety and reliability of the power supply protection system are guaranteed to the greatest extent and the losses and risks are minimized.

[0117] Based on the same inventive concept, the present invention provides a warning system for anomalies in a power supply protection system, including:

[0118] A data acquisition module for acquiring the topological structure, real-time operation data, and device status data of each node of the power supply protection network;

[0119] An instantaneous causal entropy calculation module for calculating the conditional entropy between each pair of variables based on the topological structure, real-time operation data, and device status data after determining the time delay window and causal lag coefficient; calculating the instantaneous causal entropy value between each pair of variables by performing a temporal difference processing on all conditional entropies;

[0120] A dynamic causal adjacency matrix construction module for calculating the causal strength between variable pairs according to the instantaneous causal entropy value; calculating the instantaneous causal strength between each node through a sliding window and constructing a dynamic causal adjacency matrix to reflect the causal relationship of the network changing over time;

[0121] A comprehensive feature vector module for updating and representing node features by combining the device topological structure and the dynamic causal adjacency matrix through layer-by-layer transmission and transformation of node features; extracting multi-scale temporal features using dilated causal convolution; fusing the updated node features and multi-scale temporal features through weighted summation to obtain a comprehensive feature vector;

[0122] A predicted value sample module is used to take the comprehensive feature vector as the input of a Bayesian neural network. The Bayesian neural network obtains multiple weight samples by sampling the weight distribution, performs forward propagation calculations on each sample to obtain multiple predicted value samples, takes the mean of the multiple predicted value samples as the final prediction result, and measures the uncertainty of the prediction with the variance of the multiple predicted value samples.

[0123] A determination module is used to compare the mean and variance of the predicted value samples with the corresponding preset thresholds respectively. If the comparison results do not exceed the error range, it is determined that the prediction is accurate; if the comparison results exceed the error range, an abnormal warning is triggered.

[0124] Based on the same inventive concept, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, a warning method for the abnormality of the power supply protection system is implemented.

[0125] Based on the same inventive concept, the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, a warning method for the abnormality of the power supply protection system is implemented.

[0126] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for early warning of abnormal power supply system, characterized in that: include: Obtain the topological structure, real-time operation data and equipment status data of each node of the power supply network; Based on the topology, real-time operation data and device status data, after determining the time delay window and causal lag coefficient, calculate the conditional entropy between each variable pair; By performing temporal difference processing on all conditional entropies, the instantaneous causal entropy value between each pair of variables is calculated; Calculating the causal strength between the variable pairs according to the instantaneous causal entropy value; The instantaneous causal strength between nodes is calculated through a sliding window, and a dynamic causal adjacency matrix is ​​constructed to reflect the causal relationship of the network over time; Combined with the device topology and dynamic causal adjacency matrix, the node features are updated and represented by transferring and transforming them layer by layer; multi-scale time series features are extracted using dilated causal convolution; the updated node features and multi-scale time series features are fused through weighted summation to obtain a comprehensive feature vector; The comprehensive feature vector is used as an input of a Bayesian neural network, and the Bayesian neural network obtains a plurality of weight samples by sampling weight distribution; Perform forward propagation calculation on each sample to obtain multiple predicted value samples; the mean of the multiple predicted value samples can be used as the final prediction result, and the variance of the multiple predicted value samples can be used to measure the uncertainty of the prediction; The mean and variance of the predicted value samples are compared with the corresponding preset thresholds. If the comparison result does not exceed the error range, the prediction is judged to be accurate; if the comparison result exceeds the error range, an abnormal warning is triggered.

2. The method for early warning of abnormal power supply system according to claim 1, characterized in that: After obtaining the topological structure, real-time operation data and device status data of each node of the power supply network, the method further includes: The topological structure is standardized, and the standardization includes unifying the node identification format, correcting the error information in the topological connection relationship, and removing the redundant topological connection information.

3. The method for early warning of abnormal power supply system according to claim 1, characterized in that: After obtaining the topological structure, real-time operation data and device status data of each node of the power supply network, the method further includes: Mean filtering is used to remove data points with abnormal mutations in the real-time operation data, and the median method is used to fill in the missing data values ​​in the real-time operation data.

4. The method for early warning of abnormal power supply system according to claim 1, characterized in that: After obtaining the topological structure, real-time operation data and device status data of each node of the power supply network, the method further includes: The equipment status data is classified and sorted, and the equipment is divided into different status levels according to factors such as equipment type, operating years, maintenance records, etc., and a corresponding weight coefficient is assigned to each level. At the same time, an association index is established between the equipment status data and the topological structure and real-time operation data.

5. The method for early warning of abnormal power supply system according to claim 4, characterized in that: The step of establishing the association index between the device status data and the topological structure and the real-time operation data specifically includes: A composite index structure is preset, which includes three basic dimensions of device ID, node ID, and timestamp, as well as additional dimensions of device status level and real-time operation data category, so as to realize multi-dimensional fast retrieval; For each device, a unique status code is assigned to it in the composite index structure according to its status level and category, and the status code can reflect the health status, maintenance priority and other information of the device; When real-time operation data is collected, the corresponding device ID, node ID and timestamp are automatically recorded and associated with the status code to form a data entry stored in the index database; When you need to analyze the status changes of a specific device or node, you can quickly retrieve the associated real-time operating data and device status information through conditions such as device ID, node ID or time range.

6. The method for early warning of abnormal power supply system according to claim 1, characterized in that: The step of calculating the instantaneous causal strength between nodes through a sliding window and constructing a dynamic causal adjacency matrix specifically includes: Define the size and step size of the sliding window; In each sliding window, based on the calculated instantaneous causal entropy value, a weighted method is used to calculate the instantaneous causal strength between variable pairs; When constructing a dynamic causal adjacency matrix, each element of the matrix represents the instantaneous causal strength of the corresponding node pair in the current window. As the window slides, the matrix is ​​updated in real time to dynamically reflect the changes in causal relationships in the network. A time decay factor is introduced to perform weighted summation on the causal strength in the historical window to reduce the impact of old data on the current state evaluation, so that the dynamic causal adjacency matrix can reflect recent changes in network state.

7. The method for early warning of abnormal power supply system according to claim 1, characterized in that: After the abnormal warning is triggered, the method further includes: According to the degree to which the mean of the predicted value sample deviates from the corresponding threshold and the size of the variance, the anomaly is divided into three levels: slight anomaly, general anomaly and severe anomaly; If the abnormality is a minor one, an early warning message is sent to the operation and maintenance personnel to remind them to pay attention to the operating status of the corresponding nodes and devices; For general anomalies, in addition to sending warning information, it also automatically starts detailed data collection and diagnosis procedures for the relevant nodes and their surrounding associated nodes, and adjusts some parameter settings of the Bayesian neural network to improve the recognition accuracy of abnormal situations. It also feeds back the diagnosis results and adjusted network parameters to the operation and maintenance personnel to assist them in troubleshooting and decision-making; For serious abnormalities, an emergency warning signal is triggered immediately, the power supply to the relevant nodes is suspended, and the abnormal data and related information are transmitted to the remote expert diagnosis system in real time.

8. An early warning system for abnormal power supply system, characterized in that: include: A data acquisition module is used to obtain the topological structure, real-time operation data and equipment status data of each node of the power supply network; An instantaneous causal entropy calculation module, for calculating the conditional entropy between each variable pair after determining a time delay window and a causal lag coefficient based on the topological structure, real-time operation data and device status data; By performing temporal difference processing on all conditional entropies, the instantaneous causal entropy value between each pair of variables is calculated; A dynamic causal adjacency matrix construction module is used to calculate the causal strength between the variable pairs according to the instantaneous causal entropy value; calculate the instantaneous causal strength between each node through a sliding window, and construct a dynamic causal adjacency matrix to reflect the causal relationship of the network changing over time; The comprehensive feature vector module is used to combine the device topology and the dynamic causal adjacency matrix, update the node features by transferring and transforming the node features layer by layer; extract multi-scale time series features by using dilated causal convolution; and fuse the updated node features and multi-scale time series features by weighted summation to obtain a comprehensive feature vector; A prediction value sample module, used to use the comprehensive feature vector as an input of a Bayesian neural network, and the Bayesian neural network obtains a plurality of weight samples by sampling weight distribution; Perform forward propagation calculation on each sample to obtain multiple predicted value samples; the mean of the multiple predicted value samples can be used as the final prediction result, and the variance of the multiple predicted value samples can be used to measure the uncertainty of the prediction; The judgment module is used to compare the mean and variance of the predicted value samples with the corresponding preset thresholds. If the comparison result does not exceed the error range, the prediction is judged to be accurate; if the comparison result exceeds the error range, an abnormal warning is triggered.

9. An electronic device, characterized in that: The electronic device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps corresponding to the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps corresponding to the method according to any one of claims 1 to 7 are implemented.

Images