User behavior association rule-driven load state recognition enhancement method, system and device, and medium

Abstract

The invention discloses a user behavior association rule-driven load state recognition enhancement method, system and device and a medium, and belongs to the technical field of load state recognition, and the method comprises the steps: forming a situation unit comprising user behavior characteristics, situation characteristics and a load state; based on a plurality of situation units, mining user behavior features and association relationships between situation features and load states, and generating a situation rule base; screening candidate load state recognition results based on the initial confidence; determining a target load state recognition result from the candidate load state recognition results; and adjusting the initial confidence coefficient to obtain an updated confidence coefficient. According to the method, the situation unit is established by combining the user power utilization behavior characteristics, the environment situation characteristics and the load measurement data, the association rules among the user behaviors, the situation characteristics and the load state are mined and dynamically maintained, and the accuracy and stability of load state recognition under the condition that the user behaviors are complex and changeable or the environment changes can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of load state recognition, and in particular to a user behavior association rule driven load state recognition enhancement method, system, device and medium. BACKGROUND

[0002] Currently, in the field of power load state recognition, the common method is to collect the current, voltage and switching state data of the user branch by deploying current transformers, voltage transformers and switching state sensors and the like based on load measurement data, and then to determine the load state by using threshold discrimination or pattern matching algorithm. This kind of method only constructs a model based on load measurement characteristics, without considering user power consumption behavior characteristics and environmental context characteristics, resulting in unstable recognition results in the case of complex user behavior or large load state changes.

[0003] To solve the above problems, the prior art introduces user power consumption behavior event flow and environmental context characteristics, and constructs a rule base for load state recognition by statistical analysis of historical data or mining of association rules. Such a rule base is generally obtained by statistically analyzing the association relationship among user behavior, context and load state, and is matched in real-time recognition. However, most of the existing rule bases are statically generated, lack dynamic updating mechanism, and cannot adapt to user behavior and environmental changes, so that the rule credibility decreases over time. In addition, when multiple rules are matched at the same time, the existing technology usually only selects the result according to the confidence level, without fully considering the mutual dependence relationship between rules and the load state change rule, which easily leads to conflicts that cannot be reasonably solved. SUMMARY

[0004] In view of the above problems, the present application is proposed.

[0005] Therefore, the technical problem solved by the present application is: how to solve the problem that the existing load state recognition method only relies on load measurement data and fails to fully combine user power consumption behavior characteristics and environmental context characteristics, resulting in insufficient recognition accuracy and stability when user behavior is complex or the environment changes, and the problem that the existing rule base based method has static rule generation, lacks dynamic updating and maintenance mechanism, rule credibility decays over time, cannot adapt to continuously changing user behavior and environment, and lacks effective means for conflict resolution by using rule dependence relationship and load state evolution rule when multiple rules match different load states, easily leading to recognition conflicts and uncertain results.

[0006] To solve the above technical problems, the application provides the following technical scheme: a user behavior association rule driven load state recognition enhancement method, which comprises the following steps: collecting user power consumption behavior event flow and load measurement data, combining the user power consumption behavior event flow and the load measurement data to form a context unit, the context unit comprising user behavior features, context features and load states; based on a plurality of context units, mining the association relationship between the user behavior features, the context features and the load states, generating a context rule library, the context rule library comprising a plurality of association rules, each association rule corresponding to a set of user behavior features, context features, load states and initial confidence; receiving a real-time collected context unit, retrieving the association rules matching the real-time context unit from the context rule library, and filtering the candidate load state recognition results based on the initial confidence; when there are different load state recognition results in the candidate load state recognition results, calculating an evaluation value according to the historical verification results of each matching rule, and determining a target load state recognition result from the candidate load state recognition results; according to the consistency between the target load state recognition result and the real-time load measurement data, adjusting the initial confidence to obtain an updated confidence, which is used to update the confidence of the corresponding association rule in the context rule library.

[0007] As a preferred scheme of the user behavior association rule driven load state recognition enhancement method, the forming of the context unit comprises the following steps: deploying a current transformer, a voltage transformer and a switch state sensor, collecting the current, voltage and switch state data of the user branch according to a preset time interval, and performing time stamp synchronization; based on the time stamp synchronized data, when a change in the switch state is detected, the data between the adjacent two changes is taken as an event segment; the active power, the reactive power and the duration time of each event segment are calculated, and the user equipment identifier, the action type and the occurrence time corresponding to the event segment are identified; the user equipment identifier and the action type extracted in the event segment are defined as the user behavior features, the occurrence time is defined as the context features, the active power, the reactive power and the duration time are defined as the load states, and the context unit is generated in a predetermined format.

[0008] As a preferred scheme of the user behavior associated rule driven load state recognition enhancement method, the generation of the context rule base comprises: constructing a statistical matrix according to a ternary group of user behavior characteristics, context characteristics and load states, and recording the time-weighted occurrence frequency of the corresponding combination in the matrix element; the time weighting is calculated according to the acquisition time of the context unit multiplied by the time attenuation coefficient; low-frequency combinations with an occurrence frequency lower than a first threshold value in the statistical matrix are detected, the low-frequency combinations are sparsely corrected to avoid rule loss, and the confidence after correction is calculated; the confidence of all combinations in the statistical matrix is smoothed to reduce abnormal fluctuations in the confidence, and the confidence of each combination in the matrix is updated to the smoothed value; combinations with a confidence higher than a second threshold value and a matching degree with the current environment label not lower than a third threshold value are screened from the smoothed matrix, and the user behavior characteristics, context characteristics, load state and confidence of the combinations are stored in the context rule base as effective rules.

[0009] As a preferred scheme of the user behavior associated rule driven load state recognition enhancement method, the generation of the context rule base comprises: constructing a statistical matrix according to a ternary group of user behavior characteristics, context characteristics and load states, and recording the time-weighted occurrence frequency of the corresponding combination in the matrix element; the time weighting is calculated according to the acquisition time of the context unit multiplied by the time attenuation coefficient; low-frequency combinations with an occurrence frequency lower than a first threshold value in the statistical matrix are detected, the low-frequency combinations are sparsely corrected to avoid rule loss, and the confidence after correction is calculated; the confidence of all combinations in the statistical matrix is smoothed to reduce abnormal fluctuations in the confidence, and the confidence of each combination in the matrix is updated to the smoothed value; combinations with a confidence higher than a second threshold value and a matching degree with the current environment label not lower than a third threshold value are screened from the smoothed matrix, and the user behavior characteristics, context characteristics, load state and confidence of the combinations are stored in the context rule base as effective rules.

[0010] As a preferred scheme of the user behavior associated rule driven load state recognition enhancement method, wherein: the determination of the target load state recognition result from the candidate load state recognition result comprises searching the historical verification records of each associated rule corresponding to the candidate load state in the context rule library, the historical verification records include verification time, verification result, verification environment label and cumulative verification times; the historical verification records are divided into multiple sliding time windows in chronological order, and the verification accuracy, residual stability index and anomaly detection index are calculated in each time window; the reliability score of each associated rule under the current environment is predicted based on the verification accuracy, residual stability index and anomaly detection index in each sliding time window; a rule probability graph is constructed according to the relationship between the candidate load state and the corresponding associated rule, the nodes of the rule probability graph represent the associated rules, and the weights of the edges represent the co-occurrence probability, conflict probability and state transition probability between the rules; the posterior probability of each candidate load state is calculated in the rule probability graph by using the Bayesian inference method combined with the residual of real-time load measurement data; the candidate load state with the highest posterior probability and the residual of real-time load measurement data lower than the preset threshold is selected as the target load state recognition result.

[0011] As a preferred scheme of the user behavior associated rule driven load state recognition enhancement method, wherein: the adjustment of the initial confidence to obtain the updated confidence comprises determining the predicted load curve according to the target load state recognition result, the predicted load curve is generated based on the typical operating characteristics corresponding to the target load state and the load dynamic constraints; the real-time load measurement curve is collected within a time window after the target load state recognition, and the real-time measurement curve is dynamically compared with the predicted load curve; according to the dynamic comparison result, the physical consistency index between the real-time measurement curve and the predicted load curve is calculated, the consistency index includes steady-state error, transient deviation, load mutation rate and running range constraint violation; it is judged whether the consistency index meets the preset physical constraint threshold, when it is met, it is determined that the target load state is consistent with the actual load, otherwise it is determined that it is inconsistent; each associated rule participating in the current target load state recognition is searched in the context rule library, and the updated confidence is calculated based on the contribution degree and the consistency index; the updated confidence is written into the corresponding associated rule in the context rule library.

[0012] As a preferred scheme of the user behavior association rule driven load state recognition enhancement method, after the updated confidence is obtained, the method further includes: receiving a future user behavior event stream and an environment context feature of a target user within a preset time range, and generating a future context unit sequence; retrieving, from the updated context rule library, an association rule similar to each context unit in the future context unit sequence and having a similarity greater than a matching threshold, to form a prediction rule sequence; performing weighted calculation on the prediction rule sequence according to the updated confidence, to generate a load state prediction result in a future time period; and generating a future load curve based on the load state prediction result, and outputting the prediction load curve for load scheduling or load optimization control.

[0013] The application provides a user behavior association rule driven load state recognition enhancement system.

[0014] To solve the above technical problems, the application provides the following technical scheme: a user behavior association rule driven load state recognition enhancement system, comprising: a collection module, configured to collect user power consumption behavior event streams and load measurement data, and combine the user power consumption behavior event streams and the load measurement data to form context units, wherein the context units comprise user behavior features, context features and load states; a rule library generation module, configured to mine an association relationship between the user behavior features, the context features and the load states based on a plurality of context units, and generate a context rule library, wherein the context rule library comprises a plurality of association rules, each association rule corresponds to a set of user behavior features, context features, load states and an initial confidence; a real-time matching module, configured to receive a real-time collected context unit, retrieve an association rule matched with the real-time context unit from the context rule library, and filter a candidate load state recognition result based on the initial confidence; a conflict resolution module, configured to, when different load state recognition results exist in the candidate load state recognition result, calculate an evaluation value according to historical verification results of the matching rules, and determine a target load state recognition result from the candidate load state recognition result; and a confidence updating module, configured to adjust the initial confidence according to consistency between the target load state recognition result and real-time load measurement data, obtain an updated confidence, and write the updated confidence into the corresponding association rule in the context rule library.

[0015] The application provides a computer device, comprising a memory and a processor, and the memory stores a computer program.

[0016] The application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the user behavior association rule driven load state recognition enhancement method.

[0017] The beneficial effects of the present application: by combining the user electricity consumption behavior characteristics, environmental context characteristics and load measurement data to establish the context unit, mining and dynamically maintaining the association rules between user behavior, context characteristics and load state, the accuracy and stability of load state recognition under the condition of complex and changeable user behavior or environmental change can be improved; by introducing historical verification records, sliding time window, rule probability graph and Bayesian inference, the conflict problem when multiple rules are matched at the same time is effectively solved, and the reliability of target load state recognition is improved; by dynamically adjusting the rule confidence based on the consistency of real-time load measurement data and predicted load curve, and using the updated rule base for future load state prediction and load optimization, the adaptability and prediction ability of the rule base are improved, and the load state recognition and prediction results are closer to the actual load operation characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A general flowchart of a user behavior association rule driven load state recognition enhancement method provided by an embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the present application.

[0021] Embodiment 1, refer to Figure 1 For an embodiment of the present application, the embodiment provides a user behavior association rule driven load state recognition enhancement method, comprising:

[0022] Step S100: collect user electricity consumption behavior event stream and load measurement data, combine the user electricity consumption behavior event stream and the load measurement data to form a context unit, the context unit includes user behavior characteristics, context characteristics and load state;

[0023] Step S200: based on the plurality of context units, mining the correlation between the user behavior features, the context features and the load state, generating a context rule library, the context rule library including a plurality of association rules, each association rule corresponding to a set of user behavior features, context features, load state and initial confidence;

[0024] Step S300: receiving real-time collected context units, retrieving the association rules matched with the real-time context units from the context rule library, and filtering the candidate load state recognition results based on the initial confidence;

[0025] Step S400: when there are different load states in the candidate load state recognition results, calculating the evaluation value according to the historical verification results of each matched rule, and determining the target load state recognition result from the candidate load state recognition results;

[0026] Step S500: adjusting the initial confidence according to the consistency between the target load state recognition result and the real-time load measurement data, obtaining the updated confidence, and updating the confidence of the corresponding association rule in the context rule library.

[0027] The existing load state recognition method generally relies on a pre-set static rule library, and matches the user load measurement data with the rule library to judge the current load state, which is difficult to reflect the dynamic influence of user electricity consumption behavior and environmental changes on the load state. When the user's operation habit, use period and environmental characteristics change, the existing method cannot adjust the rule confidence in time, nor can it reconstruct the rule library according to the change of load state, resulting in that the recognition result deviates from the actual load state. Therefore, the embodiment adopts to arrange current transformers, voltage transformers and switch state sensors on the user power distribution branch, collects current, voltage and switch state data according to a unified time reference, and forms context units in combination with user electricity consumption event flow, records user behavior features, context features and load state as complete triplets, and uses them to mine dynamic association rules.

[0028] As shown in Figure 1 After collecting and constructing the context units, the association between the user behavior, the context and the load state is extracted by constructing a statistical matrix to generate a rule library, the candidate state is filtered according to the similarity in the real-time recognition stage, the conflict is solved through the historical verification record and the probability inference, and the target load state is finally determined. The updated target state is dynamically compared with the real-time load measurement data, the confidence of the association rule is adjusted by calculating the consistency index, and the rule library is updated, so as to realize the dynamic maintenance of the rule library and the high-precision recognition of the load state.

[0029] In the data acquisition stage, the timestamps of each sensor are unified by a preset sampling frequency, the data between two changes in switch state is defined as an event segment when a change in switch state is detected, and the active power, reactive power and duration of the event segment are calculated, while the corresponding device identifier, action type and occurrence time of the event segment are marked for constructing a context unit. In the rule mining stage, the frequency of the ternary combination is counted according to time weighting, low-frequency rules are removed, and abnormal fluctuations are corrected by a smoothing algorithm, so as to filter out effective rules meeting the confidence requirement and store them in the rule base. In the real-time recognition stage, the characteristics of the real-time context unit are extracted, the consistency score and value distance score of the rule in the rule base are calculated, and the similarity is calculated by weighted calculation, and the final load state is determined based on the similarity and historical verification results, and the confidence of the associated rules in the rule base is dynamically adjusted combined with the real-time measurement curve.

[0030] Embodiment 2 is an embodiment of the present application, which provides a user behavior associated rule driven load state recognition enhancement method based on the previous embodiment, comprising:

[0031] Specifically, step S100 collects user electricity behavior event stream and load measurement data, combines the user electricity behavior event stream and the load measurement data to form a context unit, the context unit includes user behavior characteristics, context characteristics and load state, and includes the following steps A1-A4:

[0032] A1: Deploy current transformers, voltage transformers and switch state sensors, collect current, voltage and switch state data of user branch according to a preset time interval, and synchronize the timestamps;

[0033] A2: Based on the synchronized data, when a change in switch state is detected, the data between the adjacent two changes is regarded as an event segment;

[0034] A3: Calculate the active power, reactive power and duration of each event segment, and identify the corresponding user device identifier, action type and occurrence time of the event segment;

[0035] A4: Define the user device identifier and action type extracted in the event segment as user behavior characteristics, define the occurrence time as context characteristics, define the active power, reactive power and duration as load state, and combine to generate a context unit according to a predetermined format.

[0036] It should be noted that in the present embodiment, the current transformer, the voltage transformer and the switch state sensor are preferably deployed at the user power distribution branch level, specifically at each end circuit on the side of the household power supply, to realize fine monitoring of the energy consumption behavior of local equipment. To ensure data comparability, the sampling time interval is set to 200 ms, and the time alignment of various sensor data is completed through a unified master clock source or a GPS time module. This structure avoids the problem of time drift caused by different sampling frequencies, improves the accuracy of subsequent event positioning and breakpoint analysis, and is particularly suitable for residential scenarios with frequent dynamic changes in load and sensitive behavior characteristics.

[0037] Further, in this step, the change in switch state is used as the trigger sign for dividing the event segment, which has the natural property of user behavior boundary and can avoid the segmentation error caused by setting a fixed window. The Boolean judgment model is used to judge whether the logical difference value of the two time points before and after is 1, and the minimum time interval threshold is set to 1s to exclude false judgments caused by jitter. In addition, to reduce redundancy, this division strategy is only triggered when significant user interaction occurs, ensuring that the generated event segment closely corresponds to the user's real behavior, facilitating subsequent behavior-load state modeling.

[0038] Further, the current and voltage data in the event segment are first filtered and denoised, and then the active / reactive power is calculated based on the mean value method within a 10-second time window, and the duration is obtained by counting the start and end sampling points. The device identifier is formed by combining the branch number to which the sensor belongs and the user-set label, the action type is determined by the change direction of the switch quantity (0→1 for opening and 1→0 for closing), and the occurrence time is recorded as the time of the first sampling point of the event segment. This processing method can preserve the intrinsic characteristics of load changes while eliminating the bias introduced by sampling fluctuations, providing a basis for generating stable and high-resolution three types of feature fields.

[0039] Specifically, when generating a context unit, the user behavior features (such as "water heater-on"), context features (such as "Friday 18:20"), and load states (such as "P=3.1kW, Q=0.8kVar, T=65s") are combined into a sample unit with complete semantics in JSON structure format. This structure facilitates subsequent rule matching, statistical analysis, and graph structure modeling, and can map the non-explicit association between user behavior and physical load state, which is beneficial to rule generation quality control. In addition, all fields are bound to the data source number and the collection timestamp, ensuring traceability and providing support for the invention to achieve high-reliability load identification on the basis of transparent and controllable data sources and clear structure.

[0040] Further, step S200 mines the association relationship between the user behavior features, the context features and the load state based on the plurality of context units, and generates a context rule library, the context rule library including a plurality of association rules, each association rule corresponding to a set of user behavior features, context features, load state and initial confidence, including the following steps B1-B4:

[0041] B1: Construct a statistical matrix according to the triplets of user behavior features, context features and load states from the plurality of context units, and record the time-weighted occurrence frequency of the corresponding combination in the matrix element, wherein the time weighting is calculated according to the collection time of the context unit multiplied by a time decay coefficient;

[0042] B2: Detect low-frequency combinations with occurrence frequency lower than a first threshold in the statistical matrix, perform sparsity correction on the low-frequency combinations to avoid rule missing, and calculate the confidence after correction;

[0043] B3: Smooth the confidence of all combinations in the statistical matrix to reduce abnormal fluctuations in the confidence, and update the confidence of each combination in the matrix to the smoothed value;

[0044] B4: From the smoothed matrix, select combinations with confidence higher than a second threshold and matching degree with the current environment label not lower than a third threshold, and store the user behavior features, context features, load state and confidence of the combination in the context rule library as valid rules.

[0045] Further, step S300 receives real-time collected context units, retrieves association rules matching the real-time context units from the context rule library, and selects candidate load state recognition results based on the initial confidence, including the following steps C1-C6:

[0046] C1: Receive real-time collected context units, extract user behavior features and context features in the context units, and generate a matching vector;

[0047] C2: Traverse each association rule in the context rule library, and calculate the consistency score of the matching vector with the user behavior features and context features fields in the association rule respectively;

[0048] C3: Calculate the value distance score based on the difference between the values of the matching vector and the fields in the association rule;

[0049] C4: Weight the consistency score and the value distance score according to the preset weight to obtain the similarity between the matching vector and the association rule;

[0050] C5: Select association rules with similarity not lower than a fourth threshold from the context rule library to form a matching rule set;

[0051] C6: According to the initial confidence of each association rule in the matching rule set, the association rules with a confidence higher than the fifth threshold value are selected to generate a candidate load state recognition result.

[0052] Further, step S400: when there are different load states in the candidate load state recognition result, an evaluation value is calculated according to the historical verification results of each matching rule, and a target load state recognition result is determined from the candidate load state recognition result, including the following steps D1-D6:

[0053] D1: Retrieve the historical verification records of each association rule corresponding to the candidate load state in the context rule base, and the historical verification records include verification time, verification result, verification environment label and cumulative verification times;

[0054] D2: According to the time sequence, the historical verification records are divided into multiple sliding time windows, and the verification accuracy, residual stability index and anomaly detection index are calculated in each time window;

[0055] D3: Based on the verification accuracy, residual stability index and anomaly detection index in each sliding time window, the reliability score of each association rule under the current environment is predicted;

[0056] D4: According to the relationship between the candidate load state and the corresponding association rule, a rule probability graph is constructed, and the nodes of the rule probability graph represent the association rules, and the weights of the edges represent the co-occurrence probability, conflict probability and state transition probability between rules;

[0057] D5: In the rule probability graph, the posterior probability of each candidate load state is calculated by using the Bayesian inference method combined with the residual of real-time load measurement data;

[0058] D6: The candidate load state with the highest posterior probability and the residual lower than the preset threshold value of real-time load measurement data is selected as the target load state recognition result.

[0059] Further, step S500: According to the consistency between the target load state recognition result and the real-time load measurement data, the initial confidence is adjusted to obtain the updated confidence, which is used to update the credibility of the corresponding association rule in the context rule base, including the following steps E1-E6:

[0060] E1: According to the target load state recognition result, a predicted load curve is determined, and the predicted load curve is generated based on the typical operating characteristics corresponding to the target load state and the load dynamic constraints;

[0061] E2: In a time window after the target load state recognition, real-time load measurement curves are collected, and the real-time measurement curves are dynamically compared with the predicted load curve;

[0062] E3: According to the dynamic comparison result, calculate the physical consistency index between the real-time measurement curve and the predicted load curve, the consistency index includes steady-state error, transient deviation, load mutation rate and operation range constraint violation;

[0063] E4: Judge whether the consistency index meets the preset physical constraint threshold, when it meets, it is determined that the target load state is consistent with the actual load, otherwise it is determined that it is inconsistent;

[0064] E5: Retrieve each associated rule involved in the current target load state identification in the context rule library, and update the confidence degree based on its contribution degree and consistency index;

[0065] E6: Write the updated confidence degree into the corresponding associated rule in the context rule library.

[0066] Further, after obtaining the updated confidence degree, it further includes receiving future user behavior event stream and environmental context features of the target user within a preset time range, and generating a future context unit sequence;

[0067] Retrieve the associated rules with each context unit in the future context unit sequence from the updated context rule library, and form a predicted rule sequence when the similarity is greater than a matching threshold;

[0068] According to the updated confidence degree, the predicted rule sequence is weighted and calculated to generate a load state prediction result in the future time period;

[0069] Based on the load state prediction result, a future load curve is generated, and a predicted load curve is output for load scheduling or load optimization control.

[0070] Embodiment 3 is an embodiment of the present application, which provides a user behavior associated rule driven load state identification enhancement method based on the previous embodiment, comprising:

[0071] Specifically, in step S200, in order to represent the association between user behavior, context features and load state, a three-dimensional sparse statistical matrix is constructed, which corresponds to three dimensions respectively: (1) User behavior feature dimension, specifically device action category and control state; (2) Context feature dimension, specifically event occurrence time period, date type (weekday / holiday), external temperature and humidity level or user category; (3) Load state dimension, which is divided into discrete intervals according to the average active power, reactive power and change rate (derivative) in the event period.

[0072] When counting the matrix elements, an exponential time decay function is introduced to improve the response ability to the behavior change in the near period. The weight function is defined as follows:

[0073]

[0074] where w i (t) represents the time weight of sample i; T represents the current statistical moment; t i represents the sample collection moment; λ represents the time decay coefficient.

[0075] Specifically, T, t i The time stamp can be recorded automatically by the data collection platform; λ is an empirical parameter, which is recommended to be selected through model tuning or stability analysis, and is generally set to 10 -4 ~ 10 -3 , according to the user load change frequency, sampling period and behavior mode change degree. For example, for office building load, λ = 10 -4 may be taken; for residential users, it is appropriately increased to 5 × 10 -4 ; the obtained w i (t) is used for weighted counting into the matrix frequency value.

[0076] Finally, the occurrence frequency of any triple (B k , C j , L m ) in the statistical matrix is recorded as:

[0077]

[0078] wherein, is the sample set matching the triple.

[0079] Considering that the power distribution user load behavior has the characteristics of high dimension and low frequency, that is, part of the behavior is less likely to occur in a specific situation but is representative, it is necessary to avoid losing important modes due to direct elimination of low-frequency combinations. Therefore, a first threshold θ1 is set, and the recommended value is:

[0080]

[0081] wherein, T obs is the total duration in the statistical period (for example, one month), η is the average daily event number, N c is the number of behavior feature categories, N min is the minimum correction sample number, which is generally taken as 5-10.

[0082] For low-frequency combinations, the following correction strategies are adopted respectively:

[0083] Up and down adjacent situation compensation: query the statistical frequency of the same behavior to the time adjacent situation (such as adjacent hours, same type weather label);

[0084] Similar load state smoothing: define similar states according to the center value and standard deviation of active power interval, and transfer the weight;

[0085] Structural interpolation: when two-dimensional boundary behaviors exist, fill in the missing context load state combinations.

[0086] The confidence of each combination c k,j,m Then according to its modified weighted frequency f' k,j,m And the total frequency of the behavior in the current context is normalized:

[0087]

[0088] Where ε is a regularization term to avoid zero denominator, recommended to take 10 -6 .

[0089] In continuous sampling of user behavior data, there are sudden increases or decreases in partial confidence, in order to enhance the robustness of the model, all combination confidences are introduced based on behavior structure weighted sliding window smoothing. The sliding window size w can be set according to the user activity, the commonly used value is 1-7 days.

[0090] The smoothing formula is as follows:

[0091]

[0092] Where, represents the smoothed confidence value of the triple combination (B k ,C j ,L m ) at the current time t; φ k,j,m (t') represents the original confidence value of the combination at the historical time t'; represents the weight function used to describe the time distance decay, generally selected from symmetric window functions such as Gaussian window, triangular window or simple linear window; Z represents the normalization factor, whose value is the sum of all weights in the sliding window:

[0093]

[0094] The purpose is to prevent the value from being stretched or compressed arbitrarily in the weighted average process. Through normalization, the confidence value is still within the interval [0,1], which has comparability and numerical stability.

[0095] The selection of can be Gaussian kernel:

[0096]

[0097] Where σ can be selected according to the stability of user behavior patterns, such as σ=1.5 in day units; the normalization coefficient Z is directly obtained by accumulating all in the window, which can be obtained in real time through the sliding window processing program without setting up another sampling process.

[0098] In addition, the reference user behavior structure similarity is used to construct an adjacent weight graph, and the groups with a behavior similarity greater than a certain set threshold (for example, 0.85) are synchronously averaged in weight.

[0099] When the difference between the pre-smoothing confidence and the post-smoothing confidence exceeds a set range (for example, ±0.3), a truncation strategy is used instead to control error transmission.

[0100] The post-smoothing confidence result is used to screen effective association rules. The following three indicators are used as screening criteria:

[0101] The second threshold θ2: the lower limit of the confidence, which is usually set to about 0.7, reflecting the historical accuracy of the rule matching load state. For users in a multi-load overlapping area, it can be appropriately increased to 0.8 to enhance the distinguishing ability.

[0102] The third threshold θ3: the context matching degree requirement, which is calculated by a multi-dimensional feature matching function S(C j ,C q ) to calculate the similarity between the target context C q and the rule context C j . The matching function can be calculated according to the cosine similarity, Mahalanobis distance, or weighted Hamming distance, and the typical threshold is set to 0.75-0.9.

[0103] The lower limit of the occurrence intensity f min : used to ensure statistical reliability, dynamically set in combination with the total number of samples, generally 10-30 times.

[0104] The combination (B k , C j , L m ) that meets the above three conditions is considered to have sufficient statistical significance and context adaptability, and is included in the context rule library, and the confidence, sample source window, feature type identification, and other metadata are recorded.

[0105] Further, in step S300, in step C1, the real-time context unit collected contains multiple feature fields, the user behavior features and context features are extracted therefrom and arranged in a fixed order to form a set of fields to be matched, denoted as where f i represents the value of the i-th field, and there are m fields in total. Each field type, structure, and context rule library defined field correspond one by one, and the field number m is a preset value.

[0106] In step C2, the k-th rule in the context rule library is traversed, and its field set is represented as where f represents the value of the i-th field in the rule. A consistency function is defined as:

[0107]

[0108] To express the importance of each field to the matching process, a field matching weight vector θ = {θ1, θ2, ..., θ3} is defined. m}, where θ i ∈[0,1], and The score of rule k in terms of structural consistency with the current context unit is:

[0109]

[0110] Where, η k ∈[0,1] represents the structural consistency score, k represents the situational rule number, and f i , θ represents the value of the i-th field in the current cell and the k-th rule. i δ represents the matching weight of the i-th field, and δ represents the consistency judgment function.

[0111] Preferably, the consistency judgment function is mainly used to quickly determine whether discrete fields (such as device type, operation event, and scene label) are consistent. A preferred approach is to use a multi-level matching strategy rather than traditional hard Boolean judgment. In this embodiment, to improve the ability to identify approximate matching cases, the following improved function can be used:

[0112]

[0113] Where, β i ∈(0,1) represents the matching score when field i has semantic similarity or hierarchical inheritance relationship. The methods for obtaining this score include: constructing a similarity dictionary based on a user behavior tagging system, or determining whether the similarity between two values ​​exceeds a set threshold based on a domain ontology (such as a device category tree). A preferred method is to use category embedding encoding (such as Word2Vec or behavioral feature vectors) to calculate the cosine similarity between two symbols; a similarity is considered to exceed a threshold of 0.8. A β is then set. i = Between 0.6 and 0.8.

[0114] In step C3, the set of fields in the field set that have numerical values ​​is denoted as... For each Define the distance of this field between the current context unit and the rule as... To standardize the units of measurement, a normalized upper limit value R is introduced for this field. i This refers to the upper bound of the range of values ​​for this field in historical data, or the preset maximum acceptable difference. The distance score calculation formula is as follows:

[0115]

[0116] wherein ζ k ∈ [0, 1] represents the distance similarity score, φ i ∈ [0, 1] represents the matching sensitivity weight of the i-th numerical field, R i > 0 represents the normalization factor of field i, used for standardizing the difference, with the constraint condition This score measures the overall similarity of the numerical difference between fields, the smaller the distance, the higher the score.

[0117] Preferably, the matching sensitivity weight determines the degree of influence of different numerical fields in the value difference measurement. To ensure the weight has interpretability and actual validity, the feature influence measurement method is preferably used. The steps are as follows:

[0118] (1) Use existing historical samples to construct a lightweight supervised model (such as random forest) to predict the load state;

[0119] (2) Statistics the influence degree of each numerical field on the model output, such as feature importance index;

[0120] (3) Standardize the influence values of all fields to make them normalized to [0, 1] as the value of φ i ;

[0121] (4) If there is no support for the unsupervised model, the inverse proportion of the variance of the field can be weighted, that is, the field with large fluctuation range is given low sensitivity to express its fault tolerance in state recognition.

[0122] In step C4, to fuse the structural consistency and the numerical distance similarity, a fusion weight coefficient is introduced.

[0123]

[0124] wherein γ k ∈ [0, 1] represents the overall similarity score of rule k, represents the fusion weight factor, which controls the proportion of structure matching and numerical matching, η k , ζ k represent the aforementioned structure and distance scores.

[0125] The selection of weight can be set according to the experience of the behavior characteristic stability of the target scene. For example, for industrial loads with stable feature values, 0.7-0.9 can be taken; for residential users with obvious preference fluctuations, 0.4-0.6 can be set.

[0126] Preferably, the fusion weight factor is used to balance the contribution of structural fields (discrete consistency) and numerical fields (feature distance) to the similarity, which has a decisive influence. Preferably, its value is set by cross-validation performance optimization method:

[0127] (1) Construct a candidate model, test a plurality of values on the real sample set respectively;

[0128] (2) Compare the system recognition accuracy, F1 value or residual curve matching degree under different values;

[0129] (3) Select the that performs best in the target user type or typical power consumption situation as the current configuration;

[0130] (4) If adaptive adjustment is required, scene labels can be introduced as input variables to establish a regression model of (user type, device density).

[0131] This way can realize optimal dynamic control of fusion proportion according to different application situations.

[0132] In step C5, according to the set similarity screening lower threshold value γ th ∈(0,1) (i.e. the fourth threshold value), all rules that satisfy: γ k ≥γ th are selected as the preliminary matching rule set. This value is used to control the relevance density of the preliminary rule set, and is recommended to be set in the interval of 0.65-0.85, and is obtained according to the false recognition rate and acceptable redundancy control in simulation evaluation.

[0133] In step C6, in order to further improve the recognition accuracy, the rule initial confidence δ k ∈[0,1] is introduced, which is derived from the comprehensive evaluation of the appearance frequency, correction value and verification history of the three tuple combination in the statistical matrix. Set the confidence screening lower limit δ th ∈(0,1) (i.e. the fifth threshold value), sort all matching rules that satisfy: δ k ≥δ th , and keep the corresponding rules of the highest k as the candidate load state recognition rules. The recommended value of δ th is in the range of 0.7-0.9, which is determined according to the error tolerance of the target application and the historical model verification accuracy.

[0134] Preferably, the fourth threshold value is used to control the screening strictness of the matching rule, which affects the coverage range and false recognition rate of the matching rule, and is preferably determined by a data-driven quantile selection method:

[0135] (1) Construct a confidence-similarity joint distribution based on the similarity score between a historically verified context unit and its correct matching rule;

[0136] (2) Select the lower bound quantile corresponding to the target identification recall rate (e.g., 90%) as γ th , ensuring sufficient coverage;

[0137] (3) For specific industries, higher screening standards can be set (e.g., power dispatch load identification requires more stringent standards), corresponding to a setting between 0.75-0.85;

[0138] (4) To avoid mismatch risks, a dynamic fine-tuning mechanism can be set: when the identification accuracy decreases within a short time window, automatically reduce γ th to improve fault tolerance.

[0139] In the fifth threshold value, the initial confidence threshold directly determines whether to adopt a certain rule into the candidate set, and its setting is preferably based on the historical identification stability-confidence mapping method:

[0140] (1) Divide all rules in the rule base into several levels (e.g., 0.5-0.6, 0.6-0.7,...) according to the initial confidence;

[0141] (2) Calculate the average identification accuracy, residual fluctuation range, and prediction consistency corresponding to each confidence level;

[0142] (3) Select the minimum confidence level corresponding to the target level (e.g., 85%) of identification accuracy as δ th ;

[0143] (4) If the actual application has a high penalty for misidentification (e.g., medical, precision manufacturing loads), the value can be increased to 0.85-0.9.

[0144] (5) This threshold value can also be dynamically adjusted by the dispatch optimization system feedback identification reliability during the model deployment phase.

[0145] Further, in step S400, to determine a unique target state from multiple candidate load states, the embodiment introduces a historical verification-driven reliability evaluation and rule-based reasoning mechanism, as follows:

[0146] For each candidate load state L j , retrieve all associated rule sets supporting the state from the context rule base Each rule is associated with a set of historical verification records, including verification timestamp t m , verification result y m (value 1 indicates correct identification, 0 indicates error), and environment label vector e mand the cumulative verification times of the rule

[0147] Then, the verification records of each rule are sorted by timestamps and divided into fixed-size sliding time windows W h Each window contains a set of consecutive verification records. The following three indicators are calculated within each window:

[0148] Verification accuracy:

[0149]

[0150] Residual stability indicator:

[0151]

[0152] Anomaly detection indicator:

[0153]

[0154] wherein, represents the rule The accuracy rate of the rule in the window W h , is the ratio of the residual standard deviation σ r to the residual mean μ r , which measures the stability of the rule inference result residual, is the ratio of the number of anomaly detections N abn to the total number of records, which reflects the anomaly probability of the rule in this time period.

[0155] For the current target context label vector e now , select the set of time windows in the historical verification records whose labels are closest to the target label vector and calculate the reliability score of the current rule based on the weighted average strategy:

[0156]

[0157] wherein, represents the rule The reliability of the rule in the current context, are its average verification accuracy, average stability indicator and average anomaly indicator in the matching environment window, respectively. Coefficients η1, η2, η3 represent the weight factors of the three indicators, satisfying η1+η2+η3=1, and the preferred values are η1=0.5, η2=0.3, η3=0.2.

[0158] To model the relationship between rules, a rule probability graph G=(V,E) is constructed, where V is the set of all candidate rule nodes, and E is the set of edges between rules. The edge weight is calculated using the following formula:

[0159] w ij = θ1·P co (i,j)- θ2·P cf (i,j)+ θ3·P tr (i,j)

[0160] where w ij represents the relationship strength between the rule R i and R j , P co (i,j) is the frequency of the simultaneous occurrence of the two (co-occurrence probability), P cf (i,j) is the frequency of the output of different states of the two (conflict probability), P tr (i,j) is the probability of the transition of the prediction results of the two on the time series (state switching probability), and the weight coefficients θ1, θ2, and θ3 satisfy θ1+ θ2+ θ3= 1, and the preferred values are θ1= 0.4, θ2= 0.4, and θ3= 0.2.

[0161] Combined with the residual information of real-time load measurement data, the Bayesian inference method is used to calculate the posterior probability of each candidate state L j :

[0162]

[0163] where P(L j ) is the prior probability of the state L j , and the calculation method is the weighted average credibility of the corresponding rule:

[0164]

[0165] represents the probability of observing the residual o k under the given rule and state, which is expressed by using a Gaussian model as:

[0166]

[0167] where o k is the residual between the actual measurement value and the predicted value of the rule, μ k and σ k are the mean and standard deviation of the historical residual of the rule, respectively, and are calculated based on the sliding window historical samples.

[0168] Finally, the state L * with the maximum posterior probability is selected from all candidate states, and the mean square error (MSE) of the current residual o * of the state L * does not exceed the set threshold ε max , and the state L * is taken as the target load state of this identification. The threshold ε maxThe 95th percentile value of the residual distribution of the model training phase is preferably set.

[0169] Further, in step S500, to improve the dynamic adaptability of the association rules and the reliability of the state recognition results, a confidence iterative updating method based on physical consistency feedback is adopted. The method includes the following specific processes:

[0170] First, according to the target load state recognition result L * , its corresponding typical state feature vector is obtained from the rule base The predicted load curve is constructed The prediction curve is generated by the following function:

[0171]

[0172] wherein, P(t) represents the active power value at the prediction time point t, P0 represents the static average active power of the state, a P represents the dynamic power fluctuation amplitude, T d represents the typical operating period (seconds) of the state, represents the internal phase offset (radians) of the state, and t0 represents the starting time of the target state.

[0173] Then, within the time window [t0, t0+T w ] (such as 30 seconds), real-time load measurement data P obs (t) is collected and dynamically compared with the prediction curve. The comparison result generates a residual curve ∈(t):

[0174]

[0175] wherein, ∈(t) represents the measurement error residual at time t, and P obs (t) represents the real-time collected active load power.

[0176] On the basis of residual analysis, the following four physical consistency indicators are further extracted:

[0177] The steady-state error indicator (average deviation) is:

[0178]

[0179] wherein, δ ss represents the normalized error in the steady-state stage, T s represents the steady-state evaluation interval length, t s represents the starting time point of entering the steady state (usually t0+0.5T d after recognition).

[0180] Transient response error index:

[0181]

[0182] where δ tr represents the maximum rate of change error during the transient period, T r represents the transient period length (e.g. 5-10 seconds), d∈(t) / dt represents the derivative of the residual with respect to time, and represents the rate of change.

[0183] Load mutation rate index:

[0184]

[0185] where γ represents the second order rate of change of the load curve per unit time, d 2 P obs (t) / dt 2 represents the acceleration of the load measurement, which is used to reflect the mutation trend.

[0186] Boundary violation rate index:

[0187]

[0188] where represents the proportion of time that the actual measured power violates the typical operating interval, 1{·} represents the indicator function, which is 1 if true and 0 if false, and [P min , P max ] represents the operating power boundary allowed in this state.

[0189] If the above indexes are all lower than their corresponding preset thresholds (e.g. ε ss , ε tr , γ max , ), then the target load state is determined to be physically consistent, and a consistency identifier χ = 1 is set, otherwise χ = 0.

[0190] Next, the rule set participating in the state recognition is extracted For each rule R i , its contribution factor is calculated

[0191]

[0192] where represents the weighted contribution of rule R i in the current recognition, s i represents the matching similarity of rule R i with the current context unit (refer to step S300 similarity calculation), represents the matching similarity of rule R iinitial confidence of each rule R The normalized weight reflecting the overall contribution.

[0193] Finally, the confidence of each rule R i is dynamically revised in combination with the consistency result, using the following exponential feedback formula:

[0194]

[0195] wherein, represents the updated confidence, η represents the confidence update gain factor, usually in the range of 0.05-0.2, χ represents the consistency identifier, 1 if consistent, otherwise 0, represents the rule contribution degree.

[0196] The updating mechanism ensures that when the recognition result is consistent with the actual measurement height, the confidence of the main contribution rule is enhanced; if the recognition result deviates significantly, the confidence of the wrong rule is reduced through the punishment mechanism, and the dynamic self-adaptive evolution ability of the rule base is maintained. After the update is completed, the i of all rules R is replaced by the original confidence, written back to the context rule base, and used as a trusted basis for the next round of recognition process.

[0197] Embodiment 4 is an embodiment of the present application, which provides a user behavior association rule driven load state recognition enhancement system, comprising: a collection module for collecting user electricity behavior event stream and load measurement data, and combining the user electricity behavior event stream with the load measurement data to form a context unit, the context unit comprising user behavior features, context features and load state;

[0198] A rule base generation module is configured to mine the association relationship between user behavior features, context features and load state based on a plurality of context units, and generate a context rule base, the context rule base comprising a plurality of association rules, each association rule corresponding to a set of user behavior features, context features, load state and initial confidence.

[0199] A real-time matching module is configured to receive a real-time collected context unit, retrieve association rules matching the real-time context unit from the context rule base, and filter candidate load state recognition results based on the initial confidence.

[0200] A conflict resolution module is configured to determine a target load state recognition result from the candidate load state recognition results when different load states exist in the candidate load state recognition results, according to the evaluation value calculated according to the historical verification results of each matching rule.

[0201] The confidence updating module is configured to adjust the initial confidence according to consistency between the target load state recognition result and the real-time load measurement data, obtain an updated confidence, and write the updated confidence into a corresponding associated rule in the context rule library.

[0202] The embodiment further provides an electronic device suitable for a user behavior associated rule driven load state recognition enhancement method, including a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the user behavior associated rule driven load state recognition enhancement method.

[0203] The embodiment further provides a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the user behavior associated rule driven load state recognition enhancement method.

[0204] The storage medium provided by the embodiment and the user behavior associated rule driven load state recognition enhancement method provided by the above embodiment belong to the same inventive concept, and the technical details not described in the embodiment can be referred to the above embodiment, and the embodiment and the above embodiment have the same beneficial effects.

[0205] From the above description about the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary universal hardware, and of course can be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH memory, a hard disk or an optical disk, and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.

[0206] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A user behavior association rule-driven method for enhancing load status identification, characterized in that: include, Collect user electricity consumption behavior event streams and load measurement data, and combine user electricity consumption behavior event streams and load measurement data to form scenario units. Scenario units include user behavior characteristics, scenario characteristics, and load status. Based on multiple context units, the correlation between user behavior features, context features and load status is mined to generate a context rule base. The context rule base includes multiple association rules, each of which corresponds to a set of user behavior features, context features, load status and initial confidence level. Receive real-time collected context units, retrieve association rules matching the real-time context units from the context rule base, and filter candidate load state identification results based on initial confidence. When different load states exist in the candidate load state identification results, the evaluation value is calculated based on the historical verification results of each matching rule, and the target load state identification result is determined from the candidate load state identification results. Based on the consistency between the target load status identification results and the real-time load measurement data, the initial confidence level is adjusted to obtain the updated confidence level, which is used to update the confidence level of the corresponding association rules in the context rule base.

2. The user behavior association rule-driven load status recognition enhancement method as described in claim 1, characterized in that: The scenario formation unit includes deploying current transformers, voltage transformers, and switch status sensors to collect current, voltage, and switch status data of user branches at preset time intervals and perform timestamp synchronization. Based on the data synchronized with timestamps, when a change in switch state is detected, the data between two adjacent changes is treated as an event segment. For each event segment, calculate the active power, reactive power, and duration, and identify the user equipment identifier, action type, and occurrence time corresponding to the event segment; The user equipment identifier and action type extracted from the event segment are defined as user behavior features, the occurrence time is defined as situation features, and the active power, reactive power and duration are defined as load status. These are then combined according to a predetermined format to generate situation units.

3. The user behavior association rule-driven load status recognition enhancement method as described in claim 2, characterized in that: The generated context rule base includes constructing a statistical matrix by tuples of multiple context units according to user behavior characteristics, context characteristics and load status. The matrix elements record the time-weighted frequency of occurrence of the corresponding combination. The time weighting is calculated by multiplying the collection time of the context unit by a time decay coefficient. Low-frequency combinations that appear less than the first threshold in the detection statistics matrix are subjected to sparsity correction to avoid missing rules, and the corrected confidence level is calculated. The confidence scores of all combinations in the statistical matrix are smoothed to reduce abnormal fluctuations in confidence scores, and the confidence scores of each combination in the matrix are updated to the smoothed values. From the smoothed matrix, select combinations with confidence scores higher than the second threshold and a matching degree with the current environment label not lower than the third threshold. Store the user behavior characteristics, contextual characteristics, load status, and confidence scores of the combinations into the contextual rule base as valid rules.

4. The user behavior association rule-driven load status recognition enhancement method as described in claim 3, characterized in that: The process of filtering candidate load state identification results based on initial confidence includes receiving real-time collected context units, extracting user behavior features and context features from the context units, and generating a matching vector. Iterate through each association rule in the context rule base and calculate the consistency score between the vector to be matched and the user behavior feature and context feature fields in the association rule; Calculate the value distance score based on the difference between the vector to be matched and the values ​​of each field in the association rule; The consistency score and the value distance score are weighted according to preset weights to obtain the similarity between the vector to be matched and the association rule. Select association rules with similarity no lower than the fourth threshold from the context rule base to form a matching rule set; Sort the association rules in the matching rule set from highest to lowest initial confidence level, select the association rules with confidence levels higher than the fifth threshold, and generate candidate load status identification results.

5. The user behavior association rule-driven load status recognition enhancement method as described in claim 4, characterized in that: Determining the target load state identification result from the candidate load state identification results includes retrieving historical verification records of each associated rule corresponding to the candidate load state from the context rule base. The historical verification records include verification time, verification result, verification environment label, and cumulative number of verifications. The historical verification records are divided into multiple sliding time windows in chronological order, and the verification accuracy, residual stability index and anomaly detection index are calculated in each time window. The reliability score of each association rule in the current environment is predicted based on the verification accuracy, residual stability index and anomaly detection index within each sliding time window. A rule probability graph is constructed based on the relationship between candidate load states and corresponding association rules. The nodes in the rule probability graph represent association rules, and the weights of the edges represent the co-occurrence probability, conflict probability, and state transition probability between rules. The posterior probability of each candidate load state is calculated by using Bayesian inference methods combined with the residuals of real-time load measurement data in the regular probability graph. The candidate load state with the highest posterior probability and a residual difference from the real-time load measurement data below a preset threshold is selected as the target load state identification result.

6. The user behavior association rule-driven load status recognition enhancement method as described in claim 5, characterized in that: The process of adjusting the initial confidence level to obtain the updated confidence level includes determining the predicted load curve based on the target load state identification result. The predicted load curve is generated based on the typical operating characteristics and load dynamic constraints corresponding to the target load state. Real-time load measurement curves are collected within a time window after the target load status is identified, and the real-time measurement curves are dynamically compared with the predicted load curves. Based on the dynamic comparison results, the physical consistency index between the real-time measurement curve and the predicted load curve is calculated. The consistency index includes steady-state error, transient deviation, load mutation rate, and violation of operating range constraints. Determine whether the consistency index meets the preset physical constraint threshold. If it does, the target load state is considered to be consistent with the actual load; otherwise, it is considered to be inconsistent. Retrieve each associated rule in the context rule base that participates in the identification of the current target load status, and calculate the update confidence based on its contribution and consistency index; Write the updated confidence level into the corresponding association rule in the context rule base.

7. The user behavior association rule-driven load status recognition enhancement method as described in claim 6, characterized in that: After obtaining the updated confidence level, the method further includes receiving the target user's future user behavior event stream and environmental context features within a preset time range, and generating a future context unit sequence. Retrieve association rules from the updated context rule base that have a similarity greater than the matching threshold with each context unit in the future context unit sequence to form a prediction rule sequence; The prediction rule sequence is weighted according to the updated confidence level to generate load status prediction results for future time periods. The future load curve is generated based on the load state prediction results, and the predicted load curve is output for load scheduling or load optimization control.

8. A load status recognition enhancement system driven by user behavior association rules, employing the load status recognition enhancement method driven by user behavior association rules as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to collect user electricity consumption behavior event streams and load measurement data, and combine user electricity consumption behavior event streams and load measurement data to form a scenario unit. The scenario unit includes user behavior characteristics, scenario characteristics, and load status. The rule base generation module is used to mine the correlation between user behavior features, context features and load status based on multiple context units, and generate a context rule base. The context rule base includes multiple association rules, and each association rule corresponds to a set of user behavior features, context features, load status and initial confidence level. The real-time matching module is used to receive real-time acquired context units, retrieve association rules that match the real-time context units from the context rule base, and filter candidate load state identification results based on the initial confidence level. The conflict resolution module is used to calculate the evaluation value based on the historical verification results of each matching rule when there are different load states in the candidate load state identification results, and to determine the target load state identification result from the candidate load state identification results. The confidence update module is used to adjust the initial confidence based on the consistency between the target load status identification result and the real-time load measurement data, obtain the updated confidence, and write the updated confidence into the corresponding association rule in the context rule base.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the user behavior association rule-driven load status recognition enhancement method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the user behavior association rule-driven load status recognition enhancement method according to any one of claims 1 to 7.

Images