Multi-feature power grid data fusion method based on dynamic quality estimation
By using a multi-feature classification framework with dynamic quality estimation and consensus-aware residual gating, the problem of heterogeneity in the quality of multi-source data in power systems is solved, the robustness and accuracy in complex scenarios are improved, and the robustness and interpretability of low-quality information are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies cannot effectively handle the dynamic quality heterogeneity of multi-source heterogeneous data in power systems, resulting in the degradation of fusion effects in complex scenarios and a lack of robustness and interpretability for low-quality information.
We adopt a multi-feature classification framework based on dynamic quality estimation. Through sample-level three-dimensional dynamic quality assessment and consensus-aware residual gating, we adaptively allocate weights to suppress low-quality information and retain beneficial complementary information.
It significantly improves robustness and accuracy under noisy, missing, and adversarial perturbation conditions, ensuring clear and traceable decision-making processes and real-time deployment with low computational overhead.
Smart Images
Figure CN121786722A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-feature analysis of power data, and more specifically to a multi-feature classification method for power data. Background Technology
[0002] With the high proportion of renewable energy grid connection, the deep integration of energy-consuming electrification and information and communication technologies, power system operation is characterized by strong fluctuations, high uncertainty, and frequent events. Facing this complex environment, multi-feature (multi-modal) power analysis has gradually become an important technical approach for core tasks such as situation identification, disturbance / fault type discrimination, anomaly detection, and alarm management. On the one hand, it requires the simultaneous use of multi-source heterogeneous data, including PMU high-frequency measurements, SCADA steady-state quantities, AMI / energy-consuming behavior, equipment health and alarm records, meteorological and load forecasts, and grid topology. On the other hand, it must address the significant differences in sampling frequency, time alignment, data integrity, and noise statistics between cross-source data to form a more comprehensive, timely, and reliable understanding of the system status under real-world operating conditions.
[0003] To address these needs, research and engineering practices have led to the development of various fusion paradigms and their extensions: data-level stitching and feature-level (early) fusion can capture underlying correlations at the model front end, but they are highly sensitive to asynchronicity and alignment errors, and are prone to the curse of dimensionality and noise amplification; decision-level (late) fusion has modularity and natural robustness to missing views, but simple averaging / voting implicitly assumes equal weighting, making it susceptible to being dragged down by isolated and abnormal views; attention-based weight allocation can learn adaptively, but the source of weights is often a black box, resulting in insufficient interpretability and auditability in critical infrastructure scenarios, and unstable training when samples are sparse, views conflict, or data distribution drifts; solutions that introduce evidence theory or additional quality prediction networks are theoretically more complete, but are complex to implement, computationally and maintenance-intensive, and difficult to meet the requirements of real-time and online operation and maintenance. More importantly, real power grids generally exhibit sample-level and time-varying data quality heterogeneity (such as sensor aging and saturation, communication packet loss and delay, asynchronous sampling and clock skew, adversarial disturbances and sudden changes in operating conditions). Existing methods often lack fine-grained measurements of the credibility and consensus of the current sample and the current view, making it difficult to suppress low-quality evidence and retain useful complementary information in a timely manner, thus causing the fusion effect to degrade in complex scenarios.
[0004] In conclusion, existing technologies cannot maintain robustness to low-quality information and conflicting data while ensuring that the decision-making process is clear and traceable. Summary of the Invention
[0005] To address the shortcomings of existing technologies (multi-source data in real power grid scenarios exhibit dynamic changes in sample quality and information conflicts; traditional early / late fusion is easily hampered by a small number of low-quality samples; black-box attention lacks interpretability; and evidence theory / additional quality networks are complex and costly to implement), this invention proposes a multi-feature classification framework based on dynamic quality assessment and adaptive fusion. While ensuring a clear and traceable decision-making process, it effectively improves the robustness of the method in scenarios with low-quality information, information conflicts, and missing information.
[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This solution provides a multi-feature power grid data fusion method based on dynamic quality estimation, including the following steps: S1: Processes various input power data and calculates view-level classification probability output; S2: Perform sample-level three-dimensional dynamic quality assessment based on the obtained probability output; S3: Implement consensus-aware residual gating, and obtain sample weights through multiplicative gating and normalization; S4: Dynamically fuse multiple power data based on the calculated sample weights.
[0007] The beneficial effects of this invention are as follows: without relying on an additional quality prediction network, interpretable quality assessment and differential weighting of each view's evidence are performed at the sample level, significantly improving robustness and accuracy under noise, missing, and adversarial perturbation conditions; consensus-aware residual gating avoids decision-making hijacked by a single high-confidence view that conflicts with the group, effectively improving identification stability under complex operating conditions; and the modular decision-level fusion and low computational overhead implementation path facilitate real-time deployment and maintenance in critical power infrastructure.
[0008] Furthermore, step S1 includes the following sub-steps: S11: Construct an independent sub-classifier for each feature view and compute class logits. S12: Calculate the posterior probability distribution using the activation function Softmax.
[0009] Further, in step S11, for each feature view Build independent subclassifiers Input raw and preprocessed features Output categorical logit evidence (logits) .in For sample index, For view Feature dimensions, This represents the number of categories.
[0010] Furthermore, the calculation formula in step S12 is as follows: in It is a sample In view The posterior probability vector on, It is the sample number. These are view labels. This represents a function that normalizes a vector exponentially. Indicates sample In view Logarithmic evidence vector on, Indicates sample In view The above belongs to the first The probability of each category It is a category label. Represented by natural constant Scalar as base Take the index. Representing vectors The One portion, Indicates the category label From 1 to Sum of the exponent terms, It is the category label for summation. This indicates the total number of categories.
[0011] The beneficial effect of the above-mentioned further scheme is that, through the above processing, the corresponding classification results can be calculated using multiple features, which facilitates subsequent processing methods.
[0012] Furthermore, step S2 includes the following sub-steps: S21: Measure self-certainty using maximum a posteriori or entropy normalization to obtain confidence level; S22: Consensus is obtained by measuring the consistency with the average inner product of the probability distributions of the remaining views; S23: Use Jensen–Shannon divergence (JSD) to measure beneficial differences and threshold to suppress invalid deviations, thus obtaining uniqueness.
[0013] Furthermore, the formula for calculating the confidence level in step S21 is as follows: in It is a sample In view The maximum posterior confidence level. It is the sample number. These are view labels. Indicates the category label Take the maximum value. Indicates sample In view The above belongs to the first The probability of a class Indicates sample In view Shannon entropy over probability distribution, Represents probability Take the natural logarithm. This represents the confidence level after entropy normalization. Representing entropy and Normalize the ratio and take the complement. Represents the total number of categories Take the natural logarithm. This indicates that the confidence level is between 0 and 1.
[0014] Furthermore, the formula for calculating the consensus degree in step S22 is as follows: in It is a sample In view Consensus on It is the sample number. These are view labels. Indicates the total number of views. The reciprocal of the result after subtracting 1 is used as the average coefficient. It is the total number of views. It is the label of another view. It is a sample In view The posterior probability vector on, Indicates sample In view With View The dot product of probability vectors, This indicates the range of values for the consensus level.
[0015] Furthermore, the formula for calculating uniqueness in step S23 is as follows: in It is except for the view The mean probability of other views. It is a sample In view The posterior probability vector on; This represents the average distribution of the current view with other views. Next: It is the Kullback-Leibler divergence, which represents the difference between two probability distributions; It is a uniqueness measure that is not thresholded, representing the difference between the current view and other views; It is a uniqueness threshold used to remove invalid, minute differences, resulting in the final product. Only the differences that are useful for classification are retained.
[0016] The beneficial effects of the above-mentioned further scheme are that by dynamically evaluating and thresholding confidence, consensus, and uniqueness in three dimensions, it can accurately identify contamination and beneficial differences, suppress abnormal views, retain complementary information, alleviate the misleading effects caused by low-quality information / information conflicts / missing information, improve the stability of weight allocation and fusion, and significantly enhance the robustness and interpretability of the model.
[0017] Furthermore, step S3 includes the following sub-steps: S31: Generate sample-level differentiated gating weights from the calculated three-dimensional metric components; S32: Normalize the weights using the temperature coefficient; S33: Filter out complementary information that is worth keeping.
[0018] Furthermore, the formula for calculating the weights in step S31 is as follows: in Indicates sample In view The unnormalized gate value below indicates the importance of the view; These are hyperparameters that control the confidence level. Consensus and uniqueness The relative importance of gating.
[0019] Furthermore, the normalization formula in step S32 is as follows: in, It is a sample view The final fusion weights, temperature parameters Control the sharpness of the weight distribution. For the sample In view The unnormalized gate value on. When When taking the smaller value, the larger one This will be highlighted more strongly; the denominator is indexed by all views. Summation is performed to complete the normalization, making the total weights equal to... .
[0020] Furthermore, the formula for calculating complementary information in step S33 is as follows: Indicates a fixed sample Excluded View The arithmetic mean of the logarithmic evidence vectors of the remaining views. Indicates to Calculate the average coefficient for each view. Indicates sample In view Logarithmic evidence vector on, Indicates sample In view The residual correction term on, Represents the residual strength coefficient and satisfies , Indicates sample In view Uniqueness score, This represents the difference vector between the current view and the mean of the remaining views.
[0021] The beneficial effects of the above-mentioned further scheme are as follows: by using consensus-aware multiplicative gating and temperature normalization in conjunction with residual correction, weights are adaptively allocated based on confidence, consensus, and uniqueness, outlier views are suppressed and beneficial differences are retained. This solves the problem that static fusion is difficult to balance removing low-quality information and preserving beneficial information with lower computational cost, making the model more robust, accurate and interpretable under complex conditions.
[0022] Furthermore, step S4 includes the following sub-steps: S41: Based on the sample-level weights and complementary information obtained above, the fused classification information is obtained; S42: The classification information is transformed into a predicted probability distribution through the Softmax activation function.
[0023] Furthermore, the formula for calculating the fused classification information in step S41 is as follows: in It is a sample The fused logarithmic evidence vector, This represents the logarithmic evidence vector for each view. By weight Vector-wise weighted summation This indicates that the residual correction term for each view is... Summation of vectors one by one It is the sample number. These are view labels. This is the total number of views.
[0024] Furthermore, the formula for calculating the predicted probability distribution in step S42 is as follows: in It is a sample The fused posterior probability vector This represents a mapping that performs exponential normalization on the fused logarithmic evidence vector. Indicates sample The fusion post-hoc in the first The probability of each category Represented by natural constant The first piece of evidence for the fusion of logarithmic evidence Each component Take the index. Indicates the category label From 1 to Sum of the exponent terms, It is the sample number. and It is a category label. It represents the total number of categories.
[0025] The beneficial effects of the above-mentioned further scheme are: by using the calculated sample weights for weighting and combining them with residual correction terms, it is possible to eliminate low-quality information interference as much as possible under complex working conditions, preserve useful and valuable information, and thereby obtain more accurate classification results.
[0026] The beneficial effects of this invention are that by combining three-dimensional dynamic quality assessment with a consensus-aware residual gating mechanism, the confidence, consistency and beneficial differences of multi-source power grid data characteristics are uniformly characterized. This solves the problem that existing methods cannot maintain robustness to low-quality information and conflicting data while ensuring that the decision-making process is clear and traceable. It effectively improves the robustness of the method in scenarios with low-quality information, information conflicts and missing information. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. It should be understood that the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 A flowchart illustrating the steps of a multi-feature power grid data fusion method based on dynamic quality estimation provided in an embodiment of the present invention. Detailed Implementation
[0029] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.
[0030] like Figure 1 As shown, the present invention provides a multi-feature power grid data fusion method based on dynamic quality estimation, the implementation method of which is as follows: S1: Processes various input power data and calculates view-level classification probability output.
[0031] S11: Construct an independent subclassifier for each feature view and compute class logits.
[0032] In this embodiment, in step S11, for each feature view Build independent subclassifiers , will the Raw / preprocessed features of each sample on this view After input, the categorical logits of the view are obtained: here, Indicates the sample index (from) arrive ), Indicates the view index (from) arrive ), For the sample In view The input feature vector, The feature dimension of this view; It maps the view features to A subclassifier of the dimensional evidence space, outputting... Each dimension corresponds to an unnormalized score for a candidate category.
[0033] S12: Calculate the posterior probability distribution using the activation function Softmax.
[0034] The calculation formula in step S12 is as follows: In this formula, Indicates sample In view Logarithmic evidence vector on, For the sample In view The class posterior vector on, components Indicates belonging to the first Class probability, sum of denominators, index Traverse all One category; yes The Each component provides an independent probabilistic assessment of each sample from each view, preparing for subsequent quality evaluation and fusion.
[0035] S2: Perform sample-level three-dimensional dynamic quality assessment based on the obtained probability output.
[0036] S21: Measure self-certainty using maximum a posteriori or entropy normalization to obtain confidence level.
[0037] The formula for calculating the confidence level in step S21 is as follows: in It is a sample In view The maximum posterior confidence level. It is the sample number. These are view labels. Indicates the category label Take the maximum value. Indicates sample In view The above belongs to the first The probability of a class Indicates sample In view Shannon entropy over probability distribution, Represents probability Take the natural logarithm. This represents the confidence level after entropy normalization. Representing entropy and Normalize the ratio and take the complement. Represents the total number of categories Take the natural logarithm. This indicates that the confidence level is between 0 and 1.
[0038] S22: Consensus is obtained by measuring the consistency with the average inner product of the probability distributions of the remaining views.
[0039] The formula for calculating the consensus level in step S22 is as follows: In the formula, It is the same sample In another view The posterior vector on, the inner product Reflects the degree of similarity between the two views in terms of category distribution; for The average of all views yields the overall consistency with the population; the closer the value is to the average, the better. This indicates that the view is becoming increasingly consistent with other views.
[0040] S23: Use Jensen–Shannon divergence (JSD) to measure beneficial differences and threshold to suppress invalid deviations, thus obtaining uniqueness.
[0041] The formula for calculating uniqueness in step S23 is as follows: in It is except for the view The mean probability of other views. It is a sample In view The posterior probability vector on; This represents the average distribution of the current view with other views. Next: It is the Kullback-Leibler divergence, which represents the difference between two probability distributions; It is a uniqueness measure that is not thresholded, representing the difference between the current view and other views; It is a uniqueness threshold used to remove invalid, minute differences, resulting in the final product. Only the differences that are useful for classification are retained.
[0042] S3: Implement consensus-aware residual gating, and obtain sample weights through multiplicative gating and normalization.
[0043] S31: Generate sample-level differential gating weights from the calculated three-dimensional metric components.
[0044] The formula for calculating the weights in step S31 is as follows: in Indicates sample In view The unnormalized gate value below indicates the importance of the view; These are hyperparameters that control the confidence level. Consensus and uniqueness The relative importance of gating.
[0045] S32: Normalize the weights using the temperature coefficient.
[0046] The normalization formula in step S32 is as follows: in, It is a sample view The final fusion weights, temperature parameters Control the sharpness of the weight distribution. For the sample In view The unnormalized gate value on. When When taking the smaller value, the larger one This will be highlighted more strongly; the denominator is indexed by all views. Summation is performed to complete the normalization, making the total weights equal to... .
[0047] S33: Filter out complementary information that is worth keeping.
[0048] The formula for calculating complementary information in step S33 is as follows: Indicates a fixed sample Excluded View The arithmetic mean of the logarithmic evidence vectors of the remaining views. Indicates to Calculate the average coefficient for each view. Indicates sample In view Logarithmic evidence vector on, Indicates sample In view The residual correction term on, Represents the residual strength coefficient and satisfies , Indicates sample In view Uniqueness score, This represents the difference vector between the current view and the mean of the remaining views.
[0049] S4: Dynamically fuse multiple power data based on the calculated sample weights.
[0050] S41: Based on the sample-level weights and complementary information obtained above, the fused classification information is obtained.
[0051] The formula for calculating the fused classification information in step S41 is as follows: in It is a sample The fused logarithmic evidence vector, This represents the logarithmic evidence vector for each view. By weight Vector-wise weighted summation This indicates that the residual correction term for each view is... Summation of vectors one by one It is the sample number. These are view labels. This is the total number of views.
[0052] S42: The classification information is transformed into a predicted probability distribution through the Softmax activation function.
[0053] The formula for calculating the predicted probability distribution in step S42 is as follows: in It is a sample The fused posterior probability vector This represents a mapping that performs exponential normalization on the fused logarithmic evidence vector. Indicates sample The fusion post-hoc in the first The probability of each category Represented by natural constant The first piece of evidence for the fusion of logarithmic evidence Each component Take the index. Indicates the category label From 1 to Sum of the exponent terms, It is the sample number. and It is a category label. It represents the total number of categories.
[0054] In the embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether to implement it in hardware or software depends on the specific application and design constraints; those skilled in the art can choose different implementation paths for different applications, but this should not be considered as exceeding the scope of this invention.
[0055] Furthermore, the method provided by this invention can also be implemented in other ways. The division of the above steps and sub-steps is based on a schematic description of logical functions. In actual engineering implementation, equivalent re-divisions can be made, such as refining it into more steps and processing units, or omitting certain steps without affecting the overall function.
[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made without departing from the spirit and principles of the present invention should be included within the scope of protection of the present invention.
[0057] This invention provides a multi-feature power grid data fusion method based on dynamic quality estimation, which solves the problem of balancing robustness and computational efficiency while ensuring interpretability in the field of multi-feature power analysis.
Claims
1. A multi-feature power grid data fusion method based on dynamic quality estimation, characterized in that, The method includes: S1: Processes various input power data and calculates view-level classification probability output; S2: Perform sample-level three-dimensional dynamic quality assessment based on the obtained probability output; S3: Implement consensus-aware residual gating, and obtain sample weights through multiplicative gating and normalization; S4: Dynamically fuse multiple power data based on the calculated sample weights; Step S1 specifically includes: S11: Construct an independent sub-classifier for each feature view and compute class logits. S12: Calculate the posterior probability distribution using the activation function Softmax; Step S2 specifically includes: S21: Measure self-certainty using maximum a posteriori and entropy normalization to obtain confidence level; S22: Consensus is obtained by measuring the consistency with the average inner product of the probability distributions of the remaining views; S23: Use Jensen–Shannon divergence (JSD) to measure beneficial differences and threshold to suppress invalid deviations to obtain uniqueness; Step S3 specifically includes: S31: Generate sample-level differentiated gating weights from the calculated three-dimensional metric components; S32: Normalize the weights using the temperature coefficient; S33: Filter complementary information that is worth keeping; Step S4 specifically includes: S41: Based on the sample-level weights and complementary information obtained above, the fused classification information is obtained; S42: The classification information is transformed into a predicted probability distribution through the Softmax activation function; In step S2, a sample-level three-dimensional dynamic quality assessment is performed based on the obtained probability output. The specific calculation formula is as follows: First, the confidence level is: in It is a sample In view The maximum posterior confidence level. It is the sample number. These are view labels. Indicates the category label Take the maximum value. Indicates sample In view The above belongs to the first The probability of a class Indicates sample In view Shannon entropy over probability distribution, Represents probability Take the natural logarithm. This represents the confidence level after entropy normalization. Representing entropy and Normalize the ratio and take the complement. Represents the total number of categories Take the natural logarithm. This indicates that the confidence level is between 0 and 1; The consensus calculation formula is as follows: in It is a sample In view Consensus on It is the sample number. These are view labels. Indicates the total number of views. The reciprocal of the result after subtracting 1 is used as the average coefficient. It is the total number of views. It is the label of another view. It is a sample In view The posterior probability vector on, Indicates sample In view With View The dot product of the probability vectors, Indicates the range of values for the degree of consensus; The formula for calculating uniqueness is as follows: in It is except for the view The mean probability of other views. It is a sample In view The posterior probability vector on; It is a sample view Average distribution with other views; It is the Kullback-Leibler divergence, which represents the difference between two probability distributions; It is a uniqueness measure that is not thresholded, representing the difference between the current view and other views; It is a uniqueness threshold used to remove invalid, minute differences, resulting in the final product. Only retain the differences that are useful for classification. It is the sample number. It is the view label.
2. The multi-feature power grid data fusion method based on dynamic quality estimation according to claim 1, characterized in that, In step S31, sample-level differential gating weights are generated based on the calculated three-dimensional metric components. The calculation formula is as follows: in Indicates sample In view The unnormalized gate value below indicates the importance of the view; These are hyperparameters that control the confidence level. Consensus and uniqueness The relative importance of gating It is the sample number. It is the view label.
3. The multi-feature power grid data fusion method based on dynamic quality estimation according to claim 1, characterized in that, In steps S33 and S41, beneficial complementary residual correction terms are obtained based on the calculated uniqueness and classification information, and the calculated residual correction terms are used to guide decision-making. First, the formula for calculating the residual correction terms in S33 is as follows: Indicates a fixed sample Excluded View The arithmetic mean of the logarithmic evidence vectors of the remaining views. Indicates to Calculate the average coefficient for each view. Indicates sample In view Logarithmic evidence vector on, Indicates sample In view The residual correction term on, Represents the residual strength coefficient and satisfies , Indicates sample In view Uniqueness score, This represents the difference vector between the current view and the mean of all other views. It is the sample number. It is the view label; Next, in S41, the calculated residual correction term is used to guide decision fusion, as shown in the following formula: in It is a sample The fused logarithmic evidence vector, This represents the logarithmic evidence vector for each view. By weight Vector-wise weighted summation This indicates that the residual correction term for each view is... Summation of vectors one by one It is the sample number. These are view labels. This is the total number of views.