An intelligent charging pile operation analysis system based on an AI large model

Abstract

The application discloses an intelligent charging pile operation analysis system based on an AI large model, relates to the technical field of charging pile operation management, and comprises the following modules: an operation data collection module, which is used for obtaining a multi-source operation dataset; a proposition variable generation module, which is used for forming an operation proposition variable set; an attribution clause generation module, which is used for inputting the operation proposition variable set into a Tsetlin Machine attribution control model to generate an attribution control logic clause set; a trigger matching attribution module, which is used for determining an attribution action result; a control instruction execution module, which is used for forming a control execution record; a residual error feedback updating module, which is used for generating execution residual error feedback data; and a large model report generation module, which is used for applying bidirectional feedback to a Tsetlin Machine attribution control model corresponding to a trigger clause set and generating an operation analysis report of a target charging station based on an AI large model. The application improves the accuracy of operation anomaly identification of the target charging station.

Description

Technical Field

[0001] This invention relates to the field of charging pile operation and management technology, and in particular to an intelligent charging pile operation analysis system based on an AI large model. Background Technology

[0002] With the continuous increase in the number of new energy vehicles, the operational scale of public charging stations, park charging stations, and highway service area charging stations is constantly expanding. Charging pile operation management systems typically need to collect data on equipment operating status, charging orders, site load, electricity prices, weather, and revenue. Based on the collected results, they perform fault alarms, load monitoring, order analysis, operation and maintenance scheduling, and operation report generation. In existing technologies, some systems are already able to identify and display charging pile offline, power anomalies, order interruptions, increased queuing times, and site load fluctuations through rule thresholds, statistical analysis models, machine learning models, or large-scale model text generation methods.

[0003] However, existing charging pile operation analysis methods typically process equipment-side data and site-side data separately. There is a lack of a unified correlation expression between equipment operation anomalies, site load anomalies, order anomalies, and revenue anomalies. It is difficult to determine whether a certain operational anomaly originates from the charging pile equipment status, site resource congestion, external electricity price and weather changes, or the combined effect of multiple factors. For continuous state quantities, existing solutions mostly use fixed thresholds or manually set intervals for judgment, which makes it difficult to dynamically form a propositional expression suitable for the current charging station based on historical operational anomaly categories. This results in insufficient adaptability of anomaly identification results to different sites, different time windows, and different operational scenarios.

[0004] Therefore, how to provide an intelligent charging pile operation analysis system based on a large AI model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent charging pile operation analysis system based on an AI large model. This invention improves the accuracy of identifying operational anomalies, the reliability of matching control actions, and the traceability of operation analysis reports by cross-side propositionalization, factor sentence triggering, and residual feedback updating of the charging pile equipment-side status and the site-side operation status.

[0006] According to an embodiment of the present invention, an intelligent charging pile operation analysis system based on an AI large model includes: The operation data acquisition module is used to collect multi-source operation data of the target charging station within a continuous time window, and to preprocess the data to obtain a multi-source operation dataset. The propositional variable generation module is used to determine the constraint interval of continuous state variables in multi-source operation datasets and to encode the determination results into equipment-side propositional variables and site-side propositional variables to form an operation propositional variable set. The attribution clause generation module is used to input the set of operational proposition variables into the Tsetlin Machine attribution control model, drive the Tsetlin automaton to perform inclusion or exclusion state selection on the equipment-side text items and site-side text items in the set of operational proposition variables, and generate a set of attribution control logic clauses. The trigger matching attribution module is used to perform trigger matching on the set of labeled logical clauses based on the set of operational proposition variables, obtain the set of trigger clauses, and determine the attribution action result based on the integer weights of the logical clauses in the set of trigger clauses; The control command execution module is used to generate control commands based on the attribution action results and send the control commands to the charging pile operation execution terminal to form a control execution record; The residual feedback update module is used to collect target index data before and after the execution of control commands based on the control execution record, and generate execution residual feedback data based on the target index data. The large model report generation module is used to apply bidirectional feedback to the TsetlinMachine attribution control model corresponding to the trigger clause set using execution residual feedback data, update the text item status and integer weight of the logical clauses in the trigger clause set, and generate an operation analysis report of the target charging station based on the AI ​​large model.

[0007] Optionally, the operational data collection module includes: Data is continuously collected from the target charging station according to the preset time window to obtain raw multi-source operation data. Preprocessing is performed on the original multi-source operation data, including timestamp alignment, outlier removal, and missing sample value completion, to obtain preprocessed multi-source operation data. The preprocessed multi-source operational data are associated according to the operational index identifier to obtain a multi-source operational dataset.

[0008] Optionally, the propositional variable generation module includes: Extract continuous state variables from the device side and the site side from the multi-source operation dataset; An operational anomaly category identifier is generated based on the same operational index identifier, and a correspondence is established between the operational anomaly category identifier and the continuous status quantity on the equipment side and the continuous status quantity on the site side, respectively. The CAIM algorithm is used to perform supervised discretization of continuous state variables on the equipment side and the site side. The operation anomaly category identifier is used as the category item, and each continuous state variable is used as the attribute item. The interdependence of the category attributes of the category items is calculated according to the candidate split points of the continuous state variables, and the candidate split points that increase the interdependence of the category attributes are selected as the constraint split points. Based on the constraint split points, generate the equipment-side constraint discrete interval and the site-side constraint discrete interval respectively, and map the equipment-side continuous state variables and the site-side continuous state variables to the corresponding equipment-side constraint discrete interval and the site-side constraint discrete interval respectively to obtain the interval hit result; Binary encoding is performed on the interval hit results to obtain the device-side propositional variables and the site-side propositional variables; The device-side propositional variables and site-side propositional variables are associated with the same operational index identifier to form an operational propositional variable set.

[0009] Optionally, the factorization sentence generation module includes: The operational proposition variable set is input into the Tsetlin Machine attribution control model, which includes a proposition partitioning unit, a text item construction unit, a cross-side co-occurrence binding unit, a Tsetlin automaton, an attribution control clause construction unit, a label weight binding unit, and a clause storage unit. The proposition partitioning unit partitions the operational proposition variable set according to the operational index identifier to obtain the device-side proposition variable set and the site-side proposition variable set. The text item construction unit generates positive and negative text items for each equipment-side proposition variable in the equipment-side proposition variable set, and generates positive and negative text items for each site-side proposition variable in the site-side proposition variable set. The positive, negative, positive, and negative text items for the equipment side are used together as attribution control text items. The cross-side co-occurrence binding unit calculates the cross-side co-occurrence strength between device-side and site-side text items according to the synchronization value relationship between device-side propositional variables and site-side propositional variables under the same operation index identifier, and forms a cross-side text item candidate group with the cross-side co-occurrence strength meeting the preset binding conditions. The cross-side text item candidate group is assigned to the Tsetlin automaton group, and the Tsetlin automaton group performs inclusion or exclusion state selection on the device-side text items and site-side text items in the cross-side text item candidate group based on the cross-side co-occurrence strength and the current value in the operational proposition variable set, so as to obtain the cross-side text item selection result. The attribution control clause construction unit constructs attribution control logic clauses based on the cross-border text item selection results; The label weight binding unit binds the cause label, action label, and integer weight to the same attribution control logic clause; The clause storage unit associates and stores different attribution control logic clauses according to clause number, operational index identifier, and cross-side text item source, generating an attribution control logic clause set.

[0010] Optionally, the trigger matching attribution module includes: Based on the current operation index identifier, retrieve the operation proposition variable set and the attribution control logic clause set. Perform cross-side trigger determination for each attribution control logic clause. When both the device-side text item and the site-side text item in the same attribution control logic clause are satisfied by the operation proposition variable set, determine the current attribution control logic clause as a cross-side trigger clause. Perform attribution action consistency check on cross-side trigger clauses, group cross-side trigger clauses with the same cause label and the same action label into the same attribution action candidate group, and determine the trigger contribution value of the current attribution action candidate group based on the integer weight of the cross-side trigger clauses and the number of cross-side text items in each attribution action candidate group. The candidate group of attribution actions with the largest trigger contribution value is determined as the target attribution action group. The cross-side trigger clauses in the target attribution action group are determined as the trigger clause set. Attribution action results are generated based on the cause label and action label corresponding to the target attribution action group.

[0011] Optionally, the control command execution module includes: Read the attribution action results, and verify the clause number, integer weight, and cross-side text item source of the cross-side trigger clauses in the trigger clause set according to the current operation index identifier to form the basis for control generation; The control execution elements are determined based on the action labels, and the attribution source of the control instructions is determined based on the cause labels; Control commands are generated based on the control generation criteria and control execution elements, and then sent to the charging pile operation execution terminal to form a control execution record.

[0012] Optionally, the residual feedback update module includes: Read the control execution record and trigger clause set. Based on the device-side text item source and site-side text item source of each cross-side trigger clause in the trigger clause set, reversely determine the target indicator item that corresponds one-to-one with the cross-side trigger clause to form a clause indicator mapping relationship. Based on the control object identifier and action label in the control execution record, establish a pre-observation window, a post-observation window, and a control observation window, and collect pre-target indicator data, post-target indicator data, and control target indicator data respectively to generate net action effect target indicator data; Based on the clause index mapping relationship, the net effect target index data of the action is converted into execution residual feedback data oriented towards the trigger clause set.

[0013] Optionally, the large model report generation module includes: According to the current operation index identifier, the residual feedback data will be executed back to the TsetlinMachine attribution control model corresponding to the trigger clause set, and the attribution control logic clause, device-side text item, site-side text item and integer weight corresponding to the trigger clause set will be located in the Tsetlin Machine attribution control model to generate clause-level sub-side feedback records. Based on the clause-level feedback records, bidirectional feedback is applied to the Tsetlin Machine attribution control model corresponding to the trigger clause set to obtain the application result of the bidirectional feedback; Based on the results of the two-way feedback, the Tsetlin automaton literal state and integer weights corresponding to the attribution control logic clauses in the trigger clause set are updated. A large model evidence input package is constructed based on the updated literal state, integer weights, trigger clause set, attribution action results, control execution records, and execution residual feedback data. This package is then input into the AI ​​large model to generate an operational analysis report for the target charging station.

[0014] The beneficial effects of this invention are: This invention collects equipment operation data, charging order data, site load data, electricity price data, weather data, and site revenue data of the target charging station within a continuous time window, and converts the continuous state variables into equipment-side propositional variables and site-side propositional variables. This allows the charging pile equipment status and charging station operation status to be expressed in a unified set of operation propositional variables, thereby improving the correlation analysis capability between multi-source operation data and reducing the risk of misjudgment caused by relying solely on single equipment data or single site data for anomaly judgment.

[0015] This invention uses the Tsetlin Machine attribution control model to bind the co-occurrence of equipment-side and site-side text items across sides, and generates attribution control logic clauses that simultaneously contain equipment-side and site-side text items. This allows the cause of operational anomalies to be jointly determined by the operating status of the charging pile equipment and the operating status of the site, thereby improving the accuracy and interpretability of attribution for operational anomalies of the target charging station, and enabling the attribution results to establish a clear correspondence with control actions.

[0016] This invention generates control commands based on the attribution action results, and applies bidirectional feedback to the Tsetlin Machine attribution control model corresponding to the trigger clause set through control execution records and execution residual feedback data. This enables the equipment-side improvement results and site-side improvement results after the control command execution to update the text item status and integer weight in the attribution control logic clause in reverse, thereby improving the reliability of control action matching and allowing the model to continuously correct abnormal attribution and control strategies based on the actual operational execution effect.

[0017] This invention uses a large AI model to structurally summarize the trigger clause set, attribution action results, control execution records, execution residual feedback data, and model update results to generate an operation analysis report for the target charging station. This enables operators to clearly identify the causes of anomalies, the scope of affected charging piles, the effectiveness of control actions, and the basis for the next round of operational handling, thereby improving the readability, traceability, and operational decision support capabilities of the intelligent charging pile operation analysis results. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent charging pile operation analysis system based on an AI large model proposed in this invention; Figure 2 This is a schematic diagram of the TsetlinMachine cross-side attribution control for an intelligent charging pile operation analysis system based on an AI large model proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A smart charging pile operation analysis system based on an AI large model includes: The operation data acquisition module is used to collect multi-source operation data of the target charging station within a continuous time window, and to preprocess the data to obtain a multi-source operation dataset. The propositional variable generation module is used to determine the constraint interval of continuous state variables in multi-source operation datasets and to encode the determination results into equipment-side propositional variables and site-side propositional variables to form an operation propositional variable set. The attribution clause generation module is used to input the set of operational proposition variables into the Tsetlin Machine attribution control model, drive the Tsetlin automaton to perform inclusion or exclusion state selection on the equipment-side text items and site-side text items in the set of operational proposition variables, and generate a set of attribution control logic clauses. The trigger matching attribution module is used to perform trigger matching on the set of labeled logical clauses based on the set of operational proposition variables, obtain the set of trigger clauses, and determine the attribution action result based on the integer weights of the logical clauses in the set of trigger clauses; The control command execution module is used to generate control commands based on the attribution action results and send the control commands to the charging pile operation execution terminal to form a control execution record; The residual feedback update module is used to collect target index data before and after the execution of control commands based on the control execution record, and generate execution residual feedback data based on the target index data. The large model report generation module is used to apply bidirectional feedback to the TsetlinMachine attribution control model corresponding to the trigger clause set using execution residual feedback data, update the text item status and integer weight of the logical clauses in the trigger clause set, and generate an operation analysis report of the target charging station based on the AI ​​large model.

[0021] In this embodiment, the operational data collection module includes: Data is continuously collected from the target charging station according to a preset time window to obtain raw multi-source operational data. The preset time window has a time window number, a window start time, and a window end time. The raw multi-source operational data is divided into equipment operation data, charging order data, site load data, electricity price data, weather data, and site revenue data according to the data source. The equipment operation data is collected according to the charging pile number, the charging order data is collected according to the order number and the charging pile number, the site load data is collected according to the charging station number, and the electricity price data, weather data, and site revenue data are collected according to the charging station number and the time window number. Preprocessing is performed on the original multi-source operation data. The preprocessing includes timestamp alignment, outlier removal, and missing sample value completion to obtain preprocessed multi-source operation data. The timestamp alignment is used to classify the sampling times of different data sources into corresponding preset time windows. The outlier removal is used to delete sample data that exceeds the value range of the corresponding data source or is inconsistent with adjacent sample values ​​within the same time window. The missing sample value completion is used to complete the missing data based on the valid sample data within the same charging pile, the same charging station, or adjacent time windows. The preprocessed multi-source operation data are associated according to the operation index identifier to obtain a multi-source operation dataset. The operation index identifier includes the charging station number, the charging pile number, and the time window number. Each data record in the multi-source operation dataset corresponds to a target charging station, a target charging pile, and a preset time window. Each data record also includes equipment operation data, charging order data, site load data, electricity price data, weather data, and site revenue data within the corresponding time window.

[0022] In this embodiment, the propositional variable generation module includes: Continuous state data on the device side and continuous state data on the site side are extracted from the multi-source operation dataset. The continuous state data on the device side are continuous numerical data reflecting the operating status of the target charging pile, and the continuous state data on the site side are continuous numerical data reflecting the load status and operating status of the target charging station. An operation anomaly category identifier is generated based on the same operation index identifier, and a correspondence is established between the operation anomaly category identifier and the continuous status quantity on the equipment side and the continuous status quantity on the site side, respectively. The operation anomaly category identifier is used to indicate the operation anomaly category of the target charging station within the corresponding time window. The CAIM algorithm is used to perform supervised discretization of continuous state variables on the equipment side and the site side. The operation anomaly category identifier is used as the category item, and each continuous state variable is used as the attribute item. The interdependence of the category attributes of the category items is calculated according to the candidate split points of the continuous state variables, and the candidate split points that increase the interdependence of the category attributes are selected as the constraint split points. Based on the constraint split points, device-side constraint discrete intervals and site-side constraint discrete intervals are generated respectively. The continuous state variables of the device side and the continuous state variables of the site side are mapped to the corresponding device-side constraint discrete intervals and site-side constraint discrete intervals respectively to obtain the interval hit results. The interval hit results include device-side interval hit results and site-side interval hit results. Binary encoding is performed on the interval hit results. The binary encoding is to set a proposition bit for each constraint discrete interval. When a continuous state variable falls into the corresponding constraint discrete interval, the proposition bit corresponding to the current constraint discrete interval is set to the first binary state, and the proposition bits of other constraint discrete intervals corresponding to the same continuous state variable are set to the second binary state, so as to obtain the device-side proposition variable and the site-side proposition variable. The device-side propositional variables and site-side propositional variables are associated with the same operational index identifier to form an operational propositional variable set.

[0023] This invention uses the CAIM algorithm to perform supervised discretization of continuous state variables on the equipment side and the site side based on the operational anomaly category identifier. The interval hit results are then binary encoded into equipment-side propositional variables and site-side propositional variables. This enables the charging pile equipment status, site load status, and operational anomaly category to form a computable propositional association, thereby improving the accuracy of continuous operation data representation, the adaptability of anomaly feature extraction, and the reliability of subsequent attribution control logic clause generation.

[0024] In this embodiment, the factorization sentence generation module includes: The operational proposition variable set is input into the Tsetlin Machine attribution control model, which includes a proposition partitioning unit, a text item construction unit, a cross-side co-occurrence binding unit, a Tsetlin automaton, an attribution control clause construction unit, a label weight binding unit, and a clause storage unit. The proposition partitioning unit partitions the operational proposition variable set according to the operational index identifier to obtain the device-side proposition variable set and the site-side proposition variable set. The text item construction unit generates positive and negative text items for each equipment-side proposition variable in the equipment-side proposition variable set, and generates positive and negative text items for each site-side proposition variable in the site-side proposition variable set. The positive, negative, positive, and negative text items for the equipment side are used together as attribution control text items. The cross-side co-occurrence binding unit calculates the cross-side co-occurrence strength between device-side and site-side text items according to the synchronous value relationship between device-side propositional variables and site-side propositional variables under the same operating index identifier. Device-side and site-side text items whose cross-side co-occurrence strength meets the preset binding conditions are grouped into cross-side text item candidate groups. The preset binding conditions are that the device-side and site-side text items correspond to the same operating index identifier, the number of synchronous values ​​within a continuous time window reaches the co-occurrence number threshold, and the cross-side co-occurrence strength of the two reaches the preset co-occurrence strength threshold. The cross-side text item candidate group is assigned to the Tsetlin automaton group. The Tsetlin automaton group performs inclusion or exclusion state selection for the device-side text items and site-side text items in the cross-side text item candidate group based on the cross-side co-occurrence strength and the current value in the operational proposition variable set. The inclusion or exclusion state selection is as follows: For each device-side text item and site-side text item in the cross-side text item candidate group, the Tsetlin automaton determines whether the current text item should enter the current attribution control logic clause based on the binary value of the current text item under the current operational index identifier, the cross-side co-occurrence strength, and the previous round of feedback state. When the Tsetlin automaton is in the inclusion state, the corresponding text item is added to the current attribution control logic clause. When the Tsetlin automaton is in the exclusion state, the corresponding text item is excluded from the current attribution control logic clause, thus obtaining the cross-side text item selection result. The attribution control clause construction unit constructs attribution control logic clauses based on the cross-side text item selection results. Each attribution control logic clause is formed by combining device-side text items and site-side text items in the included state, and each attribution control logic clause contains at least one device-side text item and at least one site-side text item. The tag weight binding unit binds the cause tag, action tag, and integer weight to the same attribution control logic clause. The cause tag is an anomaly cause identifier determined based on the operational anomaly category identifier and the device-side and site-side text items in the attribution control logic clause. The action tag is a control action identifier determined based on the cause tag and historical control execution records. The integer weight is a clause contribution value determined based on the historical trigger count, cause tag matching count, and action execution improvement count of the attribution control logic clause. The clause storage unit associates and stores different attribution control logic clauses according to clause number, operational index identifier and cross-side text item source, generating an attribution control logic clause set; The Tsetlin Machine attribution control model does not directly input operational proposition variables into a regular Tsetlin Machine for classification. Instead, it processes operational proposition variables separately according to the device side and the site side. By restricting the range of text item combinations through cross-side co-occurrence binding, each attribution control logic clause contains both device-side and site-side text items. Furthermore, it binds cause labels, action labels, and integer weights to the same attribution control logic clause, enabling the attribution control logic clause to be used simultaneously for anomaly cause determination, control action determination, and residual feedback update after being triggered.

[0025] This invention uses the Tsetlin Machine attribution control model to process operational propositional variables separately on the equipment side and the site side. It generates attribution control logic clauses through cross-side co-occurrence binding, inclusion-exclusion selection, and label weight binding, so that the cause of the anomaly, the control action, and the contribution value of the clause form a unified logical expression, which improves the accuracy of charging pile operation anomaly attribution, the consistency of control action generation, and the traceability of model feedback updates.

[0026] In this embodiment, the trigger matching attribution module includes: Based on the current operation index identifier, retrieve the operation proposition variable set and the attribution control logic clause set. Perform cross-side trigger determination for each attribution control logic clause. When both the device-side text item and the site-side text item in the same attribution control logic clause are satisfied by the operation proposition variable set, determine the current attribution control logic clause as a cross-side trigger clause. In the cross-side trigger determination process, attribution control logic clauses that are satisfied only on the device side or only on the site side are determined as invalid trigger clauses. This is to exclude single-side triggering cases and prevent the TsetlinMachine attribution control model from generating valid attribution results based solely on abnormal device-side or site-side states. Since the attribution control logic clauses in this invention are used to characterize the coupling relationship between the charging pile equipment operating state and the charging station operating state, only when both the device-side and site-side text items in the same attribution control logic clause are satisfied by the set of operating proposition variables will the corresponding attribution control logic clause be used as a cross-side trigger clause to participate in the determination of attribution action results. This reduces misjudgments caused by single device fluctuations or single site load changes on the attribution control results and improves the consistency between abnormal attribution and control action generation. Perform attribution action consistency check on cross-side trigger clauses, group cross-side trigger clauses with the same cause label and the same action label into the same attribution action candidate group, and determine the trigger contribution value of the current attribution action candidate group based on the integer weight of the cross-side trigger clauses and the number of cross-side text items in each attribution action candidate group. The candidate group of attribution actions with the largest trigger contribution value is determined as the target attribution action group. The cross-side trigger clauses in the target attribution action group are determined as the trigger clause set. Attribution action results are generated based on the cause label and action label corresponding to the target attribution action group.

[0027] This invention eliminates single-sided triggering scenarios where only the device side or only the site side is satisfied by cross-side triggering determination, ensuring that the attribution action result must originate from an attribution control logic clause that is satisfied by both the device-side and site-side text items. Furthermore, it determines the target attribution action group through cause labels, action labels, and integer weights, thereby improving the accuracy of charging pile operation anomaly attribution, the consistency of control action matching, and the anti-interference capability of single device fluctuations on the attribution result.

[0028] In this embodiment, the control command execution module includes: Read the attribution action results, and verify the clause number, integer weight, and cross-side text item source of the cross-side trigger clauses in the trigger clause set according to the current operation index identifier to form the basis for control generation; The control execution elements are determined based on the action tags, and the attribution source of the control command is determined based on the cause tags. The control execution elements include the control action type, the scope of the control object, the range of control parameters, and the execution time limit. The determination process includes matching the action tags with the action identifiers in the action mapping relationship to obtain the control action type, determining the scope of the control object based on the charging station number and charging pile number in the current operation index identifier, determining the control parameter values ​​within the control parameter range based on the trigger contribution value, and determining the execution time limit based on the action priority corresponding to the action tag and the running status of the control object. The scope of the control object includes the target charging station, the target charging pile, and the charging pile operation execution terminal corresponding to the current operation index identifier. The control parameter range is the boundary of parameter values ​​that the charging pile operation execution terminal can execute. The action mapping relationship is the correspondence between the action tags and the control execution elements. Control commands are generated based on the control generation criteria and control execution elements. The control commands are then sent to the charging pile operation execution terminal, and a control execution record is formed. The control execution record includes the time of command issuance, the execution terminal's reception status, the execution terminal's response data, and the control object identifier.

[0029] In this embodiment, the residual feedback update module includes: Read the control execution record and trigger clause set. Based on the device-side text item source and site-side text item source of each cross-side trigger clause in the trigger clause set, reversely determine the target indicator item corresponding to each cross-side trigger clause to form a clause indicator mapping relationship. The clause indicator mapping relationship includes clause number, device-side text item source, site-side text item source, device-side target indicator item, site-side target indicator item, and indicator improvement direction. Based on the control object identifier and action label in the control execution record, a pre-observation window, a post-observation window, and a control observation window are established. The control observation window corresponds to a control operation index identifier that has not received the same control instruction and has the same station type, time period type, and weather interval as the current operation index identifier. Pre-observation target indicator data, post-observation target indicator data, and control target indicator data are collected respectively to generate net action effect target indicator data. According to the clause indicator mapping relationship, the net effect target indicator data of the action is converted into execution residual feedback data oriented to the trigger clause set. The conversion includes writing the net effect target indicator data of the action into the corresponding cross-side trigger clause in the trigger clause set according to the clause indicator mapping relationship, comparing the direction of change of the net effect of the action of the equipment-side target indicator item with the direction of improvement of the equipment-side indicator to generate an equipment-side improvement identifier, comparing the direction of change of the net effect of the action of the site-side target indicator item with the direction of improvement of the site-side indicator to generate a site-side improvement identifier, generating a cross-side consistent improvement identifier and a sub-side feedback identifier based on the equipment-side improvement identifier and the site-side improvement identifier, and combining the current operation index identifier, trigger clause set, cause label, action label, equipment-side improvement identifier, site-side improvement identifier, cross-side consistent improvement identifier, and sub-side feedback identifier into execution residual feedback data.

[0030] In this embodiment, the large model report generation module includes: According to the current operation index identifier, the residual feedback data will be executed back to the TsetlinMachine attribution control model corresponding to the trigger clause set, and the attribution control logic clause, device-side text item, site-side text item and integer weight corresponding to the trigger clause set will be located in the Tsetlin Machine attribution control model to generate clause-level sub-side feedback records. Based on the clause-level cross-side feedback records, bidirectional feedback is applied to the Tsetlin Machine attribution control model corresponding to the trigger clause set. The application results of the bidirectional feedback include: when the cross-side consistent improvement indicator is established, enhanced feedback is applied to both the device-side text item and the site-side text item; when the device-side improvement indicator is established and the site-side improvement indicator is not established, enhanced feedback is applied to the device-side text item and suppressed feedback is applied to the site-side text item; when the site-side improvement indicator is established and the device-side improvement indicator is not established, enhanced feedback is applied to the site-side text item and suppressed feedback is applied to the device-side text item; when neither the device-side improvement indicator nor the site-side improvement indicator is established, suppressed feedback is applied to both the device-side text item and the site-side text item, thus obtaining the application results of the bidirectional feedback. Based on the results of the two-way feedback, update the Tsetlin automaton literal state and integer weights corresponding to the attribution control logic clauses in the trigger clause set. Construct a large model evidence input package based on the updated literal state, integer weights, trigger clause set, attribution action results, control execution records, and execution residual feedback data, and input it into the AI ​​large model to generate an operation analysis report for the target charging station. The AI ​​model is a generative artificial intelligence model with natural language understanding, structured evidence induction, and text generation capabilities. It receives the trigger clause set, attribution action results, control execution records, execution residual feedback data, text item status update results, and integer weight update results output by the Tsetlin Machine attribution control model. It converts the above structured data into an operation analysis report for charging pile operation management, which explains the cause of the anomaly at the target charging station, the scope of the affected charging piles, the effect of control action execution, the improvement status on the equipment side and the site side, and the basis for the next round of operation and disposal. The cause label, action label, control command, and two-way feedback results are not changed during the report generation process. The operational analysis report is used to determine whether the equipment availability, resource utilization, user waiting status, order fulfillment status, load distribution, and operational revenue of the target charging station have improved after the execution of control commands, and to provide a basis for power allocation, charging pile scheduling, electricity price adjustment, operation and maintenance dispatch, and anomaly review.

[0031] This invention executes the Tsetlin Machine attribution control model corresponding to the residual feedback data back-point trigger clause set, and applies bidirectional feedback to the device-side and site-side text items. This allows the text item status and integer weights to be updated based on the actual control execution effect. At the same time, the AI ​​big model transforms the model update results, control execution results, and operational improvement results into a traceable operational analysis report, thereby improving the accuracy of charging pile anomaly attribution, the effectiveness of control feedback, and the efficiency of operational handling decisions.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to an operational analysis scenario of a public charging station with fast charging zones, slow charging zones, and a centralized operation backend. The station is equipped with 36 DC fast charging piles and 24 AC slow charging piles. The backend collects data on equipment operation, charging orders, station load, electricity price, weather, and revenue in 15-minute continuous time windows, forming 2880 time windows. During high-load periods, issues such as output power fluctuations, order interruptions, increased queuing times, and decreased single-pile availability arise. Traditional rule-based threshold systems struggle to identify anomalies caused by the combined effects of equipment status, station load, and resource occupancy.

[0033] The propositional variable generation module extracts continuous state variables from both the equipment and site sides, and uses the CAIM algorithm with operational anomaly category identifiers as category items for supervised discretization. In implementation, the output power fluctuation rate interval boundaries are 8.5% and 17.0%, the fast charging resource occupancy rate interval boundaries are 65.0% and 85.0%, and the average waiting time interval boundaries are 8 minutes and 15 minutes. The interval hit results are binary encoded to form equipment-side and site-side propositional variables, resulting in an operational propositional variable set.

[0034] The attribution clause generation module inputs the set of operational propositional variables into the Tsetlin Machine attribution control model. The model processes the device-side and site-side propositional variables separately and calculates the cross-side co-occurrence strength. During implementation, 186 device-side text items and 142 site-side text items were generated, forming 218 cross-side text item candidate groups, resulting in 312 attribution control logic clauses. Taking a DC fast charging pile as an example, the output power fluctuation rate increased from 6.2% to 21.8% within four consecutive time windows, while the occupancy rate of the fast charging area reached 92.6% during the same period, and the average waiting time reached 18.7 minutes. The model triggered 9 cross-side trigger clauses, 6 of which pointed to the congestion of fast charging resources leading to a decrease in equipment availability.

[0035] After the control command was executed, the system downgraded the scheduling priority of 3 fast charging piles with power fluctuation rates exceeding 17.0% by one level, upgraded the scheduling priority of 6 fast charging piles with stable output and idle time exceeding 6 minutes by one level, and adjusted the continuous order interval for a single pile to 2 minutes. Within the post-observation window, the average occupancy rate of the fast charging area decreased from 92.6% to 84.3%, the average number of vehicles in queue decreased from 14 to 9, the average waiting time decreased from 18.7 minutes to 12.4 minutes, and the number of order interruptions decreased from 11 times per window to 5 times. The residual feedback update module generated equipment-side improvement indicators, site-side improvement indicators, and cross-side consistency improvement indicators, and returned the text items and integer weights in the trigger clause set.

[0036] In this embodiment, a traditional rule-based threshold analysis method is used as a control scheme, and the system described in this invention is used as a verification scheme. Both schemes use the same multi-source operational data input. The traditional rule-based threshold analysis method determines the cause of anomalies based on fixed alarm thresholds and manual reports, while the system described in this invention generates attribution action results based on equipment-side propositional variables, site-side propositional variables, cross-side attribution control logic clauses, and execution residual feedback data. Specific comparison data is shown in Table 1: Table 1 Comparison of the operational effectiveness of smart charging piles under different operational analysis methods.

[0037] As shown in Table 1, under the same multi-source operational data input conditions, the system of this invention achieves higher accuracy in operational anomaly identification, anomaly cause attribution, and control action matching compared to the traditional rule-based threshold analysis method. Specifically, the accuracy rate of operational anomaly identification increased from 78.6% to 91.8%, the accuracy rate of anomaly cause attribution increased from 72.4% to 89.5%, and the accuracy rate of control action matching increased from 69.8% to 86.9%. Simultaneously, the false positive rate of single-sided triggering decreased from 18.9% to 6.7%, indicating that cross-sided triggering judgment can reduce reliance solely on settings. Misjudgments caused by standby or site status; In terms of operational effectiveness, the average user queuing time decreased from 16.8 minutes to 11.3 minutes, the order interruption rate decreased from 5.6% to 3.1%, and the average availability of charging piles increased from 92.1% to 96.4%, indicating that the present invention can improve the operational stability and service efficiency of charging piles; The traceability rate of the report conclusions increased from 76.2% to 95.8%, indicating that the operational analysis report generated by the AI ​​big model based on trigger clauses, control execution records, and execution residual feedback data has a better evidence correspondence.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart charging pile operation analysis system based on an AI large-scale model, characterized in that, include: The operation data acquisition module is used to collect multi-source operation data of the target charging station within a continuous time window, and to preprocess the data to obtain a multi-source operation dataset. The propositional variable generation module is used to determine the constraint interval of continuous state variables in multi-source operation datasets and to encode the determination results into equipment-side propositional variables and site-side propositional variables to form an operation propositional variable set. The attribution clause generation module is used to input the set of operational proposition variables into the Tsetlin Machine attribution control model, drive the Tsetlin automaton to perform inclusion or exclusion state selection on the equipment-side text items and site-side text items in the set of operational proposition variables, and generate a set of attribution control logic clauses. The trigger matching attribution module is used to perform trigger matching on the set of labeled logical clauses based on the set of operational proposition variables, obtain the set of trigger clauses, and determine the attribution action result based on the integer weights of the logical clauses in the set of trigger clauses; The control command execution module is used to generate control commands based on the attribution action results and send the control commands to the charging pile operation execution terminal to form a control execution record; The residual feedback update module is used to collect target index data before and after the execution of control commands based on the control execution record, and generate execution residual feedback data based on the target index data. The large model report generation module is used to apply bidirectional feedback to the TsetlinMachine attribution control model corresponding to the trigger clause set using execution residual feedback data, update the text item status and integer weight of the logical clauses in the trigger clause set, and generate an operation analysis report of the target charging station based on the AI ​​large model.

2. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The operational data collection module includes: Data is continuously collected from the target charging station according to the preset time window to obtain raw multi-source operation data. Preprocessing is performed on the original multi-source operation data, including timestamp alignment, outlier removal, and missing sample value completion, to obtain preprocessed multi-source operation data. The preprocessed multi-source operational data are associated according to the operational index identifier to obtain a multi-source operational dataset.

3. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The propositional variable generation module includes: Extract continuous state variables from the device side and the site side from the multi-source operation dataset; An operational anomaly category identifier is generated based on the same operational index identifier, and a correspondence is established between the operational anomaly category identifier and the continuous status quantity on the equipment side and the continuous status quantity on the site side, respectively. The CAIM algorithm is used to perform supervised discretization of continuous state variables on the equipment side and the site side. The operation anomaly category identifier is used as the category item, and each continuous state variable is used as the attribute item. The interdependence of the category attributes of the category items is calculated according to the candidate split points of the continuous state variables, and the candidate split points that increase the interdependence of the category attributes are selected as the constraint split points. Based on the constraint split points, generate the equipment-side constraint discrete interval and the site-side constraint discrete interval respectively, and map the equipment-side continuous state variables and the site-side continuous state variables to the corresponding equipment-side constraint discrete interval and the site-side constraint discrete interval respectively to obtain the interval hit result; Binary encoding is performed on the interval hit results to obtain the device-side propositional variables and the site-side propositional variables; The device-side propositional variables and site-side propositional variables are associated with the same operational index identifier to form an operational propositional variable set.

4. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The factorization sentence generation module includes: The operational proposition variable set is input into the Tsetlin Machine attribution control model, which includes a proposition partitioning unit, a text item construction unit, a cross-side co-occurrence binding unit, a Tsetlin automaton, an attribution control clause construction unit, a label weight binding unit, and a clause storage unit. The proposition partitioning unit partitions the operational proposition variable set according to the operational index identifier to obtain the device-side proposition variable set and the site-side proposition variable set. The text item construction unit generates positive and negative text items for each equipment-side proposition variable in the equipment-side proposition variable set, and generates positive and negative text items for each site-side proposition variable in the site-side proposition variable set. The positive, negative, positive, and negative text items for the equipment side are used together as attribution control text items. The cross-side co-occurrence binding unit calculates the cross-side co-occurrence strength between device-side and site-side text items according to the synchronization value relationship between device-side propositional variables and site-side propositional variables under the same operation index identifier, and forms a cross-side text item candidate group with the cross-side co-occurrence strength meeting the preset binding conditions. The cross-side text item candidate group is assigned to the Tsetlin automaton group, and the Tsetlin automaton group performs inclusion or exclusion state selection on the device-side text items and site-side text items in the cross-side text item candidate group based on the cross-side co-occurrence strength and the current value in the operational proposition variable set, so as to obtain the cross-side text item selection result. The attribution control clause construction unit constructs attribution control logic clauses based on the cross-border text item selection results; The label weight binding unit binds the cause label, action label, and integer weight to the same attribution control logic clause; The clause storage unit associates and stores different attribution control logic clauses according to clause number, operational index identifier, and cross-side text item source, generating an attribution control logic clause set.

5. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The trigger matching attribution module includes: Based on the current operation index identifier, retrieve the operation proposition variable set and the attribution control logic clause set. Perform cross-side trigger determination for each attribution control logic clause. When both the device-side text item and the site-side text item in the same attribution control logic clause are satisfied by the operation proposition variable set, determine the current attribution control logic clause as a cross-side trigger clause. Perform attribution action consistency check on cross-side trigger clauses, group cross-side trigger clauses with the same cause label and the same action label into the same attribution action candidate group, and determine the trigger contribution value of the current attribution action candidate group based on the integer weight of the cross-side trigger clauses and the number of cross-side text items in each attribution action candidate group. The candidate group of attribution actions with the largest trigger contribution value is determined as the target attribution action group. The cross-side trigger clauses in the target attribution action group are determined as the trigger clause set. Attribution action results are generated based on the cause label and action label corresponding to the target attribution action group.

6. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The control command execution module includes: Read the attribution action results, and verify the clause number, integer weight, and cross-side text item source of the cross-side trigger clauses in the trigger clause set according to the current operation index identifier to form the basis for control generation; The control execution elements are determined based on the action labels, and the attribution source of the control instructions is determined based on the cause labels; Control commands are generated based on the control generation criteria and control execution elements, and then sent to the charging pile operation execution terminal to form a control execution record.

7. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The residual feedback update module includes: Read the control execution record and trigger clause set. Based on the device-side text item source and site-side text item source of each cross-side trigger clause in the trigger clause set, reversely determine the target indicator item that corresponds one-to-one with the cross-side trigger clause to form a clause indicator mapping relationship. Based on the control object identifier and action label in the control execution record, establish a pre-observation window, a post-observation window, and a control observation window, and collect pre-target indicator data, post-target indicator data, and control target indicator data respectively to generate net action effect target indicator data; Based on the clause index mapping relationship, the net effect target index data of the action is converted into execution residual feedback data oriented towards the trigger clause set.

8. The intelligent charging pile operation analysis system based on an AI large model according to claim 1, characterized in that, The large model report generation module includes: According to the current operation index identifier, the residual feedback data will be executed back to the TsetlinMachine attribution control model corresponding to the trigger clause set, and the attribution control logic clause, device-side text item, site-side text item and integer weight corresponding to the trigger clause set will be located in the Tsetlin Machine attribution control model to generate clause-level sub-side feedback records. Based on the clause-level feedback records, bidirectional feedback is applied to the Tsetlin Machine attribution control model corresponding to the trigger clause set to obtain the application result of the bidirectional feedback; Based on the results of the two-way feedback, the Tsetlin automaton literal state and integer weights corresponding to the attribution control logic clauses in the trigger clause set are updated. A large model evidence input package is constructed based on the updated literal state, integer weights, trigger clause set, attribution action results, control execution records, and execution residual feedback data. This package is then input into the AI ​​large model to generate an operational analysis report for the target charging station.

Images