Artificial intelligence-based business platform process optimization method and system

By extracting features from the power business middleware process data and coordinating optimization with intelligent optimization models, the problem of low efficiency in power business middleware process management has been solved, realizing intelligent analysis and optimization of processes, and improving execution efficiency and resource utilization.

CN120875509APending Publication Date: 2025-10-31INNER MONGOLIA ELECTRIC POWER (GRP) CO LTD DIGITAL RES BRANCH
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202511031250.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-25
Publication Date
2025-10-31

AI Technical Summary

Technical Problem

The existing power business middleware process management method lacks the ability to deeply mine and intelligently analyze process data during actual operation, resulting in low process execution efficiency, waste of resources, difficulty in dynamically adjusting processes based on real-time data, and inability to promptly identify and resolve potential problems.

Method used

By acquiring process data during the operation of the power business middle platform, process features are extracted, a pre-built intelligent optimization model for power processes is invoked for collaborative optimization reasoning, a set of optimization constraints is generated, and deployed to the process execution engine. Operational feedback data is collected to iteratively update model parameters, adjust node execution timing, and reconstruct node dependencies.

Benefits of technology

It enables intelligent analysis and optimization decisions for processes, improves process execution efficiency, reduces resource waste, and continuously enhances the accuracy and effectiveness of process optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120875509A_ABST
    Figure CN120875509A_ABST
Patent Text Reader

Abstract

The invention provides an artificial intelligence-based business platform process optimization method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: firstly obtaining an electric power process data set generated by the operation of an electric power business platform, and enabling the electric power process data set to comprise a plurality of process instances composed of business operation nodes and node execution logs; performing flow feature extraction on the electric power flow data set to obtain node behavior features and node association features, calling a pre-constructed electric power flow intelligent optimization model to perform collaborative optimization reasoning, and generating an optimization constraint condition set; based on the optimization constraint condition set, generating an electric power flow optimization strategy including node execution time sequence adjustment and a node dependency relationship reconstruction scheme, deploying the electric power flow optimization strategy to a flow execution engine, then collecting flow operation feedback data after implementation, and iteratively updating model reasoning parameters, so as to obtain an electric power flow optimization strategy; therefore, intelligent optimization of the process is realized based on actual data, the process execution efficiency is improved, and a dynamic intelligent power business process management system is constructed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a business middle platform process optimization method and system based on artificial intelligence. Background Technology

[0002] In the digital transformation of the power industry, the power business middleware platform, as a core support platform, bears the important mission of integrating various power business resources and achieving efficient business collaboration. However, with the continuous expansion and increasing complexity of power businesses, the business processes carried by the power business middleware platform are becoming increasingly numerous and complex. Existing power business process management methods are mostly based on preset fixed rules and human experience for process design and optimization, lacking the ability to deeply mine and intelligently analyze process data during actual operation.

[0003] In actual operation, power business processes are affected by various factors, such as changes in equipment status, fluctuations in user demand, and external environmental interference, leading to frequent problems such as low process execution efficiency and resource waste. For example, certain business operation nodes may become process bottlenecks due to excessively long execution times, or unreasonable trigger dependencies between operation nodes may cause excessively long process waiting times. Traditional methods are unable to dynamically adjust processes based on real-time data, cannot promptly identify and resolve potential problems in the processes, and fail to meet the power business's needs for efficient, flexible, and intelligent process management. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a business middleware process optimization method based on artificial intelligence, the method comprising: Acquire a set of power process data generated during the operation of the power business middleware. The set of power process data contains multiple process instances. Each process instance consists of continuously executed business operation nodes and corresponding node execution logs. The node execution logs include the operation start time, operation end time, operation triggering conditions, and operation result status. Process features are extracted from the power process data set to obtain the node behavior features and node association features of each process instance. The node behavior features include the execution time distribution pattern of operation nodes, and the node association features include the triggering dependency relationship between operation nodes. The pre-built intelligent optimization model for power processes is invoked to perform collaborative optimization reasoning on the node behavior characteristics and node association characteristics, generating a set of optimization constraints for the process instance; Based on the set of optimization constraints, a power process optimization strategy is generated that includes a node execution timing adjustment scheme and a node dependency reconstruction scheme. The power process optimization strategy is deployed to the process execution engine of the power business platform, the process operation feedback data after the implementation of the power process optimization strategy is collected, and the inference parameters of the power process intelligent optimization model are iteratively updated through the process operation feedback data.

[0005] In another aspect, embodiments of the present invention also provide a business middleware process optimization system based on artificial intelligence, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code. The processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0006] Based on the above, by acquiring power process data sets during the operation of the power business middleware platform, and extracting process features from these data sets, it is possible to deeply analyze the behavioral and relational characteristics of nodes in each process instance. This effectively presents the distribution patterns of execution time of operation nodes and the triggering dependencies between nodes. A pre-built intelligent optimization model for power processes is invoked for collaborative optimization reasoning, generating a set of optimization constraints for each process instance. This enables intelligent analysis and optimization decision-making for the process, allowing for the development of targeted optimization strategies based on the characteristics of different process instances. Power process optimization strategies are generated based on the set of optimization constraints and deployed to the process execution engine. This effectively adjusts the execution sequence of nodes and reconstructs node dependencies, significantly improving process execution efficiency and reducing resource waste. Simultaneously, by collecting process operation feedback data and iteratively updating the model's inference parameters, the model can continuously adapt to changes in power business, continuously improving the accuracy and effectiveness of process optimization. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the business middle platform process optimization method based on artificial intelligence provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of the business middle platform process optimization system based on artificial intelligence provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a business middle platform process optimization method based on artificial intelligence, provided in one embodiment of the present invention. The following is a detailed description of this business middle platform process optimization method based on artificial intelligence.

[0010] Step S110: Obtain the power process data set generated during the operation of the power business middleware. The power process data set contains multiple process instances. Each process instance consists of continuously executed business operation nodes and corresponding node execution logs. The node execution logs include the operation start time, operation end time, operation triggering conditions, and operation result status.

[0011] Taking a large-scale commercial area's power operation platform as an example, this commercial area encompasses numerous shopping malls, office buildings, and entertainment venues, resulting in diverse and complex power supply demands. In daily power operations, a large amount of relevant data is generated from power supply assurance work in different areas of the commercial area. For instance, the commercial area is divided into different functional zones, such as shopping areas, office areas, and leisure areas. Power supply-related operations, such as equipment maintenance and fault handling, are performed in each functional area, forming a series of process instances.

[0012] Taking a typical power equipment maintenance process in a shopping area as an example, this includes continuously executed business operation nodes such as equipment inspection, fault diagnosis, and parts replacement. For the equipment inspection operation node, the node execution log records the start time, i.e., the time when the inspector begins to check the equipment; the end time, i.e., the time when the inspection work is completed; the operation triggering condition may be based on the equipment's running time since the last inspection reaching a set time, or according to the periodic inspection requirements stipulated by the commercial area's power management regulations; the operation result status may be that the equipment is operating normally, minor abnormal signs are found, or there is a potential fault risk, etc. Numerous similar process examples and their node execution logs together constitute the power process data set.

[0013] Step S120: Extract process features from the power process data set to obtain node behavior features and node association features for each process instance. The node behavior features include the distribution pattern of the execution time of the operation nodes, and the node association features include the triggering dependency relationship between the operation nodes.

[0014] In the context of a large-scale commercial area power business platform, in order to better optimize power business processes, it is necessary to perform in-depth feature extraction operations on the acquired power process data set.

[0015] Step S121: Perform type classification processing on the process instances in the power process data set, and divide the process instances into equipment inspection type, fault handling type and electricity bill settlement type.

[0016] In the power business scenario of this commercial area, there are various types of power business processes. Equipment inspection processes are designed to ensure the continuous and stable operation of power equipment. These involve comprehensive inspections of various power equipment according to set cycles, such as regular inspections of transformers and distribution boxes. Examples of these inspection processes fall under the equipment inspection type. Fault handling processes are examples of the series of operational procedures from fault detection to fault resolution when power equipment malfunctions, such as the emergency repair process for a sudden power outage in a shopping mall. Electricity billing processes are examples of the processes that periodically calculate the electricity consumption of each user in the commercial area and generate electricity bills, involving operations such as electricity data collection and electricity bill calculation. Taking a regular inspection of the power equipment in the office area as an example, this process example clearly falls under the equipment inspection type.

[0017] Step S122: Perform time series alignment processing on the node execution logs of each type of process instance, map the start time and end time of the operation to a unified time axis, and generate node timeline data containing time axis coordinates.

[0018] Taking an equipment inspection process as an example, suppose an inspection of electrical equipment in different areas of a commercial district includes multiple operation nodes. For the visual inspection node of electrical equipment in an office building, the start time of the operation is marked as a specific time point, and the end time is marked as another specific time point; for the performance testing node of the same office building's electrical equipment, the start time is yet another specific time point, and the end time is yet another specific time point. These times are uniformly mapped onto a timeline with a specific start time as the zero point, such as midnight of the current day. The duration from midnight to the start time of the visual inspection node, and the duration to the end time, and the duration from midnight to the start time of the performance testing node, and so on, are all mapped onto this unified timeline. This generates node timeline data containing these timeline-based duration coordinates, effectively showing the sequential order and positional relationship of each operation node in the time dimension.

[0019] Step S123: Perform sliding window analysis on the node timeline data, extract continuous node time segments with a preset time window length, calculate the average execution time, longest execution time and shortest execution time of the node in each time segment, and generate basic time series statistics of node behavior characteristics.

[0020] Step S1231: Extract the time series of the operation start time and the time series of the operation end time from the node timeline data.

[0021] Continuing with the example of the node timeline data for the inspection of power equipment in different areas of the commercial district, the start times of all operation nodes are organized into a time series from this node timeline data. At the same time, the end times of all operation nodes are organized into another time series. These time series record the specific position of the start and end times of each operation node on a unified time axis.

[0022] Step S1232: Calculate the execution time of each operation node. The execution time is the time difference between the operation end time and the operation start time.

[0023] For corresponding elements in the two extracted time series, the time difference is calculated. For example, for the first operation node (appearance inspection node), its execution time is the time obtained by subtracting the start time from the end time of the operation; for the second operation node (performance testing node), its execution time is the time obtained by subtracting the start time from the end time of the operation. Through the above calculation, the execution time of each operation node in this inspection process is obtained, thus reflecting the actual time spent by each operation node.

[0024] Step S1233: Arrange the execution duration sequence according to the time axis to form a continuous execution duration sequence.

[0025] The calculated execution times of each operation node are arranged in chronological order along the timeline. Assuming there are other operation nodes in the entire inspection process, the execution times of each operation node are arranged sequentially according to the order of their operations into a continuous execution time series. This execution time series effectively demonstrates the sequential relationship of the execution times of each operation node in the entire inspection process.

[0026] Step S1234: Use a sliding window to slide over the long time sequence during execution. The sliding window slides one node at a time, and the coverage length of the sliding window is the preset time window length.

[0027] A preset time window length is set, and a sliding operation is performed on the aforementioned continuous execution time series based on this time window length. The step size for each slide is set to one node, meaning that the window moves the position of one operation node at a time. For example, starting from the execution time of the first operation node, the window covers the execution times of the first and second operation nodes, then the window moves one node to cover the execution times of the second and third operation nodes, and so on, to perform a comprehensive traversal analysis of the entire execution time series in this way.

[0028] Step S1235: Perform statistical calculations on the execution time data within each sliding window to obtain the average, maximum, and minimum execution time within the sliding window.

[0029] For the execution duration data contained within each sliding window, statistical calculations are performed. For example, when the window covers the execution durations of the first and second operation nodes, the average of these two durations is calculated by adding them together and then dividing by two; simultaneously, the maximum and minimum values ​​of these two durations are identified. When the window moves to cover the execution durations of the second and third operation nodes, the same calculations are performed to obtain the average, maximum, and minimum execution durations within that window. By calculating for each sliding window, key statistical information on the execution duration within different time segments is obtained.

[0030] Step S1236: Associate the average, maximum and minimum values ​​with the starting node identifier corresponding to the sliding window and store them to generate basic time series statistics entries containing node identifier, time window position, average execution time, longest execution time and shortest execution time.

[0031] For each sliding window, the calculated average, maximum, and minimum values ​​are associated with the identifier of the starting node corresponding to that sliding window. For example, after the window covers the execution times of the first and second operation nodes and the statistical values ​​are calculated, these statistical values ​​are associated with the identifier of the first operation node. The node identifier, the position information of the sliding window in the entire time series, and the calculated average execution time, longest execution time, and shortest execution time are recorded to form a basic time series statistical entry. In this way, a corresponding entry is generated for each sliding window, and the entry records in detail the key characteristics of the execution time of the operation nodes within each time segment and the corresponding node information.

[0032] Step S1237: Arrange the basic time series statistics entries according to the execution order of the process instance to generate a basic time series statistics sequence with a time order relationship.

[0033] The generated basic time-series statistics are arranged according to the actual execution order of the operation nodes in the process instance. For example, in the entire inspection process instance, the corresponding basic time-series statistics are arranged in the order in which the operation nodes are performed, thereby generating a basic time-series statistics sequence with an effective time sequence relationship. This basic time-series statistics sequence fully presents the changes in the execution time of the operation nodes in different time segments of the entire process instance.

[0034] Step S124: Perform semantic parsing on the operation triggering conditions in the node execution log, extract the set of preceding node identifiers in the triggering conditions, and construct a directed graph of triggering dependencies between nodes. The directed graph of triggering dependencies contains node identifiers and triggering edges.

[0035] In the scenario of equipment inspection process in a large-scale commercial power business platform, an in-depth analysis is conducted on the operation triggering conditions in the node execution log.

[0036] Step S1241: Extract the operation trigger condition field from the node execution log. The operation trigger condition field contains trigger rules described in natural language.

[0037] Taking a specific operation node in the equipment inspection process as an example, we can find the field recording the operation trigger conditions in its node execution log. For instance, in the node execution log for a certain test operation on electrical equipment in a shopping mall, the operation trigger condition field records trigger rules described in natural language, such as "This test on the equipment in this mall can only be performed after the comprehensive performance test of the equipment in the previous area is completed and the test results are normal." By extracting the above field, we can obtain the trigger condition information for each operation node.

[0038] Step S1242: Perform natural language processing on the triggering rule to identify the logical keywords and node identifiers therein.

[0039] The extracted natural language descriptions of the triggering rules are analyzed using natural language processing techniques. For example, for the rule "This test can only be performed on the equipment in this mall after the comprehensive performance test of the equipment in the previous area is completed and the test results are normal," analysis can identify logical keywords such as "and," as well as the node identifier corresponding to "comprehensive performance test of the equipment in the previous area." Through the above method, logical keywords and related node identifiers are accurately identified from each triggering rule.

[0040] Step S1243: Based on the logical keywords and node identifiers, determine the prerequisite nodes that need to be completed in the triggering conditions, and form a set of prerequisite node identifiers.

[0041] Based on the identified logical keywords and node identifiers, further analysis is performed to determine the prerequisite nodes that must be completed to trigger the current operation node. For example, based on the triggering rules and the identified information, the node "comprehensive performance testing of the previous area's equipment" is determined to be a prerequisite node for the current operation node, and its node identifier is included in the prerequisite node identifier set. The triggering rules for each operation node are analyzed and processed in the above manner, forming a prerequisite node identifier set for each operation node. These prerequisite node identifier sets clarify the prerequisite condition relationships between each operation node.

[0042] Step S1244: Create a graph node for each operation node, and the identifier of the graph node is consistent with the identifier of the operation node.

[0043] During the construction of the trigger dependency directed graph, a corresponding graph node is created for each operation node in the equipment inspection process. For example, for the operation node "a certain inspection of electrical equipment in a shopping mall", a graph node is created, and the identifier of the graph node is set to be completely consistent with the identifier of the operation node to ensure effective correspondence.

[0044] Step S1245: Add a directed edge between the previous node in each set of previous node identifiers and the current operation node. The starting point of the directed edge is the previous node identifier, and the ending point is the current operation node identifier.

[0045] For each operation node's set of preceding node identifiers, directed edges are added between the created graph nodes. For example, if an operation node's set of preceding node identifiers includes a preceding node called "Comprehensive Performance Detection of Devices in the Previous Region," a directed edge is added between the graph node representing "Comprehensive Performance Detection of Devices in the Previous Region" and the graph node representing the current operation node. The starting point of this directed edge is the identifier of the "Comprehensive Performance Detection of Devices in the Previous Region" graph node, and the ending point is the identifier of the current operation node's graph node. Through the above operations, the triggering dependency direction between nodes is clarified.

[0046] Step S1246: Perform the above processing on the triggering conditions of all operation nodes to generate a trigger dependency directed graph containing graph nodes and directed edges. The trigger dependency directed graph is used to represent the triggering order constraint relationship between operation nodes.

[0047] All operation nodes in the equipment inspection process are processed according to the above steps. From extracting and analyzing the trigger conditions of each operation node, to creating graph nodes and adding directed edges, a complete trigger dependency directed graph is finally generated. This trigger dependency directed graph contains all the graph nodes corresponding to the operation nodes and the directed edges between them that represent the trigger dependency relationship, effectively showing the trigger order constraint relationship between operation nodes in the entire equipment inspection process.

[0048] Step S125: Perform path traversal processing on the trigger dependency directed graph, count all possible trigger paths from the starting node to the ending node, calculate the trigger success probability of each trigger path, and generate the dependency path probability distribution of node association features.

[0049] Further analysis is conducted based on the trigger-dependent directed graph constructed by the equipment inspection process of the power business platform in a large-scale commercial area.

[0050] Starting from the initial node of the trigger-dependent directed graph, a path traversal algorithm is used to fully traverse the graph. For example, assuming the initial node is "Start inspecting a certain area of ​​the commercial district", the algorithm explores all possible paths along the directed edges sequentially until the terminal node is reached, such as "Complete the inspection of all equipment in this area". During the traversal, every possible trigger path from the initial node to the terminal node is recorded.

[0051] For each recorded trigger path, its success probability is estimated based on historical data or relevant experience. For example, for a particular path, the execution status of the operation nodes involved in that path is referenced in previous similar inspection processes, and various factors are considered to estimate the success probability of triggering this path. The above probability calculation is performed on all statistically recorded trigger paths.

[0052] Each triggering path and its corresponding success probability are organized to form a dependency path probability distribution, which effectively shows the distribution of success probability of different triggering paths in the equipment inspection process.

[0053] Step S126: Standardize and integrate the basic time-series statistics and the probability distribution of dependent paths to generate corresponding node behavior features and node association features.

[0054] In the context of a large-scale commercial power business platform, the existing basic time-series statistics and dependency path probability distributions are further processed to generate key features.

[0055] First, the basic time-series statistics need to be standardized. Since these statistics contain different types of data, such as average execution time, longest execution time, and shortest execution time, they may have different dimensions and ranges. To effectively integrate these data, standardization is necessary. For example, using standardization methods from relevant technologies, the average execution time, longest execution time, and shortest execution time can be transformed into a defined numerical range, making the data comparable in terms of dimensions and scale.

[0056] The dependency path probability distribution is also processed accordingly. Then, the standardized basic time-series statistics and dependency path probability distribution are integrated according to their corresponding dimensional positions to generate corresponding node behavior features and node association features. The node behavior features and node association features comprehensively reflect the patterns in the execution duration of operating nodes and the probability distribution of triggering dependencies between nodes.

[0057] Step S130: Call the pre-built intelligent optimization model of power process to perform collaborative optimization reasoning on the node behavior characteristics and node association characteristics, and generate a set of optimization constraints for the process instance.

[0058] In the scenario of a large-scale commercial area power business middle platform, the extracted node behavior features and node association features are used to conduct in-depth analysis and reasoning with the help of a pre-built intelligent power process optimization model to generate the set of constraints required for optimization.

[0059] Step S131: Input the node behavior features and node association features into the feature fusion layer of the power process intelligent optimization model, calculate the association weight between the node behavior features and node association features through the feature cross-attention mechanism, and generate a fused feature representation.

[0060] The node behavior features and node association features extracted from the power process data set are input into the feature fusion layer of the intelligent power process optimization model. In the feature fusion layer, a feature cross-attention mechanism is used to analyze the relationship between these two types of features. Specifically, this mechanism analyzes each element in the node behavior features and node association features, determining the degree of their interrelationship through a series of calculations, thereby obtaining association weights. For example, for the average execution time element in the node behavior features and the probability element of a certain trigger path in the node association features, their mutual influence in the entire power business process is analyzed, and the corresponding association weights are calculated. By performing the above analysis and calculation on all elements, a complete set of association weights is obtained. Then, based on these association weights, the node behavior features and node association features are fused to generate a fused feature representation, which integrates information from both node behavior and association.

[0061] Step S132: Perform time-series anomaly detection processing on the fused feature representation through the time-series optimization sub-model of the power process intelligent optimization model, and generate a node execution duration adjustment target.

[0062] In the scenario of a large-scale commercial power business platform, the generated fusion feature representation is input into the time-series optimization sub-model of the power process intelligent optimization model. This sub-model focuses on detecting whether there are time-series anomalies by analyzing information related to time series, and then generating node execution duration adjustment targets.

[0063] Step S1321: Input the fused feature representation into the temporal distribution modeling layer of the temporal optimization sub-model, and obtain the mean parameter and variance parameter by fitting the probability distribution of node execution duration through a Gaussian mixture model.

[0064] The fused feature representation from the feature fusion layer is input into the time-series distribution modeling layer of the time-series optimization sub-model. In this layer, a Gaussian mixture model is used to fit the probability distribution of node execution duration. Specifically, the Gaussian mixture model comprehensively considers various information related to node execution duration in the fused feature representation, such as the execution duration of different operating nodes in different time periods, and the distribution of these durations along the entire power business process timeline. Through learning and analysis of a large amount of similar data, the time-series optimization sub-model can simulate the probability distribution pattern of node execution duration, thereby deriving the mean and variance parameters. The mean parameter reflects the average level of node execution duration, while the variance parameter reflects the degree of dispersion of execution duration around the mean.

[0065] Step S1322: Calculate the absolute deviation between the execution time of each node and the mean parameter, compare the absolute deviation with a preset variance multiple threshold, and identify nodes whose absolute deviation exceeds the threshold as abnormal nodes.

[0066] For each node's execution time, its absolute deviation from the mean parameter is calculated. Specifically, for each node, the mean parameter is subtracted from its execution time, and the absolute value is taken. For example, if a node's execution time is one value and the mean parameter is another, the absolute value of the difference is calculated. This absolute deviation is then compared to a pre-set variance multiple threshold. This variance multiple threshold is determined based on the actual needs and experience of the power business process and is used to define the range between normal and abnormal. If the absolute deviation of a node's execution time from the mean parameter exceeds this variance multiple threshold, then the node is identified as an abnormal node, meaning its execution time deviates significantly from the normal pattern.

[0067] Step S1323: Perform historical data backtracking on the execution time of the abnormal node, and extract the execution time distribution of the abnormal node under different business scenarios. The business scenarios include peak time scenarios, off-peak time scenarios, and fault-prone time scenarios.

[0068] For identified abnormal nodes, historical data backtesting is performed on their execution duration. From the historical power flow data set, the distribution of execution duration for these abnormal nodes under different business scenarios is searched and extracted. These business scenarios mainly cover peak-hour scenarios, such as peak electricity consumption in commercial areas where a surge in commercial activity leads to a significant increase in electricity demand, potentially affecting the execution duration of power equipment. During off-peak hours, electricity demand is relatively low, and equipment execution durations may exhibit different characteristics. In scenarios with frequent faults, the execution duration of related operations will also change when power equipment experiences numerous faults. By analyzing the distribution of execution duration under these different scenarios, a more comprehensive understanding of the changing patterns of abnormal node execution durations can be obtained.

[0069] Step S1324: Based on the execution time distribution under different business scenarios, determine the reasonable execution time range of abnormal nodes under each scenario. The reasonable execution time range includes the mean and standard deviation corresponding to the business scenario.

[0070] Based on the execution duration distribution of extracted abnormal nodes across different business scenarios, a reasonable execution duration range for each scenario is determined. For each business scenario, a mean and standard deviation are calculated through analysis of the execution duration data. The mean represents the typical level of abnormal node execution duration in that business scenario, while the standard deviation reflects the degree of fluctuation in execution duration around the mean. The mean minus a predetermined multiple of the standard deviation is used as the lower limit, and the mean plus a predetermined multiple of the standard deviation is used as the upper limit (the multiple is determined based on the actual business situation), thereby determining the reasonable execution duration range.

[0071] Step S1325: Use the upper and lower limits of the reasonable execution time range as the node execution time adjustment target. The node execution time adjustment target is used to constrain the execution time of abnormal nodes in the corresponding business scenario.

[0072] The upper and lower limits of the reasonable execution time range determined under different business scenarios are set as the node execution time adjustment targets. These targets clarify the range within which the execution time of abnormal nodes should be constrained in the corresponding business scenarios. For example, in peak period scenarios, the execution time of abnormal nodes should be adjusted and controlled within the reasonable execution time range corresponding to that scenario, thereby optimizing the execution time of nodes in the power business process and ensuring the efficient operation of the entire process.

[0073] Step S133: Use the dependency optimization sub-model of the power process intelligent optimization model to perform dependency redundancy detection processing on the fused feature representation and generate dependency correction targets.

[0074] In the context of a large-scale commercial power business platform, the dependency optimization sub-model performs dependency redundancy detection processing on the fusion feature representation, aiming to identify and correct unnecessary dependencies between nodes and generate dependency correction targets.

[0075] Step S1331: Extract edge information of the trigger-dependent directed graph from the fused feature representation, wherein the edge information includes the starting point, ending point and trigger success probability of the edge.

[0076] The dependency optimization sub-model extracts edge information related to the directed graph of trigger dependencies from the fused feature representation. The fused feature representation contains rich information about the dependencies between nodes in the power business process, from which the dependency optimization sub-model can accurately extract the start and end identifiers of each edge, as well as the corresponding trigger success probability. For example, for an edge representing a dependency between two operation nodes, the dependency optimization sub-model can explicitly obtain the start node identifier, the end node identifier, and the trigger success probability corresponding to this edge. This trigger success probability reflects the likelihood of triggering from the start node to the end node.

[0077] Step S1332: Perform threshold filtering on the trigger success probability and retain edges with trigger success probabilities lower than a preset threshold as candidate redundant edges.

[0078] The sub-model performs threshold filtering on the extracted trigger success probabilities. A reasonable threshold is preset, and the trigger success probabilities of all edges are compared with this threshold. Edges with trigger success probabilities lower than the preset threshold are retained as candidate redundant edges. This is because edges with low trigger success probabilities have a relatively small impact on the final result in the entire business process and are more likely to be redundant, so they are considered as objects that may need to be deleted or corrected.

[0079] Step S1333: Perform process impact analysis on the candidate redundant edges, simulate the change in the execution path of the process instance after deleting the candidate redundant edges, and calculate the success rate of process execution after the path change.

[0080] For the selected candidate redundant edges, a dependency optimization sub-model performs a process impact analysis. Based on the process structure information and dependency logic between nodes contained in the fused feature representation, the sub-model simulates how the execution path of the process instance from the start node to the end node will change after these candidate redundant edges are removed. For example, there might have been a path from node A to node B via a candidate redundant edge. After removing this edge, the process might need to use other nodes and edges to advance from A to B. The dependency optimization sub-model analyzes this path change in detail. After simulating the path change, the dependency optimization sub-model calculates the process execution success rate after the path change. This requires comprehensive consideration of multiple factors, including the execution conditions of each node on the new path, dependencies, and the execution status of similar paths in historical data. For example, the new path might involve nodes that are only triggered under certain conditions. The dependency optimization sub-model assesses the likelihood of these conditions being met in actual business scenarios based on historical experience and relevant information in the fused feature representation, thereby estimating the process execution success rate.

[0081] Step S1334: Select candidate redundant edges whose success rate does not decrease significantly after deletion, and use them as the final redundant edges.

[0082] The sub-model filters out edges whose process execution success rate does not significantly decrease after deleting candidate redundant edges, and identifies them as final redundant edges. Here, "not significantly decreased" is determined by comparing the process execution success rate with the original process execution success rate. The dependency optimization sub-model comprehensively considers the actual needs and acceptable fluctuation range of the power business process. If the process execution success rate remains within an acceptable range after deleting a candidate redundant edge, it is considered to have a small impact on the process execution success rate and is identified as a final redundant edge. These edges are the objects that require subsequent dependency correction.

[0083] Step S1335: Perform dependency necessity verification on the final redundant edge, check whether the starting node of the redundant edge is the only predecessor node of the ending node, and if there are other predecessor nodes, mark it as a redundant edge that can be deleted.

[0084] The dependency necessity of the identified final redundant edges is verified. The sub-model checks whether the starting node of each final redundant edge is the only predecessor node of the ending node. For example, for a final redundant edge from node C to node D, the dependency optimization sub-model checks whether node D can be triggered by any other node besides node C. If there are other predecessor nodes that can trigger node D, then this redundant edge is marked as a removable redundant edge, meaning that this edge is not essential in the node dependency relationship and can be considered for deletion to optimize the process.

[0085] Step S1336: Use the start and end identifiers of the removable redundant edges as dependency correction targets. The dependency correction targets are used to guide the redundancy triggering dependencies in the deletion process instance.

[0086] The start and end points of edges marked as removable redundant edges are used as targets for dependency correction. These markers clearly identify the specific objects in the power business process instance that require dependency correction, i.e., the redundant triggering dependencies that need to be deleted. Based on these targets, the node dependencies in the process instance can be adjusted to optimize the entire power business process and improve its operational efficiency.

[0087] Step S134: Input the node execution time adjustment target and dependency correction target into the constraint generation module of the power process intelligent optimization model to determine the feasible range of the adjustment target and the execution conditions of the correction target.

[0088] In the context of a large-scale commercial power business platform, once the node execution duration adjustment target and dependency correction target are generated, they are input into the constraint generation module of the power process intelligent optimization model.

[0089] For adjusting the execution time of nodes, the constraint generation module first considers multiple factors to determine its feasible range. These factors include the overall time requirements of the power business process, the sequential relationship between different operation nodes, and the distribution of the execution time of each node in similar business scenarios in historical data. For example, in the entire power equipment inspection process, the execution time of certain nodes needs to be completed within a set time range to ensure that subsequent nodes can start on time and will not affect the completion of the entire inspection process within the specified time. At the same time, referring to the fluctuation range of the execution time of each node in many similar inspection processes in history, the upper and lower limits of the adjustment target for the execution time of each node are determined to form a feasible range.

[0090] For dependency correction objectives, the constraint generation module analyzes their execution conditions. This involves a deep understanding of the logic of the entire power business process and the node dependency information contained in the fusion feature representation. For example, if a redundant dependency edge (one of the dependency correction objectives) needs to be deleted, the constraint generation module will examine the role of the node involved in the edge in the entire process, as well as the chain reaction on the dependencies of other nodes after deleting the edge. It can determine under what circumstances it is safe to delete the edge, such as whether certain preceding nodes have completed specific operations, or whether certain conditions have been met, thereby clarifying the execution conditions of the dependency correction objective.

[0091] Step S135: Based on the feasible range and execution conditions, generate a set of optimized constraints that includes the upper limit threshold of node execution time, the lower limit threshold of node execution time, and the condition for deleting dependent edges.

[0092] In the scenario of a large-scale commercial power business platform, the feasible range and dependencies of the target are adjusted based on the node execution time determined by the constraint generation module, and the execution conditions of the target are corrected, thus starting to generate an optimized constraint set.

[0093] Regarding node execution time, the upper and lower limits corresponding to the adjustment target of each node's execution time are extracted from the feasible range and used as the upper and lower limits of the node's execution time, respectively. For example, for a certain operation node in the equipment inspection process, based on the previously determined feasible range, the upper and lower limits of the reasonable execution time range for this node under different business scenarios are obtained, and these are explicitly set as the upper and lower limits of the node's execution time in the corresponding scenarios. These thresholds will be used to limit the actual execution time of the node, ensuring that it remains within a reasonable range.

[0094] The conditions for deleting a dependency edge are determined based on the execution conditions of the dependency relationship modification target. For example, if deleting a redundant dependency edge requires certain conditions such as the preceding node completing a specific operation and the current business scenario being in a certain state, then these conditions are compiled into detailed dependency edge deletion conditions. These conditions clarify under what circumstances the corresponding redundant dependency edges can be deleted to optimize the dependencies between nodes.

[0095] The upper and lower limits of the execution time of all nodes, as well as the conditions for deleting all redundant dependent edges, are summarized to form a complete set of optimization constraints. This set of optimization constraints comprehensively covers the optimization constraints on the execution time of nodes and node dependencies in the power business process.

[0096] Step S140: Generate a power process optimization strategy based on the set of optimization constraints, which includes a node execution timing adjustment scheme and a node dependency relationship reconstruction scheme.

[0097] In the scenario of a large-scale commercial power business platform, based on the optimized set of constraints, a comprehensive power process optimization strategy is further generated. This strategy mainly includes a node execution timing adjustment scheme and a node dependency reconstruction scheme.

[0098] Step S141: parse the upper limit threshold and lower limit threshold of node execution time in the set of optimization constraints, and configure execution time constraint rules for each abnormal node. The execution time constraint rules include node identifier, business scenario identifier, upper limit value and lower limit value of execution time.

[0099] Extract the upper and lower threshold values ​​for node execution time from the set of optimization constraints. For each anomalous node identified in the previous timing anomaly detection, configure specific execution time constraint rules.

[0100] First, identify the node identifier for each abnormal node. This identifier uniquely identifies the node within the power business process, such as the identifier for a specific equipment detection node. Next, determine the business scenario identifier associated with the node, as different business scenarios may have different requirements for node execution time, such as peak and off-peak hours in commercial areas. Then, obtain the upper and lower limits of the node's execution time under the corresponding business scenario from the set of optimization constraints. For example, during peak electricity consumption hours, the upper and lower limits of an equipment maintenance node's execution time are specific time ranges. Integrate this information to form execution time constraint rules, ensuring that the execution time of each abnormal node is reasonably limited across different business scenarios.

[0101] Step S142: Parse the dependency edge deletion conditions in the set of optimization constraints, and configure dependency deletion rules for each deletable redundant edge. The dependency deletion rules include edge start point identifier, edge end point identifier, and deletion trigger condition.

[0102] The dependency edge deletion conditions in the optimization constraint set are analyzed. For each case where a redundant edge is determined to be deletable, detailed dependency deletion rules are configured for it.

[0103] The redundant edge is clearly identified by its start and end points. These identifiers effectively define the edge's position in the directed graph that triggers dependencies. For example, for a redundant edge from node A to node B, A is the start point identifier, and B is the end point identifier. Then, based on the previously defined deletion conditions, the triggering conditions for deleting the edge are determined. For instance, the edge can only be deleted when node A completes a specific operation and the current business process is at a certain stage. These detailed triggering conditions are incorporated into the dependency deletion rules. In this way, clear deletion rules are established for each removable redundant edge, enabling accurate handling of redundant dependencies in subsequent process optimizations.

[0104] Step S143: Perform conflict detection processing on the execution duration constraint rules and dependency deletion rules to check whether there are cases where the same node is constrained by multiple rules at the same time or the same dependency edge is deleted multiple times.

[0105] In the scenario of a large-scale commercial power business platform, after generating execution time constraint rules and dependency deletion rules, conflict detection needs to be performed on these two types of rules to ensure the consistency and feasibility of the optimization strategy.

[0106] Step S1431: Extract the set of node identifiers in the execution duration constraint rules, and count the number of times each node identifier in the set of node identifiers appears in the rule set.

[0107] Extract all node identifiers from the set of execution duration constraint rules to form a node identifier set. For each node identifier in this set, count its occurrences by traversing the entire set of execution duration constraint rules. For example, examine each execution duration constraint rule in turn; if a rule involves a certain node identifier, increment the count of that node identifier by one. This counts the frequency of each node identifier in the set of execution duration constraint rules.

[0108] Step S1432: Extract the set of edge identifiers from the dependency deletion rules. The edge identifiers are generated by combining the edge start identifier and the edge end identifier. Count the number of times each edge identifier appears in the rule set.

[0109] Edge identifiers are extracted from the dependency deletion rule set. Each edge identifier is composed of the start and end identifiers of each edge. For each dependency deletion rule, the start and end identifiers of the edges involved are combined to form an edge identifier set. Then, similar to counting the occurrences of node identifiers, the dependency deletion rule set is traversed to count the occurrences of each edge identifier within that set. For example, each dependency deletion rule is checked sequentially; if a rule involves a certain edge identifier, the occurrence count of that edge identifier is incremented by one, thus obtaining the frequency of each edge identifier in the dependency deletion rule set.

[0110] Step S1433: Create a conflict detection table, which includes a node conflict column and an edge conflict column.

[0111] Create a conflict detection table to record conflict information. This table has two main columns: a node conflict column and an edge conflict column. The node conflict column records conflict information related to nodes, and the edge conflict column records conflict information related to dependent edges.

[0112] Step S1434: Mark the node identifiers that appear more than 1 times in the node conflict column, and mark the edge identifiers that appear more than 1 times in the edge conflict column.

[0113] Nodes that appear more than once in the execution duration constraint rule set are marked in the node conflict column of the conflict detection table; edges that appear more than once in the dependency deletion rule set are marked in the edge conflict column. For example, if a node appears multiple times in the execution duration constraint rule set, it is recorded in the corresponding position in the node conflict column; similarly, if an edge appears multiple times in the dependency deletion rule set, it is recorded in the corresponding position in the edge conflict column. This method visually displays nodes and edges that may conflict.

[0114] Step S1435: Perform rule content comparison processing on the conflict nodes and conflict edges in the conflict detection table, and analyze whether there is overlap in the constraints of the conflict rules.

[0115] For conflict nodes and edges marked in the conflict detection table, further analysis of their corresponding rule content is conducted. For conflict nodes, multiple execution duration constraint rules involving that node are examined in detail, and their constraint conditions, such as business scenarios and duration ranges, are compared. For conflict edges, the deletion trigger conditions for that edge in multiple dependency deletion rules are examined. Through careful comparison, it is determined whether there is any overlap in the constraint conditions of these conflicting rules. For example, for a conflict node, if two execution duration constraint rules set the same business scenario but their duration ranges overlap or contradict each other, the constraint conditions are considered to overlap; for a conflict edge, if the deletion trigger conditions of two dependency deletion rules for that edge are simultaneously satisfied in certain situations, the constraint conditions are also considered to overlap.

[0116] Step S1436: If the constraints of the conflicting rules overlap, record it as a valid conflict; if the constraints of the conflicting rules do not overlap, record it as no actual conflict.

[0117] Record the results of comparing the constraints of conflicting rules. If overlapping constraints are found, record this as a valid conflict, meaning there is a real conflict between these rules, which requires further processing. If the constraints of conflicting rules do not overlap after analysis, meaning they will not contradict each other in practical application, then record this as no actual conflict, and no rule adjustment is required in this case.

[0118] Step S1437: Generate conflict detection results containing conflict type, conflict object, and conflict status.

[0119] Based on the above analysis and records, conflict detection results are generated. The conflict type is distinguished as either node-related or edge-related; the conflict object is clearly identified as the specific node or edge identifier involved; and the conflict status indicates whether it is a valid conflict or no actual conflict. In this way, a complete conflict detection result is formed, allowing for targeted processing of conflicting rules.

[0120] Step S144: Prioritize conflicting rules, retain rules with higher priority than the first set priority, and adjust the scope of application of rules with lower priority than the second set priority. The first set priority is higher than the second set priority.

[0121] After obtaining the conflict detection results, the conflicting rules are prioritized. First, the criteria for judging rule priority are determined, which may be based on factors such as the rule's impact on the criticality of the power business process and the importance of the business scenarios involved. For example, rules directly related to ensuring the stability of power supply may have a higher priority, while rules in some relatively less important business scenarios may have a lower priority.

[0122] According to the established evaluation criteria, conflicting rules are prioritized. Rules with a priority higher than the first set priority are retained; these rules play a crucial role in optimizing power business processes and must be ensured to be implemented in the optimization strategy. For rules with a priority lower than the second set priority, their scope of application is adjusted. For example, if a rule may conflict with other rules in multiple business scenarios, but it still has a certain effect on process optimization in a specific sub-scenario, then its scope of application is narrowed to that sub-scenario to avoid conflicts with other high-priority rules, while still allowing it to play its optimization role in the appropriate scenario. Through the above processing, the rule conflict problem is resolved, making the optimization strategy more reasonable and effective.

[0123] Step S145: The execution time constraint rules, dependency deletion rules, and conflict resolution adjustment rules are structurally integrated to generate the power process optimization strategy.

[0124] After handling rule conflicts, the execution time constraint rules, dependency deletion rules, and rules adjusted after conflict resolution are structurally integrated to generate a complete power process optimization strategy.

[0125] First, organize all execution duration constraint rules and arrange them logically according to node identifiers or business scenarios to form an ordered set. For example, they can be categorized by business scenarios first, and then arranged according to the set order of node identifiers within each business scenario. This facilitates quick retrieval and application in practical applications.

[0126] Similarly, dependency deletion rules are also organized in a similar way, arranged according to the edge start point identifier, edge end point identifier, or the logical order related to the business process, forming an ordered set of dependency deletion rules.

[0127] For rules adjusted after conflict resolution, they are incorporated into the corresponding execution duration constraint rule set and dependency deletion rule set, respectively. For example, if the scope of application of an execution duration constraint rule is adjusted, the adjusted rule is accurately placed in the corresponding position in the execution duration constraint rule set; if the triggering condition of a dependency deletion rule is modified, the modified rule is integrated into the dependency deletion rule set.

[0128] Finally, the compiled set of execution time constraint rules, dependency deletion rules, and adjusted rule sets are integrated to form a structured whole, which is the final generated power process optimization strategy. This optimization strategy comprehensively covers the optimization rules for node execution time and node dependencies, and after conflict resolution, it has consistency and feasibility, which can effectively guide the power business platform to optimize the power process.

[0129] Step S150: Deploy the power process optimization strategy to the process execution engine of the power business platform, collect process operation feedback data after the implementation of the power process optimization strategy, and iteratively update the inference parameters of the power process intelligent optimization model through the process operation feedback data.

[0130] In the scenario of a large-scale commercial area power business platform, once the power process optimization strategy is generated, it needs to be deployed to the process execution engine of the power business platform. By collecting feedback data on process operation after the optimization strategy is implemented, the inference parameters of the power process intelligent optimization model are iteratively updated to continuously optimize the power business process.

[0131] For example, step S151: convert the power process optimization strategy into a configuration file that the process execution engine can recognize.

[0132] Power process optimization strategies exist as a set of logical rules, while the process execution engine requires a formatted configuration file to guide its operation. Therefore, the generated power process optimization strategies need to be transformed.

[0133] The execution duration constraint rules, dependency deletion rules, and post-conflict resolution adjustment rules are analyzed and organized. For example, the node identifier, business scenario identifier, upper and lower duration limits in the execution duration constraint rules for each node are reorganized according to the format required by the process execution engine; for dependency deletion rules, the edge start point identifier, edge end point identifier, and deletion trigger condition are also converted accordingly. Through a series of format conversions and data organization operations, the entire power process optimization strategy is transformed into a configuration file that conforms to the process execution engine's recognition specifications, ensuring that the process execution engine can accurately understand and apply these optimization rules.

[0134] Step S152: Update the configuration of the process definition module of the process execution engine through the configuration file, and trigger the process execution engine to execute the power process instance according to the new configuration.

[0135] The converted configuration file is then input into the process execution engine of the power business platform, specifically for updating the configuration of the process definition module of the process execution engine.

[0136] The process definition module is responsible for defining the specific execution logic and sequence of power business processes. Information in the configuration file replaces or supplements the original settings regarding node execution duration and node dependencies in the process definition module. For example, if the execution duration of a node in the original process definition module was not strictly limited, now, based on the execution duration constraints in the configuration file, a clear upper and lower limit value is set for that node; if there were some redundancies in the dependencies between nodes, the dependency edges between nodes are adjusted according to the dependency deletion rules in the configuration file. After the configuration update is completed, the process execution engine receives the new configuration information and triggers it to execute the power process instance according to the new configuration, enabling the power business process to run according to the optimization strategy.

[0137] Step S153: When the process execution engine executes the newly configured power process instance, collect the real-time node execution logs during the process. The real-time node execution logs include the actual start time of the operation, the actual end time of the operation, the actual triggering conditions, and the actual operation result status.

[0138] When the process execution engine starts executing the power process instance according to the new configuration, it initiates a data acquisition mechanism to collect real-time node execution logs during the process.

[0139] At each operational node in the power business process, the actual start time of the operation is recorded; the actual end time of the operation is recorded; the actual triggering condition is recorded, which may be consistent with the previously set triggering condition or may change due to actual circumstances; and the actual operation result status is recorded, such as operation successful, a fault occurring, or partial functional abnormality. For example, in the process of inspecting power equipment in a commercial area, for the equipment inspection operation node, the actual start time of inspection, the completion time of inspection, the actual triggering condition of the inspection operation (such as a change in equipment operating parameters), and the inspection result status (equipment normal, minor fault found, etc.) are recorded. By comprehensively collecting these real-time node execution logs, the actual operating data of the power process after the implementation of optimization strategies can be obtained.

[0140] Step S154: Perform feature extraction processing on the real-time node execution log to obtain the node behavior features and node association features after the implementation of the power process optimization strategy.

[0141] After collecting real-time node execution logs, feature extraction processing is performed to obtain key features after the implementation of the power process optimization strategy. This is similar to the previous feature extraction of the original power process data set, and will not be elaborated here.

[0142] Step S155: Calculate the standardized difference between the node behavior characteristics and node association characteristics after the implementation of the power flow optimization strategy and the node behavior characteristics and node association characteristics before optimization. The standardized difference includes the difference in execution time distribution and the difference in dependency path probability.

[0143] After obtaining the node behavior characteristics and node association characteristics after the implementation of the power process optimization strategy, they are compared with the corresponding characteristics before optimization, and the standardization difference is calculated.

[0144] To address the differences in execution duration distribution, we first analyze the execution duration distribution in the node behavior characteristics before optimization, including the distribution of average execution duration, longest execution duration, and shortest execution duration across different time segments as reflected in the previously calculated basic time-series statistics. Then, we compare the execution duration distribution after the optimization strategy is implemented, again using the basic time-series statistics. For example, we compare the changes in average execution duration within different time segments before and after optimization, as well as the differences in the longest and shortest execution durations. To ensure comparability of these differences, we employ standardization methods based on relevant techniques to process this data. For instance, we transform the difference in average execution duration for corresponding time segments before and after optimization into a value within a defined range using calculations related to the overall data, thereby accurately representing the differences in execution duration distribution.

[0145] For the dependency path probability difference, we analyze the dependency path probability distribution in the node association features before optimization, that is, the trigger success probability distribution of each trigger path. Then, we compare it with the newly generated dependency path probability distribution after the optimization strategy is implemented. For example, we examine the changes in the trigger success probability of certain key trigger paths before and after optimization. We also use standardization methods to process these probability difference data. For example, we convert the probability difference of the same trigger path before and after optimization into a measurable standardized value through calculations related to the probabilities of all paths, thus accurately reflecting the dependency path probability difference. We combine the execution time distribution difference and the dependency path probability difference to form a complete standardized difference degree. This standardized difference degree can comprehensively reflect the differences in node behavior and node association after the implementation of the power process optimization strategy compared to before optimization.

[0146] Step S156: Input the standardized difference into the parameter update module of the power process intelligent optimization model, and adjust the network weights of the feature fusion layer, time series optimization sub-model and dependency optimization sub-model in the power process intelligent optimization model through the gradient descent algorithm to complete the iterative update process of the inference parameters.

[0147] The calculated standardized variance is input into the parameter update module of the intelligent optimization model for power processes. The main function of the parameter update module is to adjust relevant parts of the model based on the input variance information, so as to improve the accuracy and effectiveness of the model in optimizing power business processes.

[0148] In the intelligent optimization model for power processes, the feature fusion layer, the time-series optimization sub-model, and the dependency optimization sub-model play crucial roles in the optimization inference process. Their network weights determine how the model processes different features and assesses their importance. The gradient descent algorithm can be used to adjust these network weights.

[0149] The basic idea of ​​the gradient descent algorithm is to determine the direction and magnitude of network weight adjustments by calculating the difference between the model's predicted results and the actual results (reflected in the standardized difference of features before and after optimization). Specifically, for the feature fusion layer, the algorithm can analyze the deficiencies in the fusion effect of node behavioral features and node association features reflected by the standardized difference. Based on these analyses, it calculates the direction (whether to increase or decrease) and approximate magnitude of adjustment for each network weight. For example, if it is found that some features do not contribute as expected to the final result during the fusion process, the algorithm will adjust the network weights related to these features accordingly, allowing them to play a more appropriate role in the fusion process.

[0150] For the temporal optimization sub-model, the gradient descent algorithm analyzes the sub-model's performance in detecting and handling temporal anomalies based on the information in the standardized variability distribution of execution time. If the execution time distribution difference indicates that the sub-model's judgment of certain anomalies is inaccurate or the adjustment target is unreasonable, it calculates how the network weights of the temporal optimization sub-model need to be adjusted to improve its ability to detect and handle temporal anomalies.

[0151] Similarly, for the dependency optimization sub-model, the gradient descent algorithm evaluates the sub-model's performance in detecting and handling dependency redundancy based on the dependency path probability differences in the standardized variability. If the dependency path probability differences indicate that the sub-model misjudges certain redundant dependencies or that the correction target is inappropriate, then the adjustment method for the network weights of the dependency optimization sub-model is determined to enable the dependency optimization sub-model to more accurately detect and correct dependency redundancy.

[0152] By employing the gradient descent algorithm described above, the network weights of the feature fusion layer, the temporal optimization sub-model, and the dependency optimization sub-model are progressively adjusted, completing the iterative update process of the inference parameters for the intelligent power process optimization model. The updated model is better able to adapt to the actual conditions of power business processes, providing more effective optimization suggestions and strategies in subsequent process optimization.

[0153] Figure 2The illustration shows exemplary hardware and software components of an AI-based business middleware process optimization system 100 that can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the AI-based business middleware process optimization system 100 and to perform the functions in this application.

[0154] The AI-based business middleware process optimization system 100 can be a general-purpose server or a special-purpose server; both can be used to implement the AI-based business middleware process optimization method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0155] For example, the AI-based business process optimization system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the AI-based business process optimization system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The AI-based business process optimization system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0156] For ease of explanation, only one processor is described in the AI-based business middleware process optimization system 100. However, it should be noted that the AI-based business middleware process optimization system 100 of this application may also include multiple processors. Therefore, the steps executed by one processor as described in this application may also be executed jointly or individually by multiple processors. For example, if the processor of the AI-based business middleware process optimization system 100 executes steps A and B, it should be understood that steps A and B may also be executed jointly by two different processors or individually by one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0157] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned business middle platform process optimization method based on artificial intelligence is implemented.

[0158] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A business middle platform process optimization method based on artificial intelligence, characterized in that, The method includes: Acquire a set of power process data generated during the operation of the power business middleware. The set of power process data contains multiple process instances. Each process instance consists of continuously executed business operation nodes and corresponding node execution logs. The node execution logs include the operation start time, operation end time, operation triggering conditions, and operation result status. Process features are extracted from the power process data set to obtain the node behavior features and node association features of each process instance. The node behavior features include the execution time distribution pattern of operation nodes, and the node association features include the triggering dependency relationship between operation nodes. The pre-built intelligent optimization model for power processes is invoked to perform collaborative optimization reasoning on the node behavior characteristics and node association characteristics, generating a set of optimization constraints for the process instance; Based on the set of optimization constraints, a power process optimization strategy is generated that includes a node execution timing adjustment scheme and a node dependency reconstruction scheme. The power process optimization strategy is deployed to the process execution engine of the power business platform, the process operation feedback data after the implementation of the power process optimization strategy is collected, and the inference parameters of the power process intelligent optimization model are iteratively updated through the process operation feedback data.

2. The business middle platform process optimization method based on artificial intelligence according to claim 1, characterized in that, The process feature extraction of the power process data set to obtain the node behavior features and node association features of each process instance includes: The process instances in the power process data set are classified into equipment inspection type, fault handling type and electricity bill settlement type. Perform time series alignment processing on the node execution logs of each type of process instance, mapping the start time and end time of the operation to a unified time axis, and generating node timeline data containing time axis coordinates; The node timeline data is subjected to sliding window analysis processing. Continuous node time segments are extracted with a preset time window length. The average execution time, longest execution time, and shortest execution time of the node within each time segment are calculated to generate basic time-series statistics of node behavior characteristics. The operation triggering conditions in the node execution log are semantically parsed to extract the set of preceding node identifiers from the triggering conditions and construct a trigger dependency directed graph between nodes. The trigger dependency directed graph contains node identifiers and trigger edges. The path traversal process is performed on the trigger dependency directed graph, all possible trigger paths from the start node to the end node are counted, the trigger success probability of each trigger path is calculated, and the dependency path probability distribution of node association features is generated. The basic time-series statistics and dependency path probability distributions are standardized and integrated to generate corresponding node behavior features and node association features.

3. The business middle platform process optimization method based on artificial intelligence according to claim 2, characterized in that, The process of performing sliding window analysis on the node timeline data involves extracting continuous node time segments with a preset time window length, calculating the average execution time, longest execution time, and shortest execution time of nodes within each time segment, and generating basic time-series statistics of node behavioral characteristics, including: Extract the timestamp sequence of the operation start time and the timestamp sequence of the operation end time from the node timeline data; Calculate the execution time of each operation node, where the execution time is the difference between the timestamp of the operation end time and the timestamp of the operation start time; The execution duration sequence is arranged in chronological order to form a continuous execution duration sequence; A sliding window is used to slide over the long time sequence during execution. The sliding window slides one node at a time, and the coverage length of the sliding window is the preset time window length. Statistical calculations are performed on the execution time data within each sliding window to obtain the average, maximum, and minimum execution times within the sliding window; The average, maximum, and minimum values ​​are associated with the starting node identifier corresponding to the sliding window and stored to generate basic time-series statistics entries containing node identifier, time window position, average execution time, longest execution time, and shortest execution time; The basic time series statistics entries are arranged according to the execution order of the process instances to generate a basic time series statistics sequence with a time order relationship.

4. The business middle platform process optimization method based on artificial intelligence according to claim 2, characterized in that, The semantic parsing of the operation triggering conditions in the node execution log, extracting the set of preceding node identifiers from the triggering conditions, and constructing a directed graph of triggering dependencies between nodes includes: Extract the operation trigger condition field from the node execution log, the operation trigger condition field containing trigger rules described in natural language; Natural language processing is performed on the triggering rules to identify logical keywords and node identifiers. Based on the logical keywords and node identifiers, determine the prerequisite nodes that must be completed in the triggering conditions, and form a set of prerequisite node identifiers; Create a graph node for each operation node, and the identifier of the graph node is consistent with the identifier of the operation node; Add a directed edge between the previous node in each set of previous node identifiers and the current operation node. The starting point of the directed edge is the previous node identifier, and the ending point is the current operation node identifier. The triggering conditions of all operation nodes are processed as described above to generate a trigger dependency directed graph containing graph nodes and directed edges. The trigger dependency directed graph is used to represent the triggering order constraint relationship between operation nodes.

5. The business middle platform process optimization method based on artificial intelligence according to claim 1, characterized in that, The method of invoking a pre-built intelligent optimization model for power processes performs collaborative optimization reasoning on the node behavior characteristics and node association characteristics to generate a set of optimization constraints for the process instance, including: The node behavior features and node association features are input into the feature fusion layer of the power process intelligent optimization model. The association weight between the node behavior features and node association features is calculated through the feature cross-attention mechanism to generate a fused feature representation. The time-series optimization sub-model of the power process intelligent optimization model is used to perform time-series anomaly detection processing on the fused feature representation to generate a node execution duration adjustment target; The dependency optimization sub-model of the power process intelligent optimization model is used to perform dependency redundancy detection processing on the fused feature representation to generate dependency correction targets; The node execution time adjustment target and dependency correction target are input into the constraint generation module of the power process intelligent optimization model to determine the feasible range of the adjustment target and the execution conditions of the correction target. Based on the feasible range and execution conditions, an optimized constraint set is generated, which includes the upper limit threshold of node execution time, the lower limit threshold of node execution time, and the condition for deleting dependent edges.

6. The business middle platform process optimization method based on artificial intelligence according to claim 5, characterized in that, The step of performing time-series anomaly detection processing on the fused feature representation through the time-series optimization sub-model of the power process intelligent optimization model, and generating a node execution duration adjustment target, includes: The fused feature representation is input into the temporal distribution modeling layer of the temporal optimization sub-model, and the probability distribution of node execution duration is fitted by a Gaussian mixture model to obtain the mean parameter and variance parameter. Calculate the absolute deviation of the execution time of each node from the mean parameter, compare the absolute deviation with a preset variance multiple threshold, and identify nodes whose absolute deviation exceeds the threshold as abnormal nodes. The execution time of the abnormal nodes is backtracked using historical data, and the distribution of the execution time of the abnormal nodes under different business scenarios is extracted. The business scenarios include peak time scenarios, off-peak time scenarios, and fault-prone time scenarios. Based on the execution time distribution under different business scenarios, the reasonable execution time range of abnormal nodes under each scenario is determined. The reasonable execution time range includes the mean and standard deviation corresponding to the business scenario. The upper and lower limits of the reasonable execution time range are used as the node execution time adjustment targets, which are used to constrain the execution time of abnormal nodes in the corresponding business scenarios.

7. The business middle platform process optimization method based on artificial intelligence according to claim 5, characterized in that, The process of using the dependency optimization sub-model of the intelligent power process optimization model to perform dependency redundancy detection on the fused feature representation and generate dependency correction targets includes: The edge information of the trigger-dependent directed graph is extracted from the fused feature representation, and the edge information includes the start point, end point and trigger success probability of the edge; The trigger success probability is subjected to threshold filtering, and edges with trigger success probabilities lower than a preset threshold are retained as candidate redundant edges. The candidate redundant edges are subjected to process impact analysis, the execution path changes of the process instance are simulated after the candidate redundant edges are deleted, and the execution success rate of the process after the path changes is calculated. Select candidate redundant edges whose success rate does not decrease significantly after deletion, and use them as the final redundant edges. The final redundant edge is subjected to dependency necessity verification. The starting node of the redundant edge is checked to see if it is the only predecessor node of the ending node. If there are other predecessor nodes, it is marked as a redundant edge that can be deleted. The start and end identifiers of the removable redundant edges are used as dependency correction targets, which are used to guide the redundancy triggering dependencies in the deletion process instance.

8. The business middle platform process optimization method based on artificial intelligence according to claim 1, characterized in that, The power process optimization strategy, which generates a set of optimization constraints and includes node execution timing adjustment schemes and node dependency reconstruction schemes, includes: The upper and lower threshold values ​​of node execution time in the set of optimization constraints are analyzed, and execution time constraint rules are configured for each abnormal node. The execution time constraint rules include node identifier, business scenario identifier, upper limit value and lower limit value of execution time. The dependency edge deletion conditions in the set of optimization constraints are parsed, and a dependency deletion rule is configured for each deletable redundant edge. The dependency deletion rule includes the edge start point identifier, the edge end point identifier, and the deletion trigger condition. The execution time constraint rules and dependency deletion rules are subjected to conflict detection processing to check whether there are cases where the same node is constrained by multiple rules at the same time or the same dependency edge is deleted multiple times. For conflicting rules, priority sorting is performed, retaining the rules with higher priority than the first set priority and adjusting the scope of application of the rules with lower priority than the second set priority, where the first set priority is higher than the second set priority; The execution time constraint rules, dependency deletion rules, and conflict resolution adjustment rules are structurally integrated to generate the power process optimization strategy.

9. The business middle platform process optimization method based on artificial intelligence according to claim 8, characterized in that, The conflict detection process for the execution time constraint rules and dependency deletion rules, checking for situations where the same node is constrained by multiple rules simultaneously or the same dependency edge is deleted multiple times, includes: Extract the set of node identifiers from the execution time constraint rules, and count the number of times each node identifier in the set of node identifiers appears in the rule set; Extract the set of edge identifiers from the dependency deletion rules. The edge identifiers are generated by combining the edge start identifier and the edge end identifier. Count the number of times each edge identifier appears in the rule set. Create a conflict detection table, which includes a node conflict column and an edge conflict column; In the node conflict column, mark the node that appears more than once; in the edge conflict column, mark the edge that appears more than once. The conflict nodes and conflict edges in the conflict detection table are compared according to the rule content to analyze whether there is overlap in the constraints of the conflict rules. If the constraints of conflicting rules overlap, they are recorded as valid conflicts; if the constraints of conflicting rules do not overlap, they are recorded as no actual conflicts. Generate conflict detection results that include conflict type, conflict object, and conflict status.

10. A business middle platform process optimization system based on artificial intelligence, characterized in that, The system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the AI-based business middleware process optimization method as described in any one of claims 1-9.

Citation Information

Cited By

  • Business process setting method based on process engine

    CN121116558A

  • A method for setting a business process based on a process engine

    CN121116558B

  • Signal acquisition method and device, storage medium and electronic equipment

    CN121509455A

  • Electric power data auditing method and system

    CN122047970A