A virtual power plant operation optimization system

Abstract

The application discloses a virtual power plant operation optimization system, and relates to the technical field of power plant operation optimization.The system comprises a virtual construction module, a power plant operation monitoring module and an optimization scheduling module.The virtual construction module digitally models the distributed energy resources connected thereto, covering the basic characteristics, operation constraints and scheduling rules of the distributed energy resources; the power plant operation monitoring module realizes dynamic monitoring and evaluation of the operation state through real-time data acquisition and analysis, and completes the tracing analysis and positioning of abnormal working conditions based on model deviation and data comparison; and the optimization scheduling module generates a scheduling strategy by using an adaptive optimization algorithm in combination with multidimensional scheduling targets, thereby improving the operation efficiency and system stability.The system has good real-time performance, intelligence and adaptability, is suitable for intelligent energy management application scenarios with wide access of distributed energy resources, and can effectively improve the operation management level and comprehensive scheduling capability of the virtual power plant.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power plant operation optimization, in particular to a virtual power plant operation optimization system. BACKGROUND

[0002] The virtual power plant operation optimization system is mainly based on a centralized energy management architecture, relies on fixed scheduling strategies and static mathematical models to uniformly coordinate and optimally control the distributed energy resources accessed. Usually, wind power, photovoltaic, energy storage, electric vehicle charging piles and controllable loads and other types of energy units are integrated, power prediction, load adjustment and energy scheduling are performed through an energy management system (EMS) or a distributed energy management platform (DEMS), and the optimization of overall power output and demand side response is realized.

[0003] However, the existing virtual power plant operation system lacks static scheduling strategies and cannot respond to the dramatic fluctuations of wind and light distributed energy output in real time; when the system faces short-period load changes or renewable energy uncertainty, the scheduling response lags, which further leads to power deviation, power grid impact and low resource utilization efficiency and other problems. SUMMARY

[0004] In view of the problems mentioned in the above background art, a virtual power plant operation optimization system is proposed.

[0005] The purpose of the present application can be achieved by the following technical scheme: the present application provides a virtual power plant operation optimization system, which comprises a power plant operation monitoring module and an optimization scheduling module; the power plant operation monitoring module obtains key parameter data in actual power plant operation, maps the actual power plant data to the input end of the virtual power plant model in a one-to-one correspondence, establishes a parameter synchronization channel; drives the virtual power plant model to run through the input data; compares and identifies the deviation between the virtual operation result and the actual operation data in real time; when the virtual operation result exceeds the corresponding set threshold, automatically triggers abnormal backtracking analysis, and outputs the path credibility; compare the corresponding credibility of each causal path, and output the most credible abnormal causal path;

[0006] The optimization scheduling module obtains an output most credible abnormal causal path for optimization processing, including process optimization and software operation optimization; wherein the process optimization includes: constructing a directed graph model containing task nodes and dependent edges to generate a legal schedulable topology structure; constructing a weight model to identify a bottleneck task; introducing a multi-layer perception machine classification model for the bottleneck task to output process path defects; and then matching a corresponding optimization strategy according to different process path defect types; the software operation optimization includes: calculating the proportion of deviation from the normal load of the usage of CPU, memory and I / O resources to obtain CPU deviation rate, memory deviation rate and I / O deviation rate; calculating a task bottleneck contribution degree index according to the influence degree of the usage on the overall delay of the current task, selecting a bottleneck task set, entering the optimization strategy matching to obtain a task feature vector, inputting the task feature vector into a strategy matching unit to match to obtain an optimal strategy; calculating a comprehensive optimization expected revenue value of the optimization strategy on the task; if the expected revenue value is greater than a preset revenue threshold, the strategy is considered effective; if not, a backup strategy reselection mechanism is triggered to re-enter the strategy matching.

[0007] Further, it further includes a virtual construction module: the virtual construction module is used for constructing a power plant virtual scene, traversing and matching all devices in a standard device model library to obtain the numbers of various devices, and generating a device parameter library; if a corresponding device shape size and operation parameter cannot be matched, custom modeling is performed; the custom modeling: according to actual drawings or point cloud data, spatial geometric parameters and device logical parameters are analyzed; input information is standardized into a platform geometric semantic format; then, according to the standardized structure diagram and point cloud data, a spatial coordinate system is established, and each module component is arranged in a virtual space according to actual corresponding coordinates and sizes; for a special device structure in a power plant, high-precision modeling is performed by using a modeling tool to simulate device shapes, support structures, interface connections and running parts, and then a one-to-one space restoration is realized by combining a material library and a texture map; and then, by accessing a real-time data source, the model is driven to change in real time.

[0008] Further, the process of outputting the path credibility is as follows:

[0009] An abnormal parameter is obtained, a causal dependency graph is called, all influence node paths are obtained by a graph traversal algorithm, nodes on each path are calculated according to the influence node paths, the correlation degree between the path nodes and the abnormal nodes is calculated by mutual information, and then the calculated correlation degree is compared with a correlation threshold value, nodes with a correlation degree greater than the correlation threshold value are collected as possible cause nodes, and statistics are formed to form a possible cause set;

[0010] Next, we verify whether the potential triggering nodes change abruptly before the anomaly points in time. Specifically, we set triggering parameters to determine the mutation point and rate of change. If the conditions are met, the parameters are determined to be valid suspicious factors with consistent time causality. We then match the set of suspicious factors that have been screened and verified with the patterns in the fault rule base. If a rule exists, we mark the path as a rule-consistent path.

[0011] When several suspicious factors and causal paths are output, the credibility of each causal path is calculated, and the path credibility is output by combining relevance, temporal consistency and rule support.

[0012] Furthermore, output the most plausible causal path of the anomaly, including:

[0013] The correlation calculation, time consistency calculation, and rule matching degree calculation are used as three-dimensional features. The Softmax regression model is used for classification, and the probability of each path belonging to the true causal relationship is output as the confidence level. The confidence levels of each causal path are then compared, and the causal path with the highest confidence level is the most reliable abnormal causal path.

[0014] Furthermore, the specific process of constructing a weighted model to identify bottleneck tasks is as follows:

[0015] A directed graph model of process dependencies is constructed, and then a directed acyclic graph is used to identify circular dependencies or illegal paths in the process, generating a legal schedulable process topology. Based on the generated legal schedulable process topology, abnormal process paths are quantitatively analyzed to obtain the key timing deviations that affect efficiency and stability. If the key timing deviation is greater than zero, it indicates that the task is slower than expected; if the key timing deviation is less than zero, it indicates that the execution is ahead of schedule.

[0016] Next, an anomaly propagation path graph is constructed, and the anomaly sensitivity of the task is calculated. Based on the obtained offset value and anomaly sensitivity, combined with the execution complexity and historical risk score, a task priority model is constructed to obtain the bottleneck task identification factor. The bottleneck task identification factor is compared with a preset threshold. If the bottleneck task identification coefficient is greater than the preset threshold, it is a bottleneck task. The bottleneck tasks obtained from the task weight model are sorted to obtain a set.

[0017] For identified bottleneck tasks, a process refactoring strategy is designed; after adjustments are made through the coordinated execution of the strategy, a real-time feedback mechanism is introduced: continuously monitoring the task execution time and newly emerging offsets; if significant bottlenecks still exist after scheduling adjustments, the next round of software operation optimization is triggered.

[0018] Furthermore, the bottleneck contribution index for the task is calculated as follows:

[0019] Through a multi-stage evaluation process, performance bottlenecks in the current running task are identified from three dimensions, including:

[0020] For critical resources such as CPU, memory, and I / O, calculate the instantaneous offset rate and short-term variance index, and then classify the usage of each resource into normal, fluctuating, and abnormal states based on the instantaneous offset rate and short-term variance index; finally, assign normal, fluctuating, and abnormal labels to each resource based on the output results.

[0021] Utilize a sliding window to analyze the changing trends of task resources, detect continuous changing trends of key resources and the ratio of trend persistence; output status classifications based on judgment rules, including upward trend, fluctuating trend, and stable trend; assign labels of stable trend, fluctuating trend, and upward trend to each task;

[0022] Construct an execution dependency graph between tasks, analyze the role of the current task in the critical path, and label the structure as primary bottleneck, secondary bottleneck, and non-critical path.

[0023] Based on resource status, task allocation, and structural results, the task bottleneck contribution index is obtained by combining rules.

[0024] Furthermore, the process of obtaining the optimal strategy is as follows:

[0025] For each task in the bottleneck task set, multidimensional features of each task are extracted and integrated into a task feature vector through a feature encoding module, serving as the task profile for the corresponding task. The task feature vector is then input into the policy matching unit for matching. A trained policy is used to select an agent, dynamically choosing the optimal policy based on historical optimization effects and task type. If multiple optimal policies exist, candidate policies are selected. After the optimization policy matching and deployment are completed, for each candidate policy, its comprehensive expected optimization benefit for the task is calculated, and the policy with the highest expected benefit is selected as the optimal optimization policy.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] 1. This invention breaks through the limitations of traditional virtual power plants that rely on static scheduling models, have delayed response, and have fixed strategies by constructing a multi-source real-time perception and dynamic optimization model fusion mechanism. It realizes continuous perception and dynamic modeling of the status of resources such as wind power, photovoltaics, and energy storage. It can automatically adjust optimization strategies based on real-time data during operation, significantly improve the timeliness and accuracy of scheduling response, effectively cope with the fluctuation of new energy output, and improve the overall operating efficiency and stability of the system.

[0028] 2. This invention introduces a multi-dimensional causal analysis and process scheduling optimization mechanism. By utilizing causal dependency graphs and resource load analysis, it accurately identifies the performance bottlenecks of the path and dynamically optimizes the task process and scheduling strategy. At the same time, it combines feedback benefit indicators during operation to continuously adjust the control logic, realize personalized optimization of scheduling strategy and self-upgrading of system operation. This significantly improves the stability, adaptability and overall operating efficiency of power plant systems when facing complex operating environments. Attached Figure Description

[0029] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0030] Figure 1 This is a schematic diagram of the module connection of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0032] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0033] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0034] Please see Figure 1 As shown, a virtual power plant operation optimization system includes: a virtual construction module, a power plant operation monitoring module, and an optimization scheduling module.

[0035] The virtual building module is used to build virtual power plant scenarios. The specific building steps are as follows:

[0036] Step 1: Collect power plant parameters on-site and establish information files, including: collecting power plant type (including thermal power, hydropower, wind power, etc.), land area, installed capacity, main unit model, number of units and substation capacity, as well as obtaining original data such as building structure drawings, equipment layout drawings, pipeline diagrams, and process flow diagrams.

[0037] Based on the original data, the power plant is divided into functional zones, including the main plant area, boiler room, turbine room, power distribution room, cooling system, fuel processing area, office area, and storage area. The purpose, equipment layout, and access flow of each functional zone are determined and marked.

[0038] Step Two: Establish a standard equipment model library (including but not limited to transformers, steam turbines, and cooling towers) based on the equipment layout diagram. Perform specific model matching to obtain the corresponding equipment dimensions and operating parameters. Then, number the obtained dimensions and operating parameters of the corresponding equipment and iterate through all equipment in the standard equipment model library to obtain the number of each equipment. Count all equipment numbers to generate an equipment parameter library. If no matching equipment dimensions and operating parameters can be found, indicating no readily available component, custom modeling is performed. Custom modeling includes: analyzing spatial geometric parameters (dimensions, coordinates, and relative positions, etc.) from actual drawings / point cloud data and matching them with the equipment logic. The system compiles parameters (such as connection order or control relationships); standardizes the input information into the platform's geometric semantic format; then establishes a spatial coordinate system based on the standardized structural diagram and point cloud data, arranging each module component (such as boiler, steam turbine, transformer, cooling system, etc.) in virtual space according to its actual relative coordinates and dimensions; high-precision modeling is performed on the unique equipment structures in power plants using custom modeling tools; the system simulates the equipment's shape, support structure, interface connections, and operating parts; a one-to-one spatial restoration is achieved based on the material library and texture mapping; and the model is driven to change in real time (including valve status, speed, and temperature) by accessing real-time data sources.

[0039] Step 3: After completing the spatial construction, alignment, and functional logic configuration, the next step is the 3D visualization modeling stage, which enhances the realism and interactive experience of the virtual power plant. Based on the texture information from photos and videos collected on-site, material texture mapping is performed on the surfaces of buildings and equipment. After the material texture mapping is completed, a realistic 3D model is generated through the rendering engine. After the material and visual information reproduction is completed, the 3D model has achieved spatial restoration and sensory realism, and can be further integrated with the interactive logic and navigation system to realize the operation and control of the virtual scene.

[0040] The power plant operation monitoring module acquires key parameter data from actual power plant operation, such as equipment status, energy consumption indicators, and environmental variables; maps the actual power plant data to the input end of the virtual power plant model in a one-to-one correspondence, establishing a parameter synchronization channel; drives the virtual power plant model to operate through the input data, generating a virtual operating status synchronized with reality; then compares the virtual operating results with the actual operating data in real time to identify deviations, anomalies, or response differences; when the virtual operating results exceed the corresponding set threshold, anomaly tracing analysis is automatically triggered.

[0041] In a specific example: Suppose that the value of a certain parameter of the virtual power plant at time t is V. v (t), the actual value corresponding to the power plant is V. r If (t), then the deviation is: ΔV(t) = |V v (t)-V r (t)∣, when ΔV(t)> If identified as an abnormal trigger point, an abnormal tracing analysis will be automatically triggered. To set a threshold.

[0042] The specific process of the anomaly tracing analysis is as follows:

[0043] Obtain the abnormal parameters, retrieve them from the causal dependency graph G=(N,E), and use a graph traversal algorithm (such as breadth-first search or depth-first search) to obtain all paths P={n1, n2, ..., n} that affect the nodes. k}, where each node n i These are parameters or control points that have a direct or indirect dependency on the abnormal parameters; based on the path P affecting the nodes, for each node n on that path... i Calculate the correlation R(n) between path nodes and abnormal nodes. i ), through mutual information R(n i )=corr(V(n i The correlation coefficient R(n) is calculated from ΔV; then the correlation coefficient R(n) is calculated from ΔV. i The correlation coefficient R(n) is compared with the relevant threshold τ to determine the correlation degree. i Nodes with values ​​greater than the relevant threshold τ are collected as potential triggering nodes, and a set of potential triggering nodes is formed.

[0044] Next, we verify whether the potential triggering node undergoes a mutation before the outlier in time. Specifically, we set the triggering parameter as x(t) and its mutation point as t. x If t x <t 异常 And the rate of change satisfies: If t is considered a valid and suspicious factor with consistent temporal causality, then the parameter is considered to be t. 异常The time of anomaly occurrence is given by δ, which is a set threshold. Then, the set of suspicious factors C = {c1, c2, ..., c...} after screening and verification is defined. m} Match it with patterns in the fault rule base, R j : {c1 → abnormal parameter}; if rule R exists j If the path is consistent with the rules, it means that the suspicious factor c1 is supported by a known failure mechanism, and the path is more credible.

[0045] When several suspicious factors and causal paths are output, the confidence level S is calculated for each causal path. This is then combined with relevance, temporal consistency, and rule support: R(n) i ), T(n i ) and M(n i Using R(n) as a 3D feature, a Softmax regression model is used for classification, outputting the probability that each path belongs to the "true causal relationship", which is the confidence level S; where R(n) i ) represents the correlation calculation value, T(n) i ) represents the time consistency calculation value, M(n) i The value of the rule matching degree is calculated; then the confidence degree S corresponding to each causal path is compared, and the most reliable abnormal causal path is output, which is the causal path with the largest confidence degree S.

[0046] The optimization scheduling module optimizes the most reliable causal path of anomalies, including process optimization and software operation optimization; specifically:

[0047] The process optimization process is as follows: A directed graph model of process dependencies F=(T, D) is constructed, where: T={T1, T2, ..., T...} p} represents the independent and dependent task operation units in the process; D={D hg} represents the sequential constraints and resource dependencies between tasks (edges can have constraints, such as mutual exclusion, shared resource occupancy, etc.); and each edge D hg Includes: Type: sequential, parallel, conditional jump; Resources: corresponding resource occupancy type; Locking delay: the expected delay time caused by lock contention; The dependent relationships between process tasks are modeled using a directed acyclic graph (DAG) structure, and a loop detection algorithm is used to identify and eliminate circular dependencies and illegal paths in the process. A topology sorting algorithm is used to generate a legal schedulable process topology.

[0048] Based on the generation of a valid schedulable process topology, a quantitative analysis is conducted on abnormal process paths to obtain the key timing deviations affecting efficiency and stability; wherein: the offset calculation process for each task is as follows: Output task T h Time offset value ;in, For task T h Actual execution completion time, For task T h The optimal completion time under ideal scheduling conditions (which can also be understood as the expected baseline time); if Δt h If Δt > 0, it means the task is slower than expected; if Δt h If the value is less than 0, it indicates that the execution is ahead of schedule.

[0049] Construct an anomaly propagation path diagram and derive the extent of anomaly impact propagation using formulas: Obtain the abnormal sensitivity of the task. Where succ(h) is the task T h All direct and indirect downstream tasks (i.e., tasks that depend on it); For each subsequent task T h Time offset; For from T h To T g In a dependency graph, path depth represents the distance traveled.

[0050] Based on offset value and abnormal sensitivity Furthermore, by combining execution complexity and historical risk scores, a task prioritization model is constructed, including:

[0051] The execution time (duration_h), dependency complexity (degree_h), risk score (risk_h, i.e., historical anomaly occurrence rate × current offset), and anomaly sensitivity for each task are considered. Interval standardization was performed separately to obtain normalized indices Dh0, Gh0, Rh0, and Sh0.

[0052] The bottleneck task identification factor is constructed by taking the maximum value of each normalized index: Wh0=max{Dh0, Gh0, Rh0, Sh0}, where Dh0 represents the normalized value of task execution time, Gh0 represents the normalized value of dependency complexity, Rh0 represents the normalized value of historical risk score, Sh0 represents the normalized value of anomaly sensitivity; and Wh0 represents the bottleneck task identification factor.

[0053] The bottleneck task identification factor is compared with a preset threshold. If the bottleneck task identification factor is greater than the preset threshold, it is a bottleneck task. The bottleneck tasks obtained from the task weight model are sorted to obtain a set {T}. b} serves as the core object for scheduling optimization.

[0054] For identified bottleneck tasks, a multilayer perceptron classification model is introduced as a defect identification model to output process path defects. Process path defects include: latency defects, conflict defects, blocking defects, fault-tolerant defects, and serial inefficiency defects. Then, corresponding optimization strategies are matched according to different types of process path defects: For latency tasks: task splitting or preloading resources to reduce initialization and computation time; For conflict tasks: reducing resource congestion by adjusting task startup timing or remapping resources; For blocking tasks: optimizing dependency path structure, introducing intermediate caches or predictive execution of pre-tasks; For fault-tolerant tasks: strengthening input verification mechanisms and fault rollback strategies; For serial inefficiency tasks: adjusting scheduling logic, enabling concurrent execution mechanisms or automatic flowchart reordering.

[0055] During the optimization strategy execution process, a real-time feedback mechanism is introduced after adjustments are made: continuously monitor the task execution time and newly occurring offsets; if bottlenecks still exist after scheduling adjustments, trigger the next round of software operation optimization.

[0056] Software operation optimization: Based on the usage of resources such as CPU, memory, and I / O, calculate the proportion of deviation from normal load to obtain CPU offset rate, memory offset rate, and I / O wait offset; calculate the task bottleneck contribution index B based on the impact of usage on the overall latency of the current task. h It is divided into three levels: high, medium and low.

[0057] A preferred specific example involves a multi-stage evaluation process that identifies performance bottlenecks in the currently running task from three dimensions: metric offset, trend evolution, and path dependency.

[0058] For key resources such as CPU, memory, and I / O, two indicators are calculated: instantaneous offset rate and short-term variance index. Based on these two indicators, the usage of each resource is divided into normal, fluctuating, and abnormal states. The state determination process is as follows: when the instantaneous offset rate is less than or equal to 1.2 times the original offset rate and the variance index is less than or equal to 1.5 times the original variance index, it is considered normal; when the instantaneous offset rate is between 1.2 and 1.5 times the original instantaneous offset rate and the variance index is between 1.5 and 3 times the original variance index, it is considered fluctuating; when the instantaneous offset rate is greater than 1.5 times the original instantaneous offset rate and the variance index is greater than 3 times the original variance index, it is considered abnormal. Based on the output results, each resource (such as CPU, memory, etc.) is assigned a label of "normal / fluctuating / abnormal".

[0059] Then, the sliding window is used to analyze the trend of task resource changes, detect the continuous change trend of key resources (whether the resource usage has been rising, falling or fluctuating repeatedly over a period of time) and the trend persistence ratio (the proportion of time a single trend (such as rising) lasts within the observation window); output the status classification according to the judgment rules, including rising trend, fluctuating trend and stable trend; similarly, assign the label "stable trend / fluctuating trend / rising trend" to each task.

[0060] The judgment rules are as follows: when the continuous change trend is continuously rising for more than 70% of the time and the current trend persistence ratio is abnormal, it is considered an upward trend; when the continuous change trend alternates between rising and falling and the trend direction changes frequently, it is considered a fluctuating trend; when the continuous change trend changes less than the preset change value within a fixed time and is in the normal or fluctuating range, it is considered a stable trend.

[0061] Construct an execution dependency graph (directed task dependency graph) between tasks and analyze the role of the current task in the critical path: if the task is located in the most time-consuming period of the critical path and the resource status is "abnormal", it is marked as "primary bottleneck"; if there is an abnormal node in the dependent task and the current task has not yet recovered, it is marked as "secondary bottleneck"; if the current task has no dependent bottleneck task or is located in a non-critical path, it is "non-bottleneck path"; similarly, based on the output results, the structure is labeled as "primary bottleneck / secondary bottleneck / non-critical path".

[0062] Based on resource status, task allocation, and structural results, the task bottleneck contribution index B is determined according to a combination of rules. h When the task tags include "item resource anomaly", "rising trend", and "main bottleneck", B h High; when the task tags include "item resource fluctuation", "trend fluctuation" and "secondary bottleneck or primary bottleneck", B h For the middle; when the task labels include "Resources Normal", "Stable Trend" and "Non-Critical Path", B h It is low.

[0063] Based on the task bottleneck contribution index B h The sorting results; select the bottleneck task set T B Then, it enters the optimization strategy matching stage: specifically:

[0064] For each task t in the bottleneck task set TB h Extract each task t h Multidimensional features, including but not limited to: 1: Task type (computation-intensive / I / O-intensive / network call, etc.). 2: Average execution time; 3: Historical anomaly rate; 4: Depth of the dependency path; 5: System resource usage (CPU, memory); 6: Waiting time for tasks to be blocked before or after; 7: Historical optimization strategy response performance; The multi-dimensional features are integrated into a task feature vector through a feature encoding module. k={ 1, 2, ..., 7}, as the corresponding task t h The task profile.

[0065] Task feature vector k is input to the policy matching unit for matching; the trained policy is used to select agent π. k)→Lp, dynamically select the optimal strategy Lp0 based on historical optimization results and task type; if there are multiple optimal strategies Lp0, then the candidate strategy is Lp1.

[0066] The strategy matching unit establishes an optimization strategy library L={L1, L2, ..., Lu}. Each strategy Lp corresponds to a set of matching conditions Cp, scope of application, and optimization objective. Examples include: scheduling priority adjustment (suitable for light blocking tasks), delayed execution merging (suitable for batch computing tasks), concurrent decomposition processing (suitable for highly coupled long-running tasks), asynchronous processing replacement (suitable for waiting tasks), resource skew optimization (suitable for resource bottleneck tasks), and anomaly isolation and rapid recovery (suitable for high error rate tasks).

[0067] After optimizing the strategy matching deployment, for candidate strategy Lp1, a delay optimization model, a resource utilization optimization model, and anomaly control optimization model are constructed respectively, and the candidate strategy is independently evaluated in task t. h The performance of each model is evaluated; the corresponding optimization metrics are output, such as task execution time, CPU utilization, and anomaly rate; based on the Pareto non-dominated ranking method in multi-objective optimization, these three metrics are used as different dimensions to construct a multi-dimensional optimization space; by selecting the Pareto optimal solution set and eliminating non-optimal strategies that are superior to other strategies in all metrics, a set of candidate strategies that take into account latency, resource consumption, and anomaly control performance is output; as the task t h The final optimization strategy is Lp0; Lp0 is applied to the bottleneck task t. hThe system continuously collects data on execution latency, resource consumption, and anomaly rate changes through a real-time monitoring module. A fuzzy comprehensive evaluation model is employed, designing fuzzy membership functions for each indicator—task execution latency change, CPU resource consumption change, and anomaly rate change—to map indicator values ​​to fuzzy evaluation levels (e.g., excellent, good, average, poor). The corresponding fuzzy evaluation vector is then calculated to form a fuzzy evaluation matrix RJ. The weight vector W is then processed with the fuzzy evaluation matrix RJ to obtain the comprehensive fuzzy evaluation result BJ for each strategy. The comprehensive fuzzy evaluation result BJ is then defuzzified (e.g., using the centroid method) to transform it into a specific score value, which serves as a monitoring indicator. If the monitoring indicators meet the revenue threshold If the return is ≥θ1 (θ1 is a preset return threshold), the strategy is considered effective; if not, the alternative strategy reselection mechanism is triggered, and the strategy matching process is restarted.

[0068] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A virtual power plant operation optimization system, comprising: The power plant operation monitoring module and the optimal scheduling module are characterized in that: the power plant operation monitoring module obtains key parameter data in actual power plant operation, maps the actual power plant data to the input end of the virtual power plant model in a one-to-one correspondence, and establishes a parameter synchronization channel; the virtual power plant model is driven to operate through the input data; the virtual operation result is compared with the actual operation data in real time to identify the deviation; when the virtual operation result exceeds the corresponding set threshold, an abnormal backtracking analysis is automatically triggered, and the output path credibility is output: abnormal parameters are obtained, a causal dependency graph is called, all influence node paths are obtained through a graph traversal algorithm, nodes on each path are calculated according to the influence node path, the correlation degree of the path node and the abnormal node is calculated through mutual information, and then the calculated correlation degree is compared with a correlation threshold; nodes with a correlation degree greater than the correlation threshold are collected as possible cause nodes, and statistics are formed to form a possible cause set; Then, whether the possible cause nodes mutate before the abnormal points in time is verified, specifically: an inducing parameter is set to judge the mutation point and the change rate; if the conditions are met, the parameter is determined as an effective suspicious factor with time causality consistency; the suspicious factor set after screening and verification is matched with the patterns in the fault rule library; if there is a rule, the path is marked as a rule-consistent path; When several suspicious factors and causal paths are output, the credibility of each causal path is calculated, the correlation, time consistency and rule support degree are combined to output the path credibility; the credibility corresponding to each causal path is compared in size, and the most credible abnormal causal path is output, including: the correlation calculation value, the time consistency calculation value and the rule matching degree calculation value are used as three-dimensional features, a Softmax regression model is used for classification, the probability that each path belongs to a real cause and effect is output as the credibility, and then the credibility corresponding to each output causal path is compared in size, and the causal path with the maximum credibility is output as the most credible abnormal causal path; The optimal scheduling module obtains the most credible abnormal causal path to perform optimization processing, including process optimization and software operation optimization; the process optimization includes: a directed graph model containing task nodes and dependency edges is constructed to generate a legal schedulable topology structure; a weight model is constructed to identify bottleneck tasks: a process dependency directed graph model is constructed, and a directed acyclic graph is used to identify the circular dependency or illegal path in the process to generate a legal schedulable process topology; based on the generated legal schedulable process topology, the abnormal process path is quantitatively analyzed to obtain the key timing deviation corresponding to the influence efficiency and stability; if the key timing deviation is greater than zero, it means that the task is slower than expected; if the key timing deviation is less than zero, it means that the execution is ahead of schedule. Reconstruct the anomaly propagation path diagram, and calculate the anomaly sensitivity of the task; according to the obtained offset value and anomaly sensitivity, combine the execution complexity and historical risk score to construct a task priority model to obtain a bottleneck task identification factor; compare the bottleneck task identification factor with a preset threshold value, if the bottleneck task identification factor is greater than the preset threshold value, it is a bottleneck task; sort the bottleneck tasks obtained by the task weight model to obtain a set; For the identified bottleneck task, design a process reconstruction strategy; through the strategy collaborative execution process, introduce a real-time feedback mechanism after adjustment: continuously monitor the task execution time and newly appeared offset; if there is still a significant bottleneck after scheduling adjustment, trigger the next round of software operation optimization; For the bottleneck task, introduce a multi-layer perception machine classification model to output process path defects; then match the corresponding optimization strategy according to different process path defect types; software operation optimization includes: calculating the proportion of deviation from normal load for the use of CPU, memory and I / O resources to obtain CPU offset rate, memory offset rate and I / O offset; according to the influence degree of the use on the overall delay of the current task, calculate the task bottleneck contribution degree index: through a multi-stage evaluation process, identify the performance bottleneck in the current running task from three dimensions, including: calculating the instantaneous offset rate and short-term variance index for CPU, memory and I / O key resources, and then dividing the use of each resource into normal state, fluctuation state and abnormal state according to the instantaneous offset rate and short-term variance index; then according to the output result, assign normal, fluctuation and abnormal labels to each resource; use a sliding window to analyze the task resource change trend, detect the key resource continuous change trend and trend persistence ratio; according to the determination rule, output the state classification, including trend rising, trend fluctuation and trend stable; assign the trend stable, trend fluctuation and trend rising labels to each task; build an execution dependency graph between tasks, analyze the role of the current task in the critical path and output the result, and assign the structural judgment label as the main bottleneck, secondary bottleneck and non-critical path; Based on the resource state, task allocation and structure result, determine the task bottleneck contribution degree index according to the rules, select the bottleneck task set, enter the optimization strategy matching, obtain the task feature vector, input the task feature vector into the strategy matching unit, and match to obtain the optimal strategy; calculate the comprehensive optimization expected return value of the optimization strategy on the task; if the expected return value is greater than the preset return threshold value, the strategy is effective; if not, trigger the alternative strategy reselection mechanism and re-enter the strategy matching.

2. The virtual power plant operation optimization system of claim 1, wherein, The virtual building module is used for building a power plant virtual scene, traversing and matching all equipment in a standard equipment model library to obtain numbers of respective equipment, and generating an equipment parameter library; if a corresponding equipment shape size and operation parameter cannot be matched, custom modeling is performed; the custom modeling is performed according to actual drawings or point cloud data to analyze spatial geometric parameters and equipment logical parameters; input information is standardized into a platform geometric semantic format; and then a spatial coordinate system is established according to the standardized structure diagram and point cloud data, and each module component is arranged in a virtual space according to actual corresponding coordinates and sizes; For a specific equipment structure in a power plant, a modeling tool is used for high-precision modeling to simulate equipment shapes, support structures, interface connections and running parts, and a one-to-one space restoration is realized in combination with a material library and a texture map; Real-time data sources are accessed to drive the model to change in real time.

3. The virtual power plant operation optimization system of claim 1, wherein, The process of obtaining the optimal strategy is as follows: For each task in the bottleneck task set, multi-dimensional features of each task are extracted, the multi-dimensional features are integrated into a task feature vector through a feature coding module, and the task feature vector is taken as a task portrait of the corresponding task; the task feature vector is input into a strategy matching unit for matching; a trained strategy selection agent is used to dynamically select an optimal strategy based on historical optimization effects and task types; if there are multiple optimal strategies, candidate strategies are selected; the optimal strategy is matched and deployed, and after completion, for the candidate strategies, a comprehensive optimization expected return value of the candidate strategies on the task is calculated, a maximum expected return value is selected as an optimal optimization strategy.

Images