A new entity credit evaluation method for power grid fusing double time scales

Abstract

The application discloses a kind of new entity credit evaluation methods of fusion double time scale of power grid, belong to the technical field of power system trusted access and network security.This method takes the continuous operation data of photovoltaic access entity as input, collects observation such as active power, temperature, humidity, total horizontal irradiance and scattered horizontal irradiance;Through the prediction consistency model of fusion iTransformer, LSTM, cross attention mechanism and KAN, the power behavior baseline under normal operating conditions is established, and the daily behavior abnormal intensity is generated according to the deviation between predicted power and actual power;Again, based on the normal sample reference distribution, the abnormal intensity is mapped to percentile, and positive evidence and negative evidence are constructed;Finally, the reputation value is updated using double time scale Beta recursion mechanism, and the dynamic reputation result with abnormal fast response capability and historical memory capability is obtained.The application can continuously model and trustably evaluate the operation behavior of new power entities such as photovoltaic after access, and can be used for access control, risk warning and abnormal disposal.

Description

[0001] This invention relates to the field of trusted access and network security technology in power systems, specifically to a novel power grid entity reputation assessment method that integrates dual time scales. In particular, it relates to a method for modeling, extracting abnormal evidence, and dynamically assessing the continuous operating behavior of novel power entities such as photovoltaic, wind power, energy storage, charging piles, and electric vehicle aggregates based on predictive consistency analysis and a dual time scale Beta recursion mechanism. Background Technology

[0002] With the advancement of green and low-carbon energy transformation and the construction of new power systems, new entities such as photovoltaics, wind power, energy storage, charging piles, electric vehicles, and their aggregates are being connected to the power grid on a large scale. These new entities are typically characterized by their large number, wide distribution, flexible access methods, and significant dynamic changes in operating status. Their operation involves not only the physical connection of traditional power equipment but also multiple stages such as communication interaction, data acquisition, dispatch control, and edge collaboration. As the scale of access continues to expand, the power grid system is gradually transforming from the traditional centralized energy supply model to a new type of power system with deep cyber-physical integration.

[0003] The integration of new entities into the power grid can improve energy efficiency and system operational flexibility, but it also exposes the grid to more complex security risks. On the one hand, entities such as distributed renewable energy sources, energy storage devices, and charging facilities typically rely on communication links to interact with dispatch platforms, edge gateways, or aggregated control systems, making the communication layer, control layer, and terminal layer potential entry points for attacks. On the other hand, the operational behavior of these new entities is closely related to external environments, dispatch instructions, and equipment status. Their output power, charging and discharging behavior, or load changes exhibit significant temporal, random, and environmental coupling characteristics, making it difficult to accurately reflect the reliability of their continuous operation based solely on their status at a single moment.

[0004] Existing power grid access security mechanisms primarily rely on identity authentication, access control, whitelist management, or threshold alarms. While these mechanisms can verify the legitimacy of an entity's identity at the initial stage of access or trigger alarms when certain indicators exceed preset thresholds, they still have certain limitations. First, static identity authentication can only prove that an entity is qualified to access at a certain moment, making it difficult to determine whether the entity maintains trustworthy behavior during continuous operation after access. Second, the results of a single anomaly detection usually only reflect the abnormal state at a local point in time or within a local time period, failing to provide a continuous evaluation of the entity's long-term operational behavior and failing to reflect the sustained impact of abnormal behavior on subsequent trustworthiness.

[0005] Taking photovoltaic (PV) grid-connected entities as an example, their output power is affected by factors such as solar irradiance, ambient temperature, humidity, season, weather changes, and historical operating conditions. Under normal circumstances, PV power variations should maintain a certain consistency with historical power patterns and environmental observation conditions. However, when equipment failures, abnormal shading, data tampering, false data injection, or malicious control occur, the actual output power may deviate significantly from the normal operating mode. Therefore, behavioral evidence reflecting the credibility of an entity's operation can be extracted by establishing a behavioral baseline under normal operating conditions and comparing the consistency between actual and predicted output. However, existing PV power prediction or anomaly detection methods focus more on prediction accuracy or anomaly identification itself, and rarely further transform prediction deviations into recursive, accumulative, and recoverable reputation values, making it difficult to directly serve the continuous and reliable grid-connected management of new grid entities.

[0006] Furthermore, while existing reputation or trust assessment methods can quantify an entity's historical behavior to some extent, they still have shortcomings in new power grid entity integration scenarios. Some methods employ simple weighted scoring or static index fusion, which struggle to handle the continuous changes in entity behavior over time. While some methods introduce time decay mechanisms, they lack an effective balance between rapid response to recent anomalies, suppression of short-term noise, and the ability to remember historical anomalies. In practical operation, the reputation system needs to both rapidly reduce the reputation value when an entity experiences sustained anomalies and prevent the reputation value from immediately returning to a high level after the anomaly is resolved, thus preserving the impact of historical breaches of trust on subsequent evaluations.

[0007] Therefore, there is an urgent need for a novel power grid entity reputation assessment method that integrates dual time scales. This method should be able to combine continuous operational data of the entity with external environmental observation data to establish a baseline for normal behavior, extract interpretable evidence of abnormal behavior, and continuously update the entity's reputation value through a dynamic recursive mechanism. This method should be able to balance rapid response to anomalies, stable normal operation, and the ability to remember historical behavior, providing a reliable basis for the access control, risk warning, anomaly handling, and dynamic management of novel power grid entities. Summary of the Invention

[0008] The purpose of this invention is to provide a novel power grid entity reputation assessment method that integrates dual time scales, aiming to overcome the shortcomings of existing technologies where static identity authentication cannot continuously reflect the trustworthy state of entity operation, and single anomaly detection results cannot characterize the long-term reputation evolution process. By constructing a reputation assessment framework based on predictive consistency analysis and a dual-time-scale Beta recursive mechanism, the continuous operational behavior of novel entities is transformed into quantifiable, recursive, and recoverable dynamic reputation values, thereby improving the power grid's continuous perception capability of abnormal access behavior, risk of breach of trust, and operational disturbances.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a novel power grid entity reputation assessment method integrating dual time scales, the reputation assessment method comprising the following steps:

[0010] S1: Collect operational observation data of new entities connected to the power grid during continuous operation and construct the original operational observation sequence;

[0011] S2: Perform time alignment, missing value processing, and daily segmentation on the original operational observation sequence to obtain a daily-level behavioral observation sequence;

[0012] S3: Construct a prediction consistency model, input the daily behavior observation sequence into the prediction consistency model, and obtain the predicted output power of the novel entity under normal operating conditions;

[0013] S4: Calculate the daily behavior deviation intensity of the novel entity based on the deviation between the predicted output power and the actual output power;

[0014] S5: Based on the normal sample reference distribution, perform percentile mapping on the daily-level behavior deviation intensity to obtain the daily-level anomaly degree;

[0015] S6: Construct positive and negative evidence based on the daily anomaly level and the credibility of the prediction consistency model;

[0016] S7: Based on the positive and negative evidence, the reputation value of the novel entity is updated using a dual-timescale Beta recursive mechanism to obtain a dynamic reputation assessment result that characterizes the credibility of the novel entity's continued operation.

[0017] 2. The reputation assessment method for novel power grid entity access behavior according to claim 1, characterized in that step S1 specifically includes the following steps:

[0018] S1.1 Collect output power data and environmental observation data of the novel entity during continuous operation;

[0019] S1.2 When the novel entity is a photovoltaic access entity, the operational observation data includes date, active power, ambient temperature, relative humidity, total horizontal irradiance, and diffuse horizontal irradiance;

[0020] S1.3 defines the observation vector at the t-th sampling time as: ,in For power, For ambient temperature, Relative humidity, Total horizontal irradiance, This represents the horizontal irradiance of the scattered radiation.

[0021] S1.4 defines a continuously running observation sequence as... ,in This represents the original operational observation sequence. This represents the total number of sampling times.

[0022] 3. The reputation assessment method for novel power grid entity access behavior according to claim 2, characterized in that step S2 specifically includes the following steps:

[0023] S2.1 Divide the original operational observation sequence into multiple daily-level behavioral observation sequences according to a preset time granularity;

[0024] S2.2 Let the first The daily behavior observation sequence of the day is ,in Indicates the first A daily behavior observation sequence for the day. This indicates the number of sampling points within a single day;

[0025] S2.3 When hourly sampling is used, take Then the first The daily behavior observation sequence of the day is represented as

[0026] S2.4 will be the first The power sequence and environmental covariate sequence for each day are defined as follows: , ,in Indicates the first The power sequence of the day, Indicates the first The sequence of environmental covariates for days, and: ,in Indicates the first The first day Each sampling time.

[0027] 4. The reputation assessment method for novel power grid entity access behavior according to claim 3, characterized in that step S3 specifically includes the following steps:

[0028] S3.1 uses a sliding time window to construct the prediction samples, assuming the time window length is... Then the first Each predicted sample is represented as The target of the prediction is ,in Indicates the first One predicted sample, Indicates the length of the sliding time window. This indicates the actual output power at the next moment;

[0029] S3.2 Divide the predicted sample into target behavior sequence and environmental covariate sequence. , ,in Represents the historical output power sequence. Represents a sequence of environmental covariates;

[0030] The prediction consistency model described in S3.3 includes a target behavior branch, an environment-driven branch, a cross-stream attention fusion module, and a nonlinear mapping module;

[0031] The target behavior branch described in S3.4 uses iTransformer to process the historical output power sequence. Encode to obtain target behavior features ,in This represents the target behavior characteristics corresponding to the historical output power sequence. This represents the iTransformer encoding function;

[0032] The environment-driven branch described in S3.5 uses LSTM to analyze the environmental covariate sequence. Encode to obtain environment-driven features ,in This represents the environmental driving characteristics corresponding to the environmental covariate sequence. This represents the LSTM encoding function.

[0033] S3.6: The target behavioral characteristics and the environmental driving features Input the cross-stream attention fusion module to obtain cross-stream fusion features. ,in Indicates cross-stream fusion characteristics, Represents the cross-stream attention fusion function;

[0034] S3.7: Input the cross-flow fusion feature into the nonlinear mapping module to obtain the predicted output power. ,in Indicates the first The predicted output power at each sampling time. This represents the KAN nonlinear mapping function.

[0035] 5. The reputation assessment method for novel power grid entity access behavior according to claim 4, characterized in that the target behavior branch in step S3 specifically includes the following steps:

[0036] S3.4.1 Maps the historical output power sequence to the hidden feature space to obtain the initial embedding representation. ,in The initial embedding representation of the historical output power sequence. Represents the power sequence embedding mapping function;

[0037] S3.4.2: In the first In the iTransformer layer encoding structure, the self-attention output is calculated. , , , , , ,in , , They represent the first The query matrix, key matrix, and value matrix of the layer. , , These represent the learnable projection matrices, Indicates the dimension of the key vector;

[0038] S3.4.3 obtains the first... Layer output ,in Indicates a feedforward network;

[0039] S3.4.4 after After layer encoding, the target behavior features are obtained. ,in Indicates the number of encoding layers in the iTransformer. This represents the temporal behavior characteristics of historical output power sequences.

[0040] 6. The reputation assessment method for novel power grid entity access behavior according to claim 4, characterized in that the environment-driven branch in step S3 specifically includes the following steps:

[0041] S3.5.1 Section The environmental observation vector at each sampling time is ,in Indicates ambient temperature. Indicates relative humidity. Indicates total horizontal irradiance. Indicates the horizontal irradiance of the scattered light;

[0042] S3.5.2 LSTM is used to process the environmental observation sequence. Perform recursive encoding for the th Environmental input at time 1 Its recursive process is as follows , , , , , ,in Indicates the input gate. Represents the Gate of Oblivion Indicates the output gate. Indicates the candidate cell state. Indicates the cell state. Indicates a hidden state. This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication;

[0043] S3.5.3 outputs the hidden state sequence within the time window as the environment-driven branch. ,in This indicates the timing environment characteristics of the output of the environment-driven branch.

[0044] 7. A reputation assessment method for novel entity access behavior in the power grid according to claim 4, characterized in that the cross-flow attention fusion module specifically includes the following steps:

[0045] S3.6.1 Based on the target behavioral characteristics As a source of query, the environmental driving features As the source of keys and values, the query matrix, key matrix, and value matrix of cross-stream attention are constructed. , , ,in , , These represent the query matrix, key matrix, and value matrix in cross-stream attention, respectively. , , These represent the learnable projection matrices in cross-current attention;

[0046] S3.6.2 Calculate the cross-stream attention weights based on the query matrix, key matrix, and value matrix. ,in Indicates cross-stream attention weights. This represents the dimension of the key vector in cross-stream attention;

[0047] S3.6.3 Calculate the cross-stream fusion features based on the cross-stream attention weights and the value matrix. ,in Indicates cross-stream fusion characteristics;

[0048] S3.6.4 The conditional correlation between the historical output power sequence and the environmental observation sequence is established through the cross-flow fusion feature, so that the prediction consistency model can dynamically reweight the historical power behavior features according to the environmental state.

[0049] 8. A reputation assessment method for novel power grid entity access behavior according to claim 4, characterized in that the nonlinear mapping module specifically includes the following steps:

[0050] S3.7.1 The cross-stream fusion features are pooled or vectorized to obtain the fusion feature vector. ,in Represents the fused feature vector. Indicates pooling or vectorization operations;

[0051] S3.7.2 A KAN network is used to perform a nonlinear mapping on the fused feature vector, wherein the KAN network's first... Layer Each output node is represented as ,in, Indicates the KAN network's... Layer One output node, Indicates the first The number of nodes in the layer Indicates the first Layer One input node, Indicates the connection of the first Layer The node and the first Layer A learnable unary function for each node;

[0052] S3.7.3 After multi-layer KAN nonlinear mapping, the predicted power at the next sampling time is output. ,in Indicates the first The predicted output power at each sampling time. Represents the KAN network mapping function;

[0053] S3.7.4 represents the overall mapping process of the prediction consistency model as follows: ,in This represents the branch encoding function for the target behavior. This indicates an environment-driven branching coding function. This represents the cross-stream attention fusion function.

[0054] 9. A reputation assessment method for novel power grid entity access behavior according to claim 1, characterized in that step S4 specifically includes the following steps:

[0055] S4.1 Calculate the first The mean absolute error between predicted output power and actual output power within a day ,in Indicates the first The mean absolute error of the day Indicates the first Heavenly The actual output power at each sampling time. Indicates the first Heavenly Predicted output power at each sampling time;

[0056] S4.2: Calculate the first... Daily power fluctuation range ,in express The maximum actual output power of the day. Indicates the first The minimum actual output power of the day;

[0057] S4.3 Based on the mean absolute error and the power fluctuation range, calculate the daily behavior deviation intensity. ,in, Indicates the first The intensity of the daily-level behavior deviation. This represents a constant used to prevent the denominator from being zero.

[0058] 10. A reputation assessment method for novel power grid entity access behavior according to claim 1, characterized in that step S5 specifically includes the following steps:

[0059] S5.1 Obtain the set of daily behavioral deviation intensities under normal sample conditions. ,in, This represents the set of behavioral deviations in the reference distribution of normal samples. Indicates the first The intensity of behavioral deviation from normal samples Indicates the number of normal samples;

[0060] S5.2 Construct an empirical cumulative distribution function based on the normal sample reference distribution. ,in This represents the empirical cumulative distribution function corresponding to the reference distribution of normal samples. This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0061] S5.3 will be the first Daily behavior deviation intensity Input the empirical cumulative distribution function to obtain the daily anomaly level. ,in Indicates the first The degree of daily anomaly;

[0062] S5.4 when The closer it is to 1, the more significant it is. The more the operation of the day deviates from the normal sample reference distribution; when The closer it is to 0, the more significant it is. The closer the daily operating behavior is to the normal sample reference distribution.

[0063] 11. A reputation assessment method for novel power grid entity access behavior according to claim 1, characterized in that step S6 specifically includes the following steps:

[0064] S6.1 Based on the prediction error of the prediction consistency model in the offline testing phase, construct the credibility of the prediction consistency model;

[0065] S6.2 Calculate the mean absolute error and root mean square error of the prediction consistency model on the offline test set, respectively. , ,in This represents the mean absolute error of the offline test set. This represents the root mean square error of the offline test set. Indicates the number of offline test samples. Indicates the actual output power. Indicates the predicted output power;

[0066] S6.3 Calculate the confidence level of the prediction consistency model based on the mean absolute error and root mean square error. ,in Indicates the reliability of the prediction consistency model;

[0067] S6.4 Construct negative evidence based on the daily anomaly level and the reliability of the prediction consistency model. ,in, Indicates the first Negative evidence from heaven;

[0068] S6.5 Construct positive evidence based on the complementary value of the daily anomaly level and the credibility of the prediction consistency model. ,in, Indicates the first The true evidence of Heaven.

[0069] 12. The reputation assessment method for novel power grid entity access behavior according to claim 1, characterized in that step S7 specifically includes the following steps:

[0070] S7.1 Set the fast timescale Beta recursion parameters and the slow timescale Beta recursion parameters;

[0071] S7.2 will be the first in the fast time scale The cumulative strength of positive evidence and the cumulative strength of negative evidence for each day are defined as follows: and , the first under a slow time scale The cumulative strength of positive evidence and the cumulative strength of negative evidence for each day are defined as follows: and ;

[0072] S7.3 updates the fast timescale Beta parameter using the following formula. , ,in Indicates the forgetting factor on a fast timescale;

[0073] S7.4 updates the slow timescale Beta parameter using the following formula. , ,in Represents the slow-timescale forgetting factor, and

[0074] S7.5 Calculates Fast-Scale Reputation Values ​​Based on Fast-Time-Scale Beta Parameters

[0075] S7.6: Calculate the slow-scale reputation value based on the slow-timescale Beta parameter. ,in Indicates the first The rapid scale of the credit rating of the day Indicates the first The slow-scale reputation value of the day;

[0076] S7.7: The smaller of the fast-scale reputation value and the slow-scale reputation value is taken as the final reputation value. ,in Indicates the first The final reputation value of the new entity in the day.

[0077] 13. A reputation assessment method for novel power grid entity access behavior according to claim 12, characterized in that the fast time-scale Beta recursion mechanism is used to improve the response speed of the reputation value to recent abnormal behavior, and the slow time-scale Beta recursion mechanism is used to retain the impact of historical abnormal behavior on the reputation value.

[0078] 14. A reputation assessment method for novel power grid entity access behavior according to claim 12, characterized in that, when the novel entity exhibits continuous abnormal behavior, the negative evidence... Continuous increase makes and Increase, thereby increasing the final reputation value The positive evidence decreases; when the novel entity resumes normal operation, the positive evidence... Increase, making and Increase, thereby increasing the final reputation value Gradual recovery.

[0079] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0080] 1. This invention proposes a novel power grid entity reputation assessment method that integrates dual time scales. By converting continuous operational data of entities into daily behavioral evidence and further recursively extrapolating it into dynamic reputation values, it extends the assessment from "static identity credibility" to "continuous operational credibility." This method can continuously evaluate the operational behavior of new power entities such as photovoltaic, wind power, energy storage, charging piles, and electric vehicle aggregates after their integration, providing quantitative basis for access control, risk warning, and anomaly handling.

[0081] 2. This invention employs a behavior deviation assessment mechanism based on predictive consistency. By integrating historical output power and environmental observation data, it establishes a behavioral baseline for novel entities under normal operating conditions. For photovoltaic grid-connected entities, this mechanism can comprehensively utilize data such as active power, ambient temperature, relative humidity, total horizontal irradiance, and diffuse horizontal irradiance to determine whether the actual output power is consistent with the normal environmental response pattern, thereby improving the accuracy and interpretability of abnormal behavior evidence extraction.

[0082] 3. This invention constructs a prediction consistency model that integrates iTransformer, LSTM, cross-current attention mechanism, and KAN. Specifically, iTransformer is used to extract temporal features from historical power sequences, LSTM is used to characterize the continuous changes in environmental variables, the cross-current attention mechanism is used to achieve interactive fusion between power behavior features and environmental driving features, and KAN is used to complete the nonlinear mapping from fused features to predicted power. This structure enhances the model's ability to represent the normal operation modes of novel entities.

[0083] 4. This invention employs a percentile mapping method based on the reference distribution of normal samples to convert the original prediction error into a dimensionless anomaly level, reducing the impact of differences in error scales under different dates, power amplitudes, and environmental conditions on the reputation assessment results. This approach enables the anomaly level to more accurately reflect the deviation of the current behavior from normal operating samples, reducing the risk of false alarms or false negatives in complex operating conditions caused by fixed threshold methods.

[0084] 5. This invention constructs positive and negative evidence based on the degree of anomaly and the reliability of the prediction consistency model, enabling both normal and abnormal behaviors to participate in the reputation update process. Positive evidence is used to support reputation recovery, while negative evidence is used to characterize the accumulation of dishonest behavior. Simultaneously, the influence of the evidence is adjusted through model reliability, thereby avoiding excessive impact of low-reliability model outputs on the reputation value.

[0085] 6. This invention employs a dual-timescale Beta recursive mechanism to update reputation values. Fast timescale recursion enhances the system's rapid response to recent anomalous behavior, while slow timescale recursion preserves the continued impact of historical behavior on reputation status. The smaller of the fast and slow timescale reputation values ​​is used as the final reputation value. This mechanism can promptly reduce the reputation value when persistent anomalies occur and achieve gradual recovery after the anomalies are resolved, balancing anomaly response speed, operational stability, and historical memory capabilities.

[0086] 7. The credit assessment results of this invention can be used for the dynamic access management of new entities in the power grid. When an entity's credit value continuously decreases, risk alarms, manual review, increased sampling frequency, or restricted access permissions can be triggered; when an entity maintains a high credit status for a long period, normal access can be maintained. This method can improve the power grid's ability to continuously perceive and handle abnormal access behavior and credit default risks of new entities. Attached Figure Description

[0087] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0088] Figure 1 is Overall framework diagram of the present invention.

[0089] Figure 2 is The method flowchart of the present invention. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0091] like Figure 1As shown, the overall framework of this invention includes an offline model building stage and an online reputation assessment stage. In the offline model building stage, normal photovoltaic time-series data is input into the prediction consistency module for training, learning the correlation between historical power state and environmental observation conditions under normal operating conditions, and obtaining and saving the prediction consistency model. Simultaneously, the mean absolute error (MAE) and root mean square error (RMSE) of the model are calculated based on the offline test set, and the reliability parameters of the prediction consistency model are determined accordingly. In the online reputation assessment stage, the observed photovoltaic time-series data of the photovoltaic entity to be evaluated is input into the saved prediction consistency model to obtain the predicted output power. The predicted output power is compared with the actual output power through online calculation to obtain the daily anomaly level reflecting the degree of deviation from the current operating behavior. Further, positive and negative evidence are constructed by combining the daily anomaly level and the model reliability, and the positive and negative evidence are input into the fast-scale Beta recursor and the slow-scale Beta recursor, respectively, to obtain the fast-scale reputation value and the slow-scale reputation value. Finally, the smaller of the two values ​​is taken as the reputation output for day d. This achieves a complete integration of offline learning of normal behavioral baselines, online detection of operational deviations, construction of positive and negative evidence, and dual-timescale reputation recursion.

[0092] like Figure 2 As shown in this embodiment, a novel power grid entity reputation assessment method integrating dual time scales includes the following steps:

[0093] S1: Collect operational observation data of new entities connected to the power grid during continuous operation and construct the original operational observation sequence;

[0094] S2: Perform time alignment, missing value processing, and daily segmentation on the original operational observation sequence to obtain a daily-level behavioral observation sequence;

[0095] S3: Construct a prediction consistency model, input the daily behavior observation sequence into the prediction consistency model, and obtain the predicted output power of the novel entity under normal operating conditions;

[0096] S4: Calculate the daily behavior deviation intensity of the novel entity based on the deviation between the predicted output power and the actual output power;

[0097] S5: Based on the normal sample reference distribution, perform percentile mapping on the daily-level behavior deviation intensity to obtain the daily-level anomaly degree;

[0098] S6: Construct positive and negative evidence based on the daily anomaly level and the credibility of the prediction consistency model;

[0099] S7: Based on the positive and negative evidence, the reputation value of the novel entity is updated using a dual-timescale Beta recursive mechanism to obtain a dynamic reputation assessment result that characterizes the credibility of the novel entity's continued operation.

[0100] In some embodiments, the novel entity is a photovoltaic (PV) grid-connected entity. During grid-connected operation, the output power of the PV grid-connected entity is influenced by factors such as historical power status, ambient temperature, humidity, total horizontal irradiance, and diffuse horizontal irradiance. By continuously collecting and analyzing the aforementioned multi-source observation data, a dynamic reputation value reflecting the reliability of the PV entity's operation can be obtained.

[0101] In some embodiments, the process of performing step S1 includes:

[0102] S1.1 Collects output power data and environmental observation data of the novel entity during continuous operation.

[0103] In this embodiment, operational observation data of the photovoltaic grid-connected entity is collected through a photovoltaic inverter monitoring system, environmental sensors, grid-side acquisition devices, or a dispatch-side data platform. The operational observation data includes date, active power, ambient temperature, relative humidity, total horizontal irradiance, and diffuse horizontal irradiance.

[0104] S1.2 When the novel entity is a photovoltaic access entity, the first... The observation vector at each sampling time is defined as ,in Indicates the first Active power at each sampling time. Indicates ambient temperature. Indicates relative humidity. Indicates total horizontal irradiance. This indicates the horizontal irradiance of the scattered radiation.

[0105] S1.3 defines a continuously running observation sequence as... ,in This represents the original operational observation sequence. This represents the total number of sampling times.

[0106] In some embodiments, step S2 specifically includes the following steps:

[0107] S2.1 Time alignment is performed on the original running observation sequence so that different variables form a unified observation vector at the same sampling time.

[0108] 2.2 Processing missing values ​​in the original operational observation sequence. For short-term missing data, interpolation between adjacent time points or mean imputation can be used to complete the data; for data with consecutive missing periods of too long, the data for the corresponding date can be removed to avoid affecting the subsequent calculation of behavioral deviation intensity.

[0109] S2.3 Divide the original operational observation sequence into multiple daily behavioral observation sequences according to a preset time granularity. Let the first... The daily behavior observation sequence of the day is ,in Indicates the first A daily behavior observation sequence for the day. This indicates the number of sampling points within a single day.

[0110] S2.4 In one specific embodiment, an hourly sampling method is used to take... Then the first The daily behavior observation sequence of the day is represented as

[0111] S2.5 will be the first The power sequence and environmental covariate sequence for each day are defined as follows: , ,in Indicates the first The power sequence of the day, Indicates the first The sequence of environmental covariates for the day, and ,in Indicates the first The first day Each sampling time.

[0112] In some embodiments, step S3 specifically includes the following steps:

[0113] S3.1 uses a sliding time window to construct the prediction samples. Let the time window length be... , No. Each predicted sample is represented as The target of the prediction is ,in Indicates the first One predicted sample, Indicates the length of the sliding time window. This indicates the actual output power at the next moment.

[0114] In one specific embodiment, take This means using historical operating data from the previous 24 hours to predict the active power output for the next hour.

[0115] S3.2 Divide the predicted sample into target behavior sequence and environmental covariate sequence. , ,in Represents the historical output power sequence. This represents a sequence of environmental covariates.

[0116] S3.3 Constructing a prediction consistency model. The prediction consistency model includes a target behavior branch, an environment-driven branch, a cross-current attention fusion module, and a nonlinear mapping module. The target behavior branch is used to extract the temporal behavior features of the historical output power sequence; the environment-driven branch is used to extract the temporal environment features of the environmental covariate sequence; the cross-current attention fusion module is used to establish the conditional correlation between the historical output power features and the environmental features; and the nonlinear mapping module is used to output the predicted power based on the fused features.

[0117] S3.3.1 In some specific embodiments, the target behavior branch uses iTransformer to process the historical output power sequence. Encode to obtain target behavior features ,in This represents the target behavior characteristics corresponding to the historical output power sequence. This represents the iTransformer encoding function.

[0118] Specifically, the historical output power sequence is first mapped to the hidden feature space to obtain the initial embedding representation. ,in The initial embedding representation of the historical output power sequence. This represents the power sequence embedding mapping function.

[0119] In the In the iTransformer layer encoding structure, the self-attention output is calculated. , , , , ,in , , They represent the first The query matrix, key matrix, and value matrix of the layer. , , These represent the learnable projection matrices, Indicates the dimension of the key vector;

[0120] The first is obtained through the feedforward network. Layer output ,in Represents a feedforward network; after After layer encoding, the target behavior features are obtained. ,in Indicates the number of encoding layers in the iTransformer. This represents the temporal behavior characteristics of historical output power sequences.

[0121] S3.3.2 In some specific embodiments, the environment-driven branch uses LSTM on the environmental covariate sequence. Encode to obtain environment-driven features ,in This represents the environmental driving characteristics corresponding to the environmental covariate sequence. This represents the LSTM encoding function.

[0122] For the Environmental input at time 1 The recursive process of LSTM is as follows: , , , , , ,in Indicates the input gate. Represents the Gate of Oblivion Indicates the output gate. Indicates the candidate cell state. Indicates the cell state. Indicates a hidden state. This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication;

[0123] The hidden state sequence within the time window is output as the environment-driven branch. ,in This indicates the timing environment characteristics of the output of the environment-driven branch.

[0124] S3.3.3 In some specific embodiments, the cross-stream attention fusion module uses target behavioral features As a source of queries, environmentally driven features As the source of keys and values, the query matrix, key matrix, and value matrix of cross-stream attention are constructed. , , ,in , , These represent the query matrix, key matrix, and value matrix in cross-stream attention, respectively. , , and represent the learnable projection matrix in cross-current attention, respectively.

[0125] Calculate the cross-stream attention weights based on the query matrix, key matrix, and value matrix. ,in Indicates cross-stream attention weights. This represents the dimension of the key vector in cross-stream attention;

[0126] Calculate the cross-stream fusion features based on the cross-stream attention weights and the value matrix. ,in This represents cross-current fusion characteristics. This method allows the model to dynamically reweight historical power behavior features based on environmental conditions, thereby improving the expressive power of the normal power baseline.

[0127] In some specific embodiments of S3.3.4, the nonlinear mapping module employs a KAN network. First, the cross-stream fusion features are pooled or vectorized to obtain the fusion feature vector. ,in Represents the fused feature vector. This indicates a pooling or vectorization operation.

[0128] KAN network Layer Each output node is represented as ,in Indicates the KAN network's... Layer One output node, Indicates the first The number of nodes in the layer Indicates the first Layer One input node, Indicates the connection of the first Layer The node and the first Layer A learnable unary function for each node;

[0129] After multi-layer KAN nonlinear mapping, the predicted power at the next sampling time is output. ,in Indicates the first The predicted output power at each sampling time. Represents the KAN network mapping function;

[0130] Therefore, the overall mapping process of the prediction consistency model is expressed as: ,in This represents the branch encoding function for the target behavior. This indicates an environment-driven branching coding function. This represents the cross-stream attention fusion function.

[0131] In some embodiments, step S4 specifically includes the following steps:

[0132] S4.1 Calculate the first The mean absolute error between predicted output power and actual output power within a day ,in Indicates the first The mean absolute error of the day Indicates the first Heavenly The actual output power at each sampling time. Indicates the first Heavenly Predicted output power at each sampling time;

[0133] S4.2 Calculate the first Daily power fluctuation range ,in express The maximum actual output power of the day. Indicates the first The minimum actual output power of the day;

[0134] S4.3 Based on the mean absolute error and the power fluctuation range, calculate the daily behavior deviation intensity. ,in Indicates the first The intensity of the daily-level behavior deviation. This represents a constant used to prevent the denominator from being zero.

[0135] In some embodiments, step S5 specifically includes the following steps:

[0136] S5.1 Obtain the set of daily behavioral deviation intensities under normal sample conditions. ,in This represents the set of behavioral deviations in the reference distribution of normal samples. Indicates the first The intensity of behavioral deviation from normal samples Indicates the number of normal samples;

[0137] S5.2 Construct an empirical cumulative distribution function based on the normal sample reference distribution. ,in This represents the empirical cumulative distribution function corresponding to the reference distribution of normal samples. This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0138] S5.3 will be the first Daily behavior deviation intensity Input the empirical cumulative distribution function to obtain the daily anomaly level. ,in Indicates the first Daily anomaly level: When The closer it is to 1, the more significant it is. The more the operation of the day deviates from the normal sample reference distribution; when The closer it is to 0, the more significant it is. The closer the daily operating behavior is to the normal sample reference distribution.

[0139] In some embodiments, step S6 specifically includes the following steps:

[0140] S6.1 Based on the prediction error of the prediction consistency model in the offline testing phase, construct the credibility of the prediction consistency model;

[0141] S6.2 Calculate the mean absolute error and root mean square error of the prediction consistency model on the offline test set, respectively. , ,in This represents the mean absolute error of the offline test set. This represents the root mean square error of the offline test set. Indicates the number of offline test samples. Indicates the actual output power. Indicates the predicted output power;

[0142] S6.3 Calculate the confidence level of the prediction consistency model based on the mean absolute error and root mean square error. ,in Indicates the reliability of the prediction consistency model;

[0143] S6.4 Construct negative evidence based on the daily anomaly level and the reliability of the prediction consistency model. ,in Indicates the first Negative evidence from heaven;

[0144] S6.5 Construct positive evidence based on the complementary value of the daily anomaly level and the credibility of the prediction consistency model. ,in Indicates the first The true evidence of Heaven.

[0145] In some embodiments, step S7 specifically includes the following steps:

[0146] S7.1 Set the fast timescale Beta recursion parameters and the slow timescale Beta recursion parameters;

[0147] S7.2 will be the first in the fast time scale The cumulative strength of positive evidence and the cumulative strength of negative evidence for each day are defined as follows: and , the first under a slow time scale The cumulative strength of positive evidence and the cumulative strength of negative evidence for each day are defined as follows: and ;

[0148] S7.3 updates the fast timescale Beta parameter using the following formula. , ,in Indicates the forgetting factor on a fast timescale;

[0149] S7.4 updates the slow timescale Beta parameter using the following formula. , ,in Represents the slow-timescale forgetting factor, and

[0150] S7.5 Calculates Fast-Scale Reputation Values ​​Based on Fast-Time-Scale Beta Parameters

[0151] S7.6: Calculate the slow-scale reputation value based on the slow-timescale Beta parameter. ,in Indicates the first The rapid scale of the credit rating of the day Indicates the first The slow-scale reputation value of the day;

[0152] S7.7: The smaller of the fast-scale reputation value and the slow-scale reputation value is taken as the final reputation value. ,in Indicates the first The final reputation value of the new entity in the day.

[0153] In some specific embodiments, when the novel entity exhibits persistent abnormal behavior, the daily level of abnormality... Increase, the negative evidence Continuous increase makes and Increase, thereby increasing the final reputation value The degree of daily anomaly decreased. When the novel entity resumes normal operation, the degree of daily anomaly... Reduce, the positive evidence Increase, making and Increase, thereby increasing the final reputation value Gradual recovery.

[0154] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," 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 invention. In this specification, 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.

[0155] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A novel entity reputation assessment method for power grids integrating dual time scales, characterized in that, The credit assessment method includes the following steps: S1.1: Collect operational observation data of new entities connected to the power grid during continuous operation and construct the original operational observation sequence; S1.2: Perform time alignment, missing value processing, and daily segmentation on the original operational observation sequence to obtain a daily-level behavioral observation sequence; S1.3: Construct a prediction consistency model, input the daily behavior observation sequence into the prediction consistency model, and obtain the predicted output power of the novel entity under normal operating conditions; S1.4: Calculate the daily behavior deviation intensity of the novel entity based on the deviation between the predicted output power and the actual output power; S1.5: Based on the normal sample reference distribution, perform percentile mapping on the daily-level behavior deviation intensity to obtain the daily-level anomaly degree; S1.6: Based on the daily anomaly level and the reliability of the prediction consistency model, construct positive and negative evidence; S1.7: Based on the positive and negative evidence, the reputation value of the novel entity is updated using a dual-timescale Beta recursive mechanism to obtain a dynamic reputation assessment result that characterizes the credibility of the novel entity's continued operation.

2. The novel power grid entity reputation assessment method integrating dual time scales as described in claim 1, characterized in that, The operational observation data includes output power data and environmental observation data; when the novel entity is a photovoltaic access entity, the operational observation data includes date, active power, ambient temperature, relative humidity, total horizontal irradiance, and diffuse horizontal irradiance.

3. The novel power grid entity reputation assessment method integrating dual time scales as described in claim 1, characterized in that, In step S2, the original operational observation sequence is divided into multiple daily behavioral observation sequences according to a preset time granularity, and the power sequence and environmental covariate sequence are extracted from them.

4. The novel power grid entity reputation assessment method integrating dual time scales as described in claim 1, characterized in that, In step S3, a sliding time window is used to construct prediction samples, and the prediction samples are divided into historical output power sequences and environmental covariate sequences; the prediction consistency model includes a target behavior branch, an environment-driven branch, a cross-current attention fusion module, and a nonlinear mapping module.

5. The novel power grid entity reputation assessment method integrating dual time scales according to claim 1, characterized in that, The prediction consistency model uses iTransformer to encode the historical output power sequence, LSTM to encode the environmental covariate sequence, and a cross-current attention fusion module to fuse power behavior features and environmental driving features.

6. The novel power grid entity reputation assessment method integrating dual time scales according to claim 1, characterized in that, The nonlinear mapping module uses a KAN network to perform nonlinear mapping on cross-stream fusion features and outputs predicted output power.

7. The novel power grid entity reputation assessment method integrating dual time scales according to claim 1, characterized in that, In steps S4 to S6, the daily behavioral deviation intensity is obtained based on the deviation between the predicted output power and the actual output power, and it is converted into the daily anomaly degree based on the normal sample reference distribution; then, positive evidence and negative evidence are constructed by combining the credibility of the prediction consistency model.

8. The novel power grid entity reputation assessment method integrating dual time scales according to claim 7, characterized in that, In step S7, the reputation value is updated using Beta recursion with both fast and slow time scales, and the smaller of the fast and slow time scale reputation values ​​is taken as the final reputation value. The fast time scale is used to improve the response speed to recent abnormal behavior, while the slow time scale is used to retain the impact of historical abnormal behavior on the reputation value.

Images