Abnormity monitoring method for panoramic smart energy cloud-side collaboration

Abstract

The invention discloses an anomaly monitoring method for panoramic smart energy cloud edge collaboration, and the method comprises the following steps: S10, edge side credibility perception calculation: carrying out the real-time analysis of locally collected energy time sequence data through an edge side device, generating a preliminary anomaly probability Pe, and carrying out the calculation of the edge side credibility based on the probability Pe and the corresponding data; calculating an edge credibility score Re corresponding to the preliminary anomaly probability Pe by analyzing the difference with a historical normal mode and the uncertainty of local model fitting; and step S20, cloud credibility enhancement analysis: the cloud performs deep analysis according to the data uploaded by the edge side, generates a cloud anomaly probability Pc, and calculates a cloud credibility score Rc corresponding to the cloud anomaly probability Pc by analyzing the novelty of the data in global distribution and the stability of model prediction based on the probability Pc and the corresponding data. According to the invention, abnormity monitoring of panoramic smart energy cloud edge collaboration can be carried out more comprehensively.

Description

Technical Field

[0001] This invention relates to the field of anomaly monitoring, and more specifically to anomaly monitoring methods for panoramic smart energy cloud-edge collaboration. Background Technology

[0002] Panoramic smart energy cloud-edge collaboration is an energy management model that integrates global computing power in the cloud with local response capabilities on the edge. Through cloud-edge data interoperability and functional division, it enables panoramic monitoring and efficient collaborative operation of the energy system. However, under the current cloud-edge collaboration architecture, the collaboration mechanism between the edge and the cloud is not yet perfect, and the linkage of data processing and analysis is insufficient.

[0003] As a core component ensuring the stable operation of the panoramic smart energy cloud-edge collaborative anomaly monitoring system, it is necessary to simultaneously meet the timeliness of real-time data processing on the edge side and the comprehensiveness of global analysis on the cloud side. However, traditional monitoring methods have significant limitations: rigid weight allocation, unable to dynamically adjust based on the judgment performance of the edge side and the cloud; insufficient collaboration between the edge side and the cloud, lacking a precise fusion mechanism based on the trust between the two parties; edge node judgment performance easily decays over time, and lacks highly adaptable knowledge update methods; and insufficient consideration of the correlation of anomaly propagation between nodes, making it difficult to achieve early prediction of cross-node anomalies. Consequently, the real-time performance, accuracy, and adaptability of monitoring cannot meet the actual needs of the panoramic smart energy system. Therefore, an anomaly monitoring method for panoramic smart energy cloud-edge collaboration is proposed. Summary of the Invention

[0004] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps: Step S10, Edge-side credibility perception calculation: The edge-side device performs real-time analysis on the locally collected energy time-series data, generates a preliminary anomaly probability Pe, and calculates the edge credibility score Re corresponding to the preliminary anomaly probability Pe based on the probability Pe and its corresponding data by analyzing the differences from historical normal patterns and the uncertainty of local model fitting. Step S20, Cloud Trustworthiness Enhancement Analysis: The cloud performs in-depth analysis on the data uploaded from the edge side to generate the cloud anomaly probability Pc. Based on the probability Pc and its corresponding data, the cloud trustworthiness score Rc corresponding to the cloud anomaly probability Pc is calculated by analyzing the novelty of the data in the global distribution and the stability of the model prediction. Step S30, Two-way credibility calibration decision: Based on the edge credibility score Re, the cloud credibility score Rc, and a dynamically updated historical calibration coefficient η, the preliminary anomaly probability Pe and the cloud anomaly probability Pc are fused to generate the final anomaly probability P. final ; Step S40, Knowledge Collaboration and Evolution: Based on the final anomaly probability P finalThe cloud-based analysis model is updated based on the feedback data, and a lightweight knowledge package adapted to specific edge nodes is generated and distributed by extracting the decision logic related to the model's response to data from specific nodes.

[0005] Furthermore, the process of calculating the edge confidence score Re in step S10 is as follows: First, the difference from the historical pattern is quantified by calculating the minimum dynamic time warping distance between the current data window Xt and a reference set Dref consisting of recent historical normal data windows, denoted as dmin. Secondly, the uncertainty of the local model fit is quantified by fitting Xt with a low-order autoregressive model and calculating the variance of its predicted residual sequence, denoted as . ; Finally, the edge confidence score Re is dmin and The formula for calculating a monotonically decreasing function is: ; Where α and β are non-negative weight coefficients obtained through training with historical data.

[0006] Furthermore, the process of calculating the cloud credibility score R_c in step S20 is as follows: First, the novelty of the data in the global distribution is quantified by mapping the data to be tested onto the global feature space in the cloud and calculating the average Euclidean distance to its k nearest neighbor samples, denoted as d. global ; Secondly, by applying m random small perturbations to the data to be tested, we obtain m anomaly probability values ​​output by the cloud model and calculate their coefficient of variation to quantify the stability of the model prediction, denoted as CV. Finally, the cloud credibility score Rc is d. global The stability CV predicted by the model is a monotonically decreasing function, and its calculation formula is as follows: ; Wherein, γ and These are the preset non-negative weighting coefficients.

[0007] Furthermore, in step S30, the final anomaly probability P is generated. final The calculation process is as follows: Query the historical calibration coefficient \eta corresponding to the current Re and Rc value range. The value of this coefficient η ranges from [-0.5, 0.5]. Based on the historical calibration coefficient η, dynamic weights are assigned to the edge side and the cloud side respectively, and a weighted average is calculated. The formula is as follows: .

[0008] Furthermore, the online update process for the historical calibration coefficient η is as follows: After obtaining a true result feedback in a judgment, if the independent judgment on the edge side is correct, set Ae=1; if the independent judgment on the edge side is incorrect, set Ae=0. If the cloud-based independent judgment is correct, set Ac=1; if the cloud-based independent judgment is incorrect, set Ac=0. Whether the edge side or the cloud independently determines whether it is correct or not is determined by comparing its output anomaly probability with a preset threshold and comparing it with the actual results; Continuous Re and Rc values ​​are mapped to discrete interval identifiers using a predefined discretization method, and these identifiers are combined to form a two-dimensional index I, which is used to locate the corresponding coefficients. ; Based on the accuracy differences in this determination, and adjusted according to the closeness of Re and Rc, the corresponding coefficients were... The update formula is as follows: (This is followed by a series of seemingly unrelated phrases and sentences, making a coherent translation impossible.) ; in, Here, θ is the learning rate, and θ is a preset parameter for adjusting update sensitivity.

[0009] Furthermore, the process of generating the lightweight knowledge package in step S40 specifically includes the following sub-steps: Path analysis: Using cloud-based analytics models, historical data of target edge nodes are propagated forward to identify the computational path in the model that has the most significant response to the data pattern of that node. Subgraph extraction: Based on the computation path, a simplified model is extracted from the cloud model. This simplified model retains the core decision-making logic of the original model for the typical pattern of the node. Rule compilation: The decision logic of the simplified model is transformed into a set of rules in the form of "IF-THEN"; the feature threshold conditions in the rules are obtained by analyzing the feature distribution of historical normal samples and confirmed abnormal samples of the target edge nodes; the confidence weight associated with the rule is determined by calculating the accuracy of the rule on the validation set.

[0010] Furthermore, the process for determining the trigger conditions for distributing lightweight knowledge packages is as follows: Calculate the average edge confidence score of the most recent M decisions for the target edge node. and its long-term benchmark obtained through historical data statistics. The relative rate of decline; Calculate the correlation coefficient between the edge detection result of this node and the cloud detection result during the same period. and its long-term benchmark obtained through historical data statistics. The difference; The performance degradation index is obtained by weighting and summing the relative decline rate and the difference. The calculation formula is as follows: ; in, These are preset weighting coefficients; When performance degradation index If the threshold is exceeded, the knowledge package will be distributed.

[0011] Furthermore, the method also includes anomaly pattern propagation analysis steps: In the historical data, when node i experiences an anomaly, node j will occur in the subsequent time window. The conditional probability of an anomaly occurring within the system is denoted as... ; Calculate the average time interval between the occurrence of an anomaly at node i and the occurrence of an anomaly at node j, denoted as . ; Based on the physical and business logic relationships between nodes, a connection strength coefficient is preset for node pair (i,j). ; Calculate the pattern influence of node i on node j This influence is used to enhance the attention paid to the state of node i when the cloud model analyzes node j, and its calculation formula is as follows: ; in, The preset time decay constant, These are the preset weighting coefficients.

[0012] Compared with existing technologies, this invention has the following advantages: This anomaly monitoring method for panoramic smart energy cloud-edge collaboration achieves accurate monitoring of energy time-series data anomalies through four core steps: edge-side credibility perception calculation, cloud-side credibility enhancement analysis, bidirectional credibility calibration decision-making, and knowledge collaboration and evolution, combined with anomaly pattern propagation analysis. Dynamic weight allocation and iterative updates of historical calibration coefficients improve the accuracy of judgment, lightweight knowledge packages are distributed on demand to adapt to the characteristics of edge nodes, and anomaly pattern influence analysis enhances the cloud model's attention to the node association status. Overall, it takes into account the real-time performance, accuracy, and adaptability of monitoring, and optimizes the overall performance of anomaly monitoring in panoramic smart energy cloud-edge collaboration. Attached Figure Description

[0013] Figure 1 This is an overall structural diagram of the present invention. Detailed Implementation

[0014] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0015] like Figure 1 As shown, this embodiment provides a technical solution: an anomaly monitoring method for panoramic smart energy cloud-edge collaboration, comprising the following steps: Step S10, Edge-side credibility perception calculation: The edge-side device performs real-time analysis on the locally collected energy time-series data, generates a preliminary anomaly probability Pe, and calculates the edge credibility score Re corresponding to the preliminary anomaly probability Pe based on the probability Pe and its corresponding data by analyzing the differences from historical normal patterns and the uncertainty of local model fitting. Step S20, Cloud Trustworthiness Enhancement Analysis: The cloud performs in-depth analysis on the data uploaded from the edge side to generate the cloud anomaly probability Pc. Based on the probability Pc and its corresponding data, the cloud trustworthiness score Rc corresponding to the cloud anomaly probability Pc is calculated by analyzing the novelty of the data in the global distribution and the stability of the model prediction. Step S30, Two-way credibility calibration decision: Based on the edge credibility score Re, the cloud credibility score Rc, and a dynamically updated historical calibration coefficient η, the preliminary anomaly probability Pe and the cloud anomaly probability Pc are fused to generate the final anomaly probability P. final ; Step S40, Knowledge Collaboration and Evolution: Based on the final anomaly probability P final The cloud-based analysis model is updated based on the feedback data, and a lightweight knowledge package adapted to specific edge nodes is generated and distributed by extracting the decision logic related to the model's response to data from specific nodes.

[0016] The process of calculating the marginal confidence score Re in step S10 is as follows: First, the difference from the historical pattern is quantified by calculating the minimum dynamic time warping distance between the current data window Xt and a reference set Dref consisting of recent historical normal data windows, denoted as dmin. Secondly, the uncertainty of the local model fit is quantified by fitting Xt with a low-order autoregressive model and calculating the variance of its predicted residual sequence, denoted as . ; Finally, the edge confidence score Re is dmin and The formula for calculating a monotonically decreasing function is: ; Where α and β are non-negative weight coefficients obtained through training with historical data; The cloud credibility score Rc is constructed from two dimensions: novelty quantification of the global feature space and stability assessment of model prediction. This not only makes up for the limited perspective of local analysis on the edge side and can identify abnormal patterns in the global scope, but also reduces the risk of accidental misjudgment of the model through random perturbation and coefficient of variation. At the same time, the preset non-negative weights γ and δ can be flexibly adjusted according to different energy data types and model characteristics, so that Rc calculation can be adapted to various panoramic smart energy monitoring scenarios, providing reliable cloud credibility support for cloud-edge collaborative decision-making.

[0017] Preset weights: γ=0.2 (global distance weight), δ=0.3 (stability weight); Global feature space in the cloud: constructed based on 1000 sets of historical normal / abnormal energy data (such as power and voltage fusion features), k=5 (number of nearest neighbor samples); Perturbation parameter: Apply a random small perturbation of ±0.05 times the standard deviation to each feature dimension of the data, m=10 (number of perturbations); Anomaly probability threshold: 0.7 (used for subsequent independent judgment reference, not included in Rc calculation); Scenario 1: Normal data (data within a global distribution); Feature vector of the data to be tested: (Corresponding to the normal output power characteristics of a photovoltaic power station during a certain period); Calculate d global : Match k=5 nearest neighbor samples from the global sample library in the cloud, with feature vectors as follows: S1=[4.1,3.7,5.0], S2=[4.3,3.9,5.2], S3=[4.2,3.8,5.1], S4=[4.0,3.6,4.9], S5=[4.4,4.0,5.3]; Calculate the Euclidean distance: ; ; ; ; Mean Euclidean distance: ; Calculate CV (model prediction stability): Ten random perturbations are applied to Xfeat, resulting in ten sets of perturbed feature vectors. These vectors are then input into the cloud model to output the anomaly probability. ; Calculate the mean: ; Calculate the standard deviation: ; Calculate the coefficient of variation: ; Calculate Rc: Substituting into the formula, we get: This indicates that the probability of anomalies in the cloud, Pc, is highly reliable.

[0018] Scenario 2: Anomalous data (novel data outside the global distribution) Feature vector of the data to be tested: (Corresponding to the abnormal power characteristics caused by photovoltaic power station faults) Calculate d global : Match k=5 nearest neighbor samples with the following feature vectors: S6=[3.9,3.5,4.8], S7=[4.5,4.1,5.4], S8=[3.8,3.4,4.7], S9=[4.6,4.2,5.5], S10=[3.7,3.3,4.6]; Calculate the Euclidean distance: Similarly, we can calculate: ; ; ; ; Mean Euclidean distance: ; Calculate CV (model prediction stability): right After applying 10 random perturbations, output the anomaly probability: ; Calculate the mean: ; Calculate the standard deviation: ; Calculate the coefficient of variation: ; Calculate Rc: Substituting into the formula, we get: Although the model's predictions are stable, the high global novelty of the data justifies the reduction in Rc, which aligns with the logic for assessing the credibility of outlier data.

[0019] The process of calculating the cloud credibility score Rc in step S20 is as follows: First, the novelty of the data in the global distribution is quantified by mapping the data to be tested onto the global feature space in the cloud and calculating the average Euclidean distance to its k nearest neighbor samples, denoted as d. global ; Secondly, by applying m random small perturbations to the data to be tested, we obtain m anomaly probability values ​​output by the cloud model and calculate their coefficient of variation to quantify the stability of the model prediction, denoted as CV. Finally, the cloud credibility score Rc is d. global The stability CV predicted by the model is a monotonically decreasing function, and its calculation formula is as follows: ; Wherein, γ and These are preset non-negative weighting coefficients; By quantifying the novelty of data through the k-nearest neighbor average Euclidean distance in the global feature space of the cloud, the field of vision limitation of edge-side local analysis is broken, and new abnormal patterns can be accurately identified in the global scope. At the same time, by applying random small perturbations to the data and calculating the coefficient of variation (CV), the stability of model prediction is quantified, effectively filtering out accidental misjudgments and improving the reliability of judgment. Moreover, the preset non-negative weights γ and δ can be flexibly adjusted according to the energy data type (such as photovoltaic, electric, gas) and model characteristics, and can be adapted to different panoramic smart energy monitoring scenarios without reconstructing the calculation logic, providing accurate and reliable cloud credibility score Rc support for cloud-edge collaborative decision-making.

[0020] Preset weights: γ=0.2 (global distance weight), δ=0.3 (stability weight); Cloud-based global feature space: Constructed based on historical data from 1000 sets of normal / abnormal photovoltaic power plants, with feature dimensions including [active power P (kW), reactive power Q (kVar), solar irradiance G (kW / m²)]. 2 )] Nearest neighbor parameter: k=5 (select 5 nearest neighbor samples); Perturbation parameter: Apply a random small perturbation of ±0.05 times the standard deviation to each dimension of the feature, with the number of perturbations m=10; Features of some normal samples in the global sample library (used to calculate d) global ): S1=[4.1,3.7,5.0], S2=[4.3,3.9,5.2], S3=[4.2,3.8,5.1], S4=[4.0,3.6,4.9], S5=[4.4,4.0,5.3], S6=[3.9,3.5,4.8], S7=[4.5,4.1,5.4], S8=[3.8,3.4,4.7], S9=[4.6,4.2,5.5], S10=[3.7,3.3,4.6]; Scenario 1: Normal data (within the global distribution, the photovoltaic power station is operating stably) Feature vector of the data to be tested: ; Calculate global novelty d globalThe formula for three-dimensional Euclidean distance is: : Match k=5 nearest neighbor samples: S1, S2, S3, S4, S5; Calculate the Euclidean distance for each sample: ; ; ; ; ; Mean Euclidean distance:

[0021] Calculate the model's predicted stability (CV): The probability of anomalies in the cloud model output after 10 perturbations is: ; Calculate the mean: ; Calculate the sample standard deviation (denominator is m-1=9): ; ; Coefficient of variation: ; Calculate Rc: Substitute into the formula : This indicates that the probability of anomalies in the cloud, Pc, is highly reliable and conforms to the logic of normal data judgment.

[0022] Scenario 2: Abnormal data (PV module failure outside of global distribution) Feature vector of the data to be tested: A sudden drop in active power and abnormally high irradiance are consistent with component failure characteristics. Calculate global novelty d global : Match k=5 nearest neighbor samples (all normal samples): S6, S7, S8, S9, S10 Calculate the Euclidean distance for each sample: ; ; ; ; ; Mean Euclidean distance: ; Calculate the model's predicted stability (CV): After 10 perturbations, the cloud model outputs the anomaly probability (fault mode stable, probability concentrated): ; Calculate the mean: ; Calculate the sample standard deviation: ; ; Coefficient of variation:

[0023] Calculate Rc: Substitute into the formula to get Although the model prediction is stable, the global novelty of the data is extremely high, the Rc is reasonably reduced, and the credibility characteristics of the abnormal data are accurately reflected.

[0024] In step S30, the final anomaly probability P is generated. final The calculation process is as follows: Query the historical calibration coefficient \eta corresponding to the current Re and Rc value range. The value of this coefficient η ranges from [-0.5, 0.5]. Based on the historical calibration coefficient η, dynamic weights are assigned to the edge side and the cloud side respectively, and a weighted average is calculated. The formula is as follows: ; By introducing dynamically updated historical calibration coefficients η, and combining edge credibility score Re and cloud credibility score Rc, dynamic weights are assigned to the edge-side anomaly probability Pe and the cloud-side anomaly probability Pc. This avoids the rigidity of fixed weights, fully utilizes historical judgment experience and dual-end credibility information, and achieves accurate weighted fusion of Pe and Pc. It takes into account the advantages of real-time edge analysis and global cloud analysis, and can dynamically optimize the weight allocation logic based on historical performance. This makes the final anomaly probability Pfinal more in line with the actual monitoring scenario, and significantly improves the accuracy and reliability of cloud-edge collaborative anomaly judgment.

[0025] The historical calibration coefficient η ranges from -0.5 to 0.5, and the preset anomaly probability threshold is 0.7 (used as an independent criterion for updating the historical η; this example directly provides the η value corresponding to the scenario). The settings for Re, Rc, Pe, and Pc in each scenario are based on the previous settings. Fusion formula: ; Scenario 1: Both ends of the decision are reliable (η=0, neutral weights); Input parameters: Edge side: (Normal data credibility score) (Preliminary probability of anomalies in normal data) Cloud side: (Normal data credibility score) (Probability of abnormal data in the cloud) Historical calibration coefficient: η=0 (the accuracy of both ends of the historical judgment is consistent, so a neutral value is taken); Calculation process: Numerator: (1+0)×0.919×0.076+(1-0)×0.919×0.076=0.919×0.076+0.919×0.076≈0.070+0.070=0.14; Denominator: (1+0)×0.919+(1-0)×0.919=0.919+0.919=1.838; Final anomaly probability: ; Results show that Pfinal≈0.076<0.7, which is considered normal data and consistent with the actual scenario. The neutral weights achieve accurate fusion when both ends are trusted.

[0026] Scenario 2: Edge-side determination is more reliable (η=0.3, biased towards edge weights); Input parameters: - Edge side: Re=0.9 (edge ​​data fluctuates little and has high reliability). (Edge detection is normal) Cloud side: Rc=0.7 (slight interference in global cloud data, slightly lower reliability), Pc=0.1 (slight tendency for misjudgment in the cloud). Historical calibration coefficient: η=0.3 (Historical data shows that the accuracy of lower edge detection is higher in this Re and Rc interval) Mole: (1+0.3)×0.900×0.080+(1-0.3)×0.700×0.100=1.3×0.072+0.7×0.070=0.0936+0.049=0.1426; Denominator: (1+0.3)×0.900+(1-0.3)×0.700=1.17+0.49=1.66; Final anomaly probability: ; Results show that Pfinal≈0.086<0.7, indicating that the data is considered normal. The weights are biased towards the edge, effectively offsetting the impact of slight misjudgments in the cloud and closely reflecting the characteristics of the actual data.

[0027] Scenario 3: Cloud-based judgment is more reliable (η=-0.4, biased towards cloud weights); Input parameters: Edge side: Re=0.6 (edge ​​devices are temporarily affected by environmental interference, resulting in low reliability), Pe=0.15 (prone to false positives at the edge). Cloud-based: Rc=0.9 (global cloud analysis is interference-free and highly reliable), Pc=0.09 (accurate cloud-based determination is normal). Historical calibration coefficient: η = -0.4 (Historical data shows that cloud-based judgment accuracy is higher in this Re and Rc range). Calculation process: Mole: (1-0.4)×0.600×0.150+(1+0.4)×0.900×0.090=0.6×0.09+1.4×0.081=0.054+0.1134=0.1674; Denominator: (1-0.4)×0.600+(1+0.4)×0.900=0.36+1.26=1.62; Final anomaly probability: ; Results show that Pfinal≈0.103<0.7, indicating that the data is normal. The weighting is biased towards the cloud side, which successfully corrects the tendency to misjudge edge devices and ensures accurate judgment results.

[0028] Scenario 4: Abnormal data detection (η=0.2, slightly biased towards the edge); Input parameters: Edge side: Re=0.85 (confidence score for outlier data edges), Pe=0.82 (preliminary determination of anomalies at the edge). Cloud-based: Rc=0.8 (Cloud credibility score for abnormal data), Pc=0.85 (Cloud-based determination of abnormality) Historical calibration coefficient: η=0.2 (this allows for more timely detection of anomalies at the lower edges of the Re and Rc intervals); Calculation process: Mole: (1+0.2)×0.850×0.820+(1-0.2)×0.800×0.850=1.2×0.697+0.8×0.68=0.8364+0.544=1.3804; Denominator: (1+0.2)×0.850+(1-0.2)×0.800=1.02+0.64=1.66; Final anomaly probability: ; Results show that Pfinal≈0.832>0.7, indicating abnormal data. This accurately captures the abnormal characteristics of photovoltaic module failures, which is consistent with the actual scenario.

[0029] The online update process for the historical calibration coefficient η is as follows: After obtaining a true result feedback in a judgment, if the independent judgment on the edge side is correct, set Ae=1; if the independent judgment on the edge side is incorrect, set Ae=0. If the cloud-based independent judgment is correct, set Ac=1; if the cloud-based independent judgment is incorrect, set Ac=0. Whether the edge side or the cloud independently determines whether it is correct or not is determined by comparing its output anomaly probability with a preset threshold and comparing it with the actual results; Continuous Re and Rc values ​​are mapped to discrete interval identifiers using a predefined discretization method, and these identifiers are combined to form a two-dimensional index I, which is used to locate the corresponding coefficients. ; Based on the accuracy differences in this determination, and adjusted according to the closeness of Re and Rc, the corresponding coefficients were... The update formula is as follows: (This is followed by a series of seemingly unrelated phrases and sentences, making a coherent translation impossible.) ; in, Here, θ is the learning rate, and θ is a preset parameter for adjusting update sensitivity. By constructing an online update mechanism for historical calibration coefficients η based on real result feedback, and using the accuracy difference (Ae-Ac) between edge and cloud independent judgments as the core adjustment basis, η is dynamically optimized by combining the proximity of Re and Rc. Furthermore, the coefficients corresponding to the current Re and Rc intervals are accurately matched through a two-dimensional index I. This allows η to absorb historical judgment experience in real time and adapt to different confidence combination scenarios. At the same time, the update magnitude is controlled by the learning rate λ and the sensitivity parameter θ to avoid over-adjustment, ensuring the stability and adaptability of η. This provides coefficient support that is more in line with actual monitoring performance for dynamic weight allocation, and continuously optimizes the accuracy of the final anomaly probability Pfinal.

[0030] The learning rate λ = 0.1 (controls the update magnitude), and the sensitivity parameter θ = 0.1 (adjusts the effect of the proximity between Re and Rc on the update). Anomaly probability threshold = 0.7 (criterion for determining whether an independent judgment is correct or not); Discretization rule: Re and Rc are both divided into two intervals: [0, 0.5] and [0.5, 1.0], with interval labels of 0 and 1 respectively. The two-dimensional index I = (Re interval label, Rc interval label) Initial η value: Preset the initial ηold corresponding to each index, such as ηold=0.2 when I=(1,1) and ηold=-0.1 when I=(0,1); Scenario 1: Edge-side detection is more accurate, Re and Rc are close; Determine basic information: Real-world scenario: Photovoltaic power station operating stably (normal data); Edge side: Pe=0.08 (<0.7, normal, consistent with reality) → Ae=1; Re=0.85 (∈[0.5,1.0], interval identifier=1) Cloud side: Pc=0.82 (>0.7, considered abnormal, inconsistent with reality) → Ac=0; Rc=0.82 (∈[0.5,1.0], interval identifier=1) Two-dimensional index: I=(1,1) (Re identifies 1, Rc identifies 1) Calculation update process: Step 1: Calculate the accuracy difference (Ae-Ac) = 1 - 0 = 1 Step 2: Calculate the proximity term between Re and Rc: ; Step 3: Substitute into the update formula: ; ; Results show that after the update, η≈0.296 (still within the range of [-0.5,0.5]). Because the edge detection is more accurate, η is slightly adjusted towards the edge. The subsequent weight allocation of the Re and Rc intervals will focus more on the edge side, which is consistent with historical performance.

[0031] Scenario 2: Cloud-based judgment is more accurate, with a large difference between Re and Rc; Determine basic information: Real-world scenario: Minor fault in photovoltaic modules (abnormal data) Edge side: Pe=0.65 (<0.7, considered normal, inconsistent with reality) → Ae=0; Re=0.42 (∈[0,0.5), interval identifier=0); Cloud side: Pc=0.78 (>0.7, considered abnormal, consistent with reality) → Ac=1; Rc=0.90 (∈[0.5,1.0], interval label=1); Two-dimensional index: I=(0,1) (Re identifies 0, Rc identifies 1); Calculation and update process: Step 1: Calculate the accuracy difference (Ae-Ac) = 0 - 1 = -1 Step 2: Calculate the proximity term between Re and Rc: ; Step 3: Substitute into the update formula: ; Results indicate that after the update, η≈-0.1 (a very small change). Due to the large difference between Re and Rc, the proximity term approaches 0, weakening the impact of accuracy differences on η, preventing η from being overly biased towards the cloud, and ensuring coefficient stability. Subsequent intervals will maintain a moderately cloud-biased weighting while also taking into account adjustment caution.

[0032] The process of generating the lightweight knowledge package in step S40 specifically includes the following sub-steps: Path analysis: Using cloud-based analytics models, historical data of target edge nodes are propagated forward to identify the computational path in the model that has the most significant response to the data pattern of that node. Subgraph extraction: Based on the computation path, a simplified model is extracted from the cloud model. This simplified model retains the core decision-making logic of the original model for the typical pattern of the node. Rule compilation: The decision logic of the simplified model is transformed into a set of rules in the form of "IF-THEN"; the feature threshold conditions in the rules are obtained by analyzing the feature distribution of historical normal samples and confirmed abnormal samples of the target edge nodes; the confidence weight associated with the rule is determined by calculating the accuracy of the rule on the validation set. By accurately identifying the core computational paths in the cloud model that adapt to specific edge node data patterns through path analysis, a simplified model is obtained by extracting and pruning redundant structures from the subgraph, and then compiled into a lightweight knowledge package with "IF-THEN" rules. This retains the core decision-making logic for typical patterns of the target node while significantly reducing the storage and computational overhead of the knowledge package, adapting to the limited computing power of edge devices. At the same time, the rule thresholds are derived from the statistical analysis of the historical sample feature distribution of the node, and the confidence level is calculated based on the specific judgment performance of the validation set. The transformation process is transparent and traceable, allowing the edge side to quickly parse and execute, taking into account the accuracy, real-time performance, and adaptability of anomaly detection, and realizing the efficient and accurate migration of cloud knowledge to the edge side.

[0033] Target edge node: Photovoltaic inverter A (responsible for photovoltaic power conversion in region 3; historical data includes three modes: normal operation, overload, and voltage instability). Cloud-based analytics model: CNN model (input features are [output power P (kW), operating temperature T (°C), input voltage U (V)], containing 3 convolutional layers and 2 fully connected layers) Sample data: 500 historical normal samples, 200 overload abnormal samples, and 150 voltage instability abnormal samples; validation set includes 100 normal samples, 30 overload samples, and 20 voltage instability samples. Core objective: To generate a lightweight knowledge package adapted to inverter A for real-time edge-side detection of two types of anomalies: overload and voltage instability. Path analysis: Input historical data of 500 normal samples and 350 abnormal samples from inverter A, and then feed them into the cloud-based CNN model for forward propagation. By using gradient visualization and neuron response intensity analysis, the computational path that most significantly responds to the inverter A data pattern was identified: input layer → convolutional layer 1 (3×3 convolutional kernel, activation response value ≥ 0.85) → pooling layer (max pooling) → convolutional layer 2 (3×3 convolutional kernel, activation response value ≥ 0.90) → fully connected layer 1 (16 neurons, top 5 absolute weights) → output layer; This path contributes 92% to the identification of inverter A's overload and voltage instability modes, making it the core decision-making path.

[0034] Step 2: Subgraph extraction: Based on the above core path, the cloud CNN model is pruned: redundant convolutional kernels with response strength below the threshold in convolutional layer 3, fully connected layer 2, and convolutional layers 1 and 2 are removed (7 out of a total of 12 convolutional kernels are removed). Retaining core parameters: the weights of the 5 effective convolutional kernels in convolutional layer 1 (example of weight matrix:) (e.g.), the weights of the four effective convolutional kernels in convolutional layer 2, and the weights of the five key neurons in fully connected layer 1 (weight values ​​are 0.68, 0.75, 0.82, 0.71, and 0.65, respectively). The simplified model has only 18% of the parameters of the original model, and the storage usage has been reduced from 2.4MB to 0.43MB. Its decision logic is: output the probability distribution of the three types of judgment results [Pnormal, Poverload, Pvoltage], and select the category with the highest probability and ≥0.8 as the valid judgment result.

[0035] Simplified model decision logic decomposition: The core decision-making logic of the simplified CNN model is as follows: local features of P, T, and U are extracted through convolutional layers, and after weighted fusion by fully connected layers, the probabilities of the three types of modes are output. When the probability of a certain type of mode is the maximum value and ≥0.8, it is determined to be that type of mode; otherwise, it is determined to be "pending confirmation" (which needs to be verified again by combining real-time data from the edge side).

[0036] Core Feature Filtering: By analyzing the weight contribution of the simplified model: Overload mode: P (power) contributes 78% of the weight, T (temperature) contributes 20% of the weight, and U (voltage) contributes 2% of the weight. Therefore, P and T are identified as the core features for overload detection. Voltage instability mode: U (voltage) has a weight contribution of 92%, while P and T have a combined weight contribution of 8%. Therefore, U is identified as the core feature for determining voltage instability. Normal mode: The weights of P, T, and U are 35%, 30%, and 35%, respectively, and all three must meet the normal feature interval.

[0037] Specific statistical calculations of feature thresholds: Based on the feature distribution of historical samples of the target node, a 95% confidence interval is used ( The threshold is calculated and determined after allowing for actual operational fluctuations. Normal samples (500 groups): P: mean Standard deviation 95% confidence interval Take the interval expansion boundaries as 3.5 (lower limit) and 5.5 (upper limit); T: Mean Standard deviation 95% confidence interval = Take the interval expansion boundaries as 40 (lower limit) and 60 (upper limit); U: Mean Standard deviation 95% confidence interval = Take the interval expansion boundaries as 220 (lower limit) and 240 (upper limit).

[0038] Overload anomaly samples (200 groups): P: mean Standard deviation 95% confidence interval = Take the lower limit of the interval as 6; T: Mean Standard deviation 95% confidence interval = Take the lower limit of the interval as 65.

[0039] Voltage instability anomaly samples (150 groups): Low voltage subset (80 groups): mean Standard deviation 95% confidence interval = Take the upper limit of the interval as 215; High voltage subset (70 groups): mean Standard deviation 95% confidence interval = Take the lower limit of the interval as 245.

[0040] Transformation and confidence calculation of the "IF-THEN" rule: Based on the mapping relationship of "features satisfying threshold → simplified model corresponding pattern probability ≥ 0.8", it is transformed into a rule, and the confidence score is calculated through the validation set (confidence score = number of times the rule is correctly judged / number of times the rule is hit): Rule 1 (Overload Anomaly): Transformation logic: When P≥6.0 and T≥65, the simplified model outputs a probability of Poverload≥0.8 of 93%, so it is transformed into "IFP≥6.0 ANDT≥65THEN is judged as an overload anomaly"; Validation set performance: After traversing 30 overloaded samples and 120 non-overloaded samples, the rule was matched in 30 samples (all of which were overloaded samples). Among them, 28 samples were judged to have Poverload≥0.8 by the simplified model (correct), and 2 samples were misjudged due to instantaneous fluctuations of the sensor, resulting in P_overload=0.75. The confidence level was 28 / 30≈0.93.

[0041] Rule 2 (Voltage Instability Anomaly): Conversion logic: When U≤215 or U≥245, the simplified model outputs Pvoltage≥0.8 with a probability of 95%, so it is converted to "IFU≤215ORU≥245THEN is judged as voltage instability abnormality"; Validation set performance: After traversing 20 sets of unstable voltage samples and 110 sets of non-unstable voltage samples, the rule was matched in 20 sets (all of which were unstable voltage samples). 19 sets were judged to have Pvoltage≥0.8 by the simplified model (correct), and 1 set was misjudged as Pvoltage=0.78 due to instantaneous voltage fluctuation. The confidence level was 19 / 20=0.95.

[0042] Rule 3 (Normal Operation): Transformation logic: When 3.5≤P≤5.5 and 40≤T≤60 and 220≤U≤240, the simplified model outputs Pnormal≥0.8 with a probability of 98%, so it is transformed into "IF3.5≤P≤5.5AND40≤T≤60AND220≤U≤240THEN is judged as normal operation"; Validation set performance: After traversing 100 normal samples and 50 abnormal samples, the rule was matched in 95 groups (all of which were normal samples). 93 groups were judged to have Pnormal ≥ 0.8 (correct) by the simplified model, and 2 groups were misjudged due to data noise, resulting in Pnormal = 0.77. The confidence level was 93 / 95 ≈ 0.98.

[0043] Storage overhead: The knowledge package contains only 3 rules and their corresponding feature thresholds and confidence parameters, occupying only 0.02KB of storage, far lower than the 2.4MB of the original CNN model; Computational efficiency: When the edge device of inverter A (microcontroller computing power, main frequency 80MHz) runs this knowledge package, a single judgment only needs to parse the rule logic and compare the values, which takes about 7ms, while running the original CNN model directly takes 52ms, which meets the requirement of real-time monitoring within 100ms on the edge side; Judgment accuracy: The knowledge package was used to test 100 sets of new data (60 sets of normal data, 25 sets of overload data, and 15 sets of unstable voltage data) on the edge side. 94 sets were correctly judged, with an accuracy rate of 94%, which is close to the 96% accuracy rate of the cloud CNN model. Only 2% of the accuracy was lost, which fully meets the actual monitoring requirements.

[0044] The process for determining the trigger conditions for the distribution of lightweight knowledge packages is as follows: Calculate the average edge confidence score of the most recent M decisions for the target edge node. and its long-term benchmark obtained through historical data statistics. The relative rate of decline; Calculate the correlation coefficient between the edge detection result of this node and the cloud detection result during the same period. and its long-term benchmark obtained through historical data statistics. The difference; The performance degradation index is obtained by weighting and summing the relative decline rate and the difference. The calculation formula is as follows: ; in, These are preset weighting coefficients; When performance degradation index If the preset threshold is exceeded, the knowledge package will be distributed. By constructing a performance degradation index based on two dimensions—"relative decline rate of average edge credibility score + difference in correlation coefficient between edge and cloud judgment results"—and combining it with preset weights to quantify changes in edge node judgment capabilities, this approach accurately captures the core issues of declining edge credibility and reduced consistency between cloud and edge judgments. Furthermore, it avoids misjudgments through dynamic comparison of historical benchmarks and recent performance, triggering the distribution of lightweight knowledge packages only when performance degradation exceeds a threshold. This ensures that the accuracy of edge node judgments remains consistently online while avoiding the cloud-edge communication overhead and edge computing power waste caused by redundant knowledge package distribution, thus achieving on-demand and precise triggering of knowledge collaboration.

[0045] Long-term benchmark (obtained from historical data statistics): Long-term average marginal credibility score Long-term cloud edge determination correlation coefficient ; Preset parameters: weighting coefficient ω=0.5, most recent decision count M=20, performance degradation threshold. (If the limit is exceeded, a knowledge package will be distributed.) Performance degradation index formula: ; (Note: The average edge confidence score for the most recent M times. (The correlation coefficient of the most recent M cloud edge determinations). Scenario 1: Stable edge performance, no triggering of distribution: Basic data (results from the last 20 judgments): The most recent 20 edge credibility scores (Re) are: [0.91, 0.93, 0.90, 0.92, 0.94, 0.91, 0.93, 0.92, 0.90, 0.91, 0.93, 0.92, 0.94, 0.90, 0.91, 0.92, 0.93, 0.91, 0.92, 0.90]; Recent 20 cloud edge detection results (0 = normal, 1 = abnormal): Edge detection: [0,0,0,1,0,1,0,0,1,0,0,1,0,0,1,0,0,1,0,0]; Cloud-based determination: [0,0,0,1,0,1,0,0,1,0,0,1,0,0,1,0,0,1,0,0]; Calculation process: Step 1: Calculate the average marginal credibility score : ; Step 2: Calculate the relative decline rate : ; Step 3: Calculate the current correlation coefficient (Pearson correlation coefficient formula:) ): Let edge detection be X, cloud detection be Y, and there be a total of 20 sets of data (n=20): Mean: ; covariance (Since X and Y are completely identical, covariance = Var(X)) variance ; Correlation coefficient: (In reality, since the data is completely consistent, ρ=1.0) Step 4: Calculate the performance degradation index : ; (Threshold) The edge node has stable performance and does not need to distribute knowledge packages, thus avoiding redundant communication.

[0046] Scenario 2: Edge performance degradation triggers a release; Basic data (results from the last 20 assessments, edge devices are affected by environmental interference): The most recent 20 edge credibility scores (Re): [0.85, 0.83, 0.87, 0.82, 0.84, 0.81, 0.86, 0.83, 0.80, 0.82, 0.85, 0.83, 0.84, 0.81, 0.80, 0.83, 0.82, 0.81, 0.84, 0.80] Recent 20 cloud edge detection results (0 = normal, 1 = abnormal): Edge detection: [0,1,0,1,0,0,1,0,1,0,1,0,0,1,0,1,0,0,1,0,0,1,0] (increased false positives); Cloud-based determination: [0,0,0,1,0,1,1,0,1,0,0,0,0,1,0,1,0,1,1,0] (precise determination); Calculation process: Step 1: Calculate the average marginal credibility score : ; Step 2: Calculate the relative decline rate : ; Step 3: Calculate the current correlation coefficient : Let edge detection be X, cloud detection be Y, and n=20: Mean:

[0047] covariance : ; ; variance ; Correlation coefficient: ; Step 4: Calculate the performance degradation index : ; Results Explanation: ≈0.2888>0.15 (threshold), the performance of the edge node degrades significantly, triggering the distribution of a lightweight knowledge package. The edge judgment logic is updated through cloud knowledge to restore the judgment accuracy.

[0048] The method also includes anomaly pattern propagation analysis steps: In the historical data, when node i experiences an anomaly, node j will occur in the subsequent time window. The conditional probability of an anomaly occurring within the system is denoted as... ; Calculate the average time interval between the occurrence of an anomaly at node i and the occurrence of an anomaly at node j, denoted as . ; Based on the physical and business logic relationships between nodes, a connection strength coefficient is preset for node pair (i,j). ; Calculate the pattern influence of node i on node j This influence is used to enhance the attention paid to the state of node i when the cloud model analyzes node j, and its calculation formula is as follows: ; in, The preset time decay constant, These are preset weighting coefficients; Through conditional probability ( Quantify the probability of abnormal propagation of nodes, combined with the average time interval ( The time decay effect and the connection strength coefficient of physical / business logic between nodes () ), accurately calculate the pattern influence of node i on j This allows cloud-based models to focus on the status of high-influence precursor nodes when analyzing target nodes, breaking the limitations of isolated monitoring of single nodes, predicting the risk of anomaly propagation in advance, and significantly improving the timeliness of cross-node anomaly identification and the comprehensiveness of global monitoring in the panoramic smart energy system.

[0049] Preset parameter: Time decay constant (Unit: hours, adapted to the daily cycle operation characteristics of photovoltaic power plants), weighting coefficient μ=0.3 (balancing the influence of propagation probability and connection strength). Anomaly time window Δt = 12 hours (to count whether node j becomes abnormal within 12 hours after node i becomes abnormal). Node definition: Node i: Inverter A (responsible for photovoltaic power conversion in area 3, core equipment); Node j: Energy storage module B (directly connected to inverter A, receiving its output electrical energy); Node k: Photovoltaic panel group C (provides input power to inverter A, and is physically closely related); Influence Formula: ; Scenario 1: Calculation of the influence of inverter A(i) on energy storage module B(j); Basic data (based on statistics of anomaly records over the past year): Conditional probability Historically, after inverter A malfunctioned, energy storage module B malfunctioned 36 times within 12 hours, and inverter A malfunctioned a total of 40 times. ; Average time interval The time intervals between the inverter A malfunction and the energy storage module B malfunction were [2, 3, 1.5, 4, 2.5, 3.5, 1, 2.2, 3.8, 4.5, ...] (a total of 36 data sets), with an average value of: Hour Connection strength coefficient Inverter A and energy storage module B are directly electrically connected, and there is an energy transmission dependency in the business. ; Calculation process: Time decay term: ; Connection strength contribution: ; Model Influence: ; (The higher the value, the stronger the influence.) When analyzing energy storage module B in the cloud, the focus will be on the operating status of inverter A. If inverter A malfunctions, the abnormal risk of energy storage module B can be predicted in advance, shortening the response time.

[0050] Scenario 2: Calculation of the influence of photovoltaic panel group C(k) on inverter A(i); Basic data (based on statistics of anomaly records over the past year): Conditional probability Historically, after photovoltaic panel C malfunctioned, inverter A malfunctioned 27 times within 12 hours, and photovoltaic panel C malfunctioned a total of 30 times. Therefore... ; Average time interval The time intervals between the malfunction of photovoltaic panel group C and the malfunction of inverter A were [0.5, 1, 0.8, 1.2, 0.6, 0.9, 1.1, 0.7, 0.4, 1.3, ...] (a total of 27 data sets), with an average value of... Hour Connection strength coefficient Photovoltaic panel group C and inverter A are directly physically connected and are the core source of energy input. (Preset) ; Calculation process: Time decay term: ; Connection strength contribution: ; Model Influence: ; The impact of inverter A on energy storage module B is greater than that of inverter A, because the impact of photovoltaic panel anomalies on inverters is more direct and the response is faster. When analyzing inverter A in the cloud, the status of photovoltaic panel group C will be monitored first. If photovoltaic panel group C has problems such as voltage abnormality or shading, it can provide early warning of overload or undervoltage anomalies in inverter A, thereby achieving source risk management.

[0051] When the photovoltaic panel group C(k) experiences an abnormal shading condition (determined as an abnormal state), the cloud model... The value reached 1.139, and the operational data of inverter A(i) will be closely monitored. Subsequent monitoring revealed that the input power of inverter A dropped below the normal threshold after 0.9 hours. The cloud-based [system / mechanism]... The early monitoring allowed for the identification of abnormal trends in inverter A 2.1 hours earlier than traditional isolated monitoring, providing maintenance personnel with ample time to address the issue and preventing the anomaly from spreading further to energy storage module B.

[0052] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0053] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0054] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An anomaly monitoring method for panoramic smart energy cloud-edge collaboration, characterized in that, Includes the following steps: Step S10, Edge-side credibility perception calculation: The edge-side device performs real-time analysis on the locally collected energy time-series data, generates a preliminary anomaly probability Pe, and calculates the edge credibility score Re corresponding to the preliminary anomaly probability Pe based on the probability Pe and its corresponding data by analyzing the differences from historical normal patterns and the uncertainty of local model fitting. Step S20, Cloud Trustworthiness Enhancement Analysis: The cloud performs in-depth analysis on the data uploaded from the edge side to generate the cloud anomaly probability Pc. Based on the probability Pc and its corresponding data, the cloud trustworthiness score Rc corresponding to the cloud anomaly probability Pc is calculated by analyzing the novelty of the data in the global distribution and the stability of the model prediction. Step S30, Two-way credibility calibration decision: Based on the edge credibility score Re, the cloud credibility score Rc, and a dynamically updated historical calibration coefficient η, the preliminary anomaly probability Pe and the cloud anomaly probability Pc are fused to generate the final anomaly probability P. final ; Step S40, Knowledge Collaboration and Evolution: Based on the final anomaly probability P final The cloud-based analysis model is updated based on the feedback data, and a lightweight knowledge package adapted to specific edge nodes is generated and distributed by extracting the decision logic related to the model's response to data from specific nodes.

2. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 1, characterized in that: The process of calculating the marginal confidence score Re in step S10 is as follows: First, the difference from the historical pattern is quantified by calculating the minimum dynamic time warping distance between the current data window Xt and a reference set Dref consisting of recent historical normal data windows, denoted as dmin. Secondly, the uncertainty of the local model fit is quantified by fitting Xt with a low-order autoregressive model and calculating the variance of its predicted residual sequence, denoted as . ; Finally, the edge confidence score Re is dmin and It is a monotonically decreasing function.

3. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 2, characterized in that: The process of calculating the cloud credibility score R_c in step S20 is as follows: First, the novelty of the data in the global distribution is quantified by mapping the data to be tested onto the global feature space in the cloud and calculating the average Euclidean distance to its k nearest neighbor samples, denoted as d. global ; Secondly, by applying m random small perturbations to the data to be tested, we obtain m anomaly probability values ​​output by the cloud model and calculate their coefficient of variation to quantify the stability of the model prediction, denoted as CV. Finally, the cloud credibility score Rc is d. global The monotonically decreasing function of the stability CV predicted by the model; Wherein, γ and These are the preset non-negative weighting coefficients.

4. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 3, characterized in that: In step S30, the final anomaly probability P is generated. final The calculation process is as follows: Query the historical calibration coefficient \eta corresponding to the current Re and Rc value range. The value of this coefficient η ranges from [-0.5, 0.5]. Based on the historical calibration coefficient η, dynamic weights are assigned to the edge side and the cloud side respectively, and a weighted average is performed to obtain the final anomaly probability P. final .

5. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 4, characterized in that: The online update process for the historical calibration coefficient η is as follows: After obtaining a true result feedback in a judgment, if the independent judgment on the edge side is correct, set Ae=1; if the independent judgment on the edge side is incorrect, set Ae=0. If the cloud-based independent judgment is correct, set Ac=1; if the cloud-based independent judgment is incorrect, set Ac=0. Whether the edge side or the cloud independently determines whether it is correct or not is determined by comparing its output anomaly probability with a preset threshold and comparing it with the actual results; Continuous Re and Rc values ​​are mapped to discrete interval identifiers using a predefined discretization method, and these identifiers are combined to form a two-dimensional index I, which is used to locate the corresponding coefficients. ; Based on the accuracy differences in this determination, and adjusted according to the closeness of Re and Rc, the corresponding coefficients were... Perform iterative corrections.

6. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 5, characterized in that: The process of generating the lightweight knowledge package in step S40 specifically includes the following sub-steps: Path analysis: Using cloud-based analytics models, historical data of target edge nodes are propagated forward to identify the computational path in the model that has the most significant response to the data pattern of that node. Subgraph extraction: Based on the computation path, a simplified model is extracted from the cloud model. This simplified model retains the core decision-making logic of the original model for the typical pattern of the node. Rule compilation: The decision logic of the simplified model is transformed into a set of "IF-THEN" rules; the feature threshold conditions in the rules are obtained by analyzing the feature distribution of historical normal samples and confirmed abnormal samples of the target edge node; The confidence weight associated with a rule is determined by calculating the accuracy of that rule on the validation set.

7. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 6, characterized in that: The process for determining the trigger conditions for the distribution of lightweight knowledge packages is as follows: Calculate the average edge confidence score of the most recent M decisions for the target edge node. and its long-term benchmark obtained through historical data statistics. The relative rate of decline; Calculate the correlation coefficient between the edge detection result of this node and the cloud detection result during the same period. and its long-term benchmark obtained through historical data statistics. The difference; The performance degradation index is obtained by weighting and summing the relative decline rate and the difference. ; When performance degradation index If the threshold is exceeded, the knowledge package will be distributed.

8. The anomaly monitoring method for panoramic smart energy cloud-edge collaboration according to claim 7, characterized in that: The method also includes anomaly pattern propagation analysis steps: In the historical data, when node i experiences an anomaly, node j will occur in the subsequent time window. The conditional probability of an anomaly occurring within the system is denoted as... ; Calculate the average time interval between the occurrence of an anomaly at node i and the occurrence of an anomaly at node j, denoted as . ; Based on the physical and business logic relationships between nodes, a connection strength coefficient is preset for node pair (i,j). ; Calculate the pattern influence of node i on node j This influence is used to enhance the focus on the state of node i when analyzing node j in the cloud model.

Images