Virtual power plant optimization scheduling method and system

Abstract

The invention relates to the technical field of power system automatic management, in particular to a virtual power plant optimal scheduling method and system.The virtual power plant optimal scheduling method comprises the steps that real-time multi-mode state data of all nodes are collected through an intelligent electric meter and a sensor, and a time sequence input tensor is constructed; a prediction tensor is generated based on the power system resource state and historical operation data, and abnormity is identified through difference operation with the real-time tensor; constructing an abnormal propagation map and identifying key linkage nodes; generating a candidate disposal path and an edge mapping strategy according to the key node data; testing a response index of each path in the virtual environment, collecting residual data, and selecting an optimal path; scheduling a processing sequence based on node priorities; and dynamically updating model parameters and a scoring function structure according to execution feedback. According to the method, the problem that tiny anomalies are difficult to perceive under the condition of complex power demands is solved, and model correction and strategy updating can be carried out according to disturbance response residual errors and simulation deviations after actual execution.

Description

Technical Field

[0001] This invention relates to the field of power system automation management technology, and in particular to a virtual power plant optimization scheduling method and system. Background Technology

[0002] Virtual power plant (VPS) optimal dispatch is an advanced new power system management technology. It aggregates dispersed distributed energy resources for unified and optimized management through intelligent coordination and control, achieving dispatch capabilities similar to traditional power plants. With the increasing penetration of renewable energy and the advancement of power market reforms, how to realize VPS as a new operating model for aggregating distributed energy resources has become an important research direction. The rapid development of IoT, industrial internet, edge computing, and artificial intelligence technologies provides strong technical support for the state perception, data acquisition, and intelligent decision-making of VPS optimal dispatch.

[0003] However, existing virtual power plant optimization scheduling methods still face the following technical challenges: the heterogeneity and asynchronous nature of multimodal data make information fusion a significant challenge, and traditional methods struggle to fully exploit potential cross-information and temporal correlations; distributed resources are diverse and have varying characteristics, making unified scheduling difficult; anomaly detection algorithms often rely on static thresholds or single model outputs, lacking the ability to structurally model anomaly propagation links in complex systems, making it difficult to identify anomaly linkages and potential risk paths; existing handling methods typically employ rule-based pre-sets or single-path execution strategies, lacking dynamic feedback and strategy self-updating mechanisms, making it difficult to adapt to high-frequency changes in equipment status and diverse evolution of anomalies; for the selection and decision-making of multiple candidate handling paths, most virtual power plant optimization scheduling systems cannot perform pre-validation in a virtual environment and lack a systematic evaluation mechanism to determine the optimal response scheduling path. These shortcomings can easily lead to system response delays, resource scheduling imbalances, and the spread of anomaly risks in high-risk or high-precision scenarios, reducing the overall stability of the power system.

[0004] Therefore, virtual power plant optimization scheduling methods and systems are needed. Summary of the Invention

[0005] To address the problems mentioned in the background section, the present invention provides the following technical solution:

[0006] The specific technical solution of this invention is as follows:

[0007] The virtual power plant optimization scheduling method includes the following steps:

[0008] Step S1. Collect real-time multimodal status data of each node through smart meters and sensors. The multimodal status data includes electrical characteristic data, equipment status data, environmental status data, and operation control data, and construct a time series input tensor.

[0009] Step S2. Generate a prediction tensor based on the power system resource status and historical operating data, and identify anomalies by performing a difference operation with the real-time tensor;

[0010] Step S3. Construct an anomaly propagation graph and identify key linkage nodes;

[0011] Step S4. Generate candidate disposal paths and edge mapping strategies based on key node data;

[0012] Step S5. Test the response metrics of each path in a virtual environment and collect residual data to select the optimal path;

[0013] Step S6. Schedule the processing order based on node priority;

[0014] Step S7. Dynamically update the model parameters and scoring function structure based on the execution feedback.

[0015] Furthermore, the edge mapping strategy includes the following:

[0016] Based on the current topology and response latency characteristics of the edge nodes, a policy path mapping function is constructed.

[0017] Each key response action in the anomaly handling path is mapped to a corresponding edge node, and a linkage isolation control sequence is generated based on the execution coupling relationship and priority weight between nodes to form a distributed edge execution path;

[0018] The linkage isolation control sequence is dynamically trimmed based on environmental disturbance factors, its own state residuals, and task conflict factors.

[0019] Furthermore, based on the control and execution mechanism of the abnormal handling path on the edge node, a dynamic graph of path execution residuals is constructed in a multi-edge node environment. The graph is based on the execution residual sequence, resource occupancy status and task response stability index formed by different edge nodes in multiple historical feedback cycles. The residual trend tensor is extracted and a node behavior stability prediction model is constructed.

[0020] Based on the stability prediction model, a migration scheduling mapping relationship for candidate disposal paths is established, and the current preferred path is mapped to an edge node with sufficient remaining resources and high predicted stability according to the strategy scoring result.

[0021] During execution, path migration decisions are dynamically adjusted based on changes in node resource load and response delays after migration. A feedback update mechanism is constructed to achieve dynamic scheduling of disposal paths among edge nodes, execution residual suppression, and collaborative stability control.

[0022] Furthermore, in the dynamic scheduling mechanism between edge nodes for anomaly handling paths, after completing the construction of the dynamic graph of path execution residuals, the path segments corresponding to high-frequency offset nodes in the graph are extracted, and the resource consumption weight and response stability index of the path segment in multiple feedback cycles are calculated.

[0023] Construct a path segment pruning priority function Used to dynamically prune each path segment based on resource constraints, response bias, and residual weights;

[0024] The remaining path segments are constructed with residual minimization and stability maximization as joint optimization objective functions. The disposal paths are reorganized based on the graph structure shortest path algorithm, and the corrected disposal paths are executed according to the new node scheduling priority sequence to improve overall execution efficiency and multi-node response stability.

[0025] Furthermore, when the output value of the node behavior stability prediction model is lower than the preset stability threshold for two consecutive scheduling cycles, the following steps are executed:

[0026] Based on the constructed path execution residual dynamic graph, the path execution residual sequence is extracted;

[0027] A residual trend tensor is constructed based on the residual sequence to identify stability decay patterns.

[0028] The retrieval system uses a pre-stored path structure template library to filter candidate templates with a feature similarity of more than 85% to the current path.

[0029] The key node sequences in the candidate template are fine-tuned in terms of position and weight to generate alternative paths that adapt to the edge node states;

[0030] The alternative path is input into the stability prediction model for evaluation. Once the model meets the criteria for three consecutive evaluations, the original path is replaced and executed.

[0031] Furthermore, after the adaptive alternative path is loaded into the task scheduling sequence of the current edge node, a scoring residual trend tensor is constructed based on the scoring residual change trend and scoring fluctuation characteristics of the alternative path in the historical period.

[0032] Based on the aforementioned scoring residual trend tensor, the residual attribution sequence of each scoring factor in the scoring function is extracted, and the penalty parameter and stability evaluation criterion of the feedback scoring function are dynamically adjusted accordingly.

[0033] When the deviation of the scoring standard exceeds the preset tolerance threshold within two consecutive scheduling cycles, the evolution mechanism of the scoring function structure is triggered, the evaluation dimension of the scoring function is reconstructed, and a scoring adjustment factor based on the residual fluctuation frequency is introduced.

[0034] The scoring adjustment factor is applied to the current scoring function to re-evaluate the adaptive alternative path. When the scoring trend stabilizes and the scoring confidence interval converges to the target interval, the alternative path is marked as a long-term scheduling candidate path, and a stability index is generated for reference in subsequent scheduling cycles.

[0035] Furthermore, after the adaptive alternative path is loaded into the scheduling sequence of the edge node, a scoring residual trend tensor is constructed based on the scoring residual change trend and scoring fluctuation frequency of the alternative path in continuous historical cycles, in order to identify the feedback lag of the current scoring mechanism to path stability.

[0036] Based on the scoring residual trend tensor, the residual attribution sequence of each scoring factor in the scoring function is extracted, its offset magnitude and contribution ratio are calculated, and the penalty parameter and path stability evaluation dimension in the scoring function are dynamically adjusted according to the attribution results.

[0037] When the deviation of the stability judgment result of the scoring function exceeds the preset tolerance threshold in two consecutive scheduling cycles, the evolution mechanism of the scoring function structure is triggered, the evaluation sub-dimension of the scoring function is reconstructed, and a scoring adjustment factor based on the fluctuation frequency of the scoring residual is introduced to enhance the sensitivity of the scoring function to path behavior anomalies.

[0038] The scoring adjustment factor is modeled by weighting the residual fluctuation period, maximum amplitude and delay trend of each scoring factor, and is applied to the current scoring function to re-score the alternative path;

[0039] When the re-scoring result converges to the target interval within a continuous period and the scoring trend meets the stability criterion, the alternative path is marked as a long-term scheduling candidate path, and a stability index parameter is generated for reference in subsequent task scheduling cycles.

[0040] The virtual power plant optimized dispatch system includes the following:

[0041] The multimodal data acquisition module collects real-time multimodal status data from each node through smart meters and sensors. The multimodal status data includes, but is not limited to, electrical characteristic data, equipment status data, environmental status data, and operation control data. It also constructs a unified multimodal feature input tensor based on time series.

[0042] The residual generation and trend recognition module is used to construct a prediction tensor based on the power system resource status and historical operation data, and perform differential processing with the input tensor to form a residual tensor, further constructing a scoring residual trend tensor and identifying scoring fluctuation characteristics.

[0043] The anomaly propagation graph construction module is used to construct an anomaly propagation path graph based on the multidimensional distribution and temporal characteristics of the residual tensor, and to identify key response nodes and linkage path structures within it.

[0044] The strategy path generation module is used to construct a feedback-driven strategy path candidate set based on the state sequence and historical handling paths of key response nodes, and to select the optimal anomaly handling path by minimizing perturbation and restoring stability objective functions.

[0045] The virtual simulation and scoring evaluation module is used to simulate candidate paths in a simulation environment containing multiple potential fault evolution characteristics, collect feedback indicators of the paths in terms of response speed, anomaly suppression efficiency and stable recovery time, and calibrate the current optimal handling path based on the comprehensive scoring function.

[0046] The edge node scheduling and linkage module is used to map the optimal anomaly handling path to the edge node, generate linkage isolation control sequence based on topology, resource status and priority relationship, and perform path migration, task pruning and residual suppression scheduling between nodes.

[0047] The scoring function evolution and path stability evaluation module is used to trigger the scoring function structure reconstruction mechanism when the scoring deviation exceeds the threshold, adjust the scoring factor weights and penalty parameters based on the scoring residual trend tensor, and solidify the alternative path as a long-term scheduling candidate path when the scoring trend of the alternative path stabilizes.

[0048] The feedback and parameter adaptive update module is used to synchronously update the path selection parameters, level discrimination threshold and scoring function structure based on the disturbance residuals, score fluctuations and node state change trends generated during the execution process, so as to realize the periodic reconstruction and evolution optimization of the anomaly handling strategy.

[0049] In summary, the present invention has the following beneficial effects:

[0050] This invention constructs a time-series feature tensor from multiple heterogeneous sensor signals and performs differential processing with power system resource status prediction data to form a residual tensor to identify abnormal offsets. This not only achieves deep fusion of multimodal signals, but also, more importantly, solves the problem of difficulty in detecting minute anomalies under complex power demand conditions.

[0051] This invention constructs an anomaly propagation path map by analyzing the temporal changes and structural distribution of residual tensors. It can accurately identify the propagation chain of abnormal events within the power system and its key response nodes, thereby realizing the leap from single-point perception to system-wide linkage analysis and providing a systematic foundation for subsequent resource strategy scheduling and node priority allocation.

[0052] This invention introduces a virtual simulation environment to pre-run and score multiple anomaly handling paths, and constructs a dynamic evaluation function based on historical data to realize the trial operation, scoring and optimization of paths. At the same time, after actual execution, the model can be corrected and the strategy updated based on the disturbance response residual and simulation deviation. Attached Figure Description

[0053] Figure 1 This is a flowchart of the virtual power plant optimization scheduling method of the present invention;

[0054] Figure 2 This is a block diagram of the virtual power plant optimized scheduling system of the present invention. Detailed Implementation

[0055] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. It should also be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the scope of the invention.

[0056] See attached document Figure 1 This embodiment provides a virtual power plant optimized scheduling method, including the following steps:

[0057] S1. Real-time multimodal status data of each node is collected through smart meters and sensors. The multimodal status data includes, but is not limited to, electrical characteristic data, equipment status data, environmental status data, and operation control data. A unified multimodal feature input tensor X is constructed based on the time series. t ;

[0058] S2. Constructing a prediction tensor based on power system resource status and historical operation data. The predicted tensor is then differentially processed with the currently acquired tensor to form a residual tensor. To capture any possible abnormal offset patterns;

[0059] S3. Based on the multidimensional structural distribution and temporal variation characteristics of the residual tensor, construct an anomaly propagation path map and identify the linkage path nodes in the propagation chain to characterize the diffusion trend of anomalies among equipment components or subsystems.

[0060] S4. Based on the state sequence of key response nodes identified in the anomaly propagation chain and their historical handling behavior paths, construct a feedback-driven strategy update path, and based on the execution effect of the strategy path in multiple feedback cycles, establish a dynamic evaluation function with the goal of minimizing disturbances and achieving stable recovery, calculate the expected value of multiple candidate handling paths and select the best alternative, and generate an adaptive and variable optimal anomaly handling path sequence.

[0061] S5. Among the multiple candidate anomaly handling paths generated, a virtual simulation environment containing multiple potential fault evolution characteristics is constructed. Each path is sequentially mapped to the simulation environment for trial operation. Multi-dimensional feedback indicators such as response speed, anomaly suppression efficiency and stable recovery time of the handling path under simulation conditions are collected. The optimal path is determined based on the path comprehensive scoring function and marked as the current anomaly handling execution path.

[0062] S6. Based on the historical anomaly trigger frequency, current anomaly level and structural position of the target node in the anomaly propagation chain, combined with the simulation scoring results in S5, construct a multi-factor fusion anomaly level weight model, and update the strategy allocation order according to node priority to realize the early execution of high-risk node handling paths and dynamic buffer scheduling of low-weight node paths.

[0063] S7. After the current anomaly handling path is completed, based on the changing trends of the disturbance response residual, simulation score deviation and edge node environmental disturbance factor during the handling process, a disturbance parameter correction function is constructed and the path scoring model, level discrimination threshold and path optimization strategy parameters are updated synchronously to improve the adaptability and convergence of the strategy path in multi-cycle anomaly environment.

[0064] In this embodiment of the invention, after all data is synchronized with the local timestamp, it is normalized and encoded with a tensor structure to construct a third-order input feature tensor with dimensions of sensor type × time window × data channel.

[0065] Using the LSTM-GRU hybrid prediction model deployed on the main control platform, a prediction tensor is generated based on historical 12-hour sliding window data;

[0066] Subsequently, element-wise difference operations are performed between the current input tensor and the prediction tensor to generate the residual tensor;

[0067] The portion of the residual value that is greater than the set threshold (mean + 2 × standard deviation) is marked as a potential outlier region.

[0068] The temporal evolution trend of the residual tensor is modeled using a structural graph. The linked nodes in each residual propagation chain are extracted using GAT (Graph Attention Network) to form an anomaly propagation map.

[0069] When multiple nodes exhibit continuous propagation of residual peaks within 20 seconds, identify the chain as the main propagation path and lock the node with the shortest response delay in the path as the key response node.

[0070] Using the three most recent anomaly handling records of key response nodes as state sequence inputs, a feedback update path based on the policy evolution function is constructed.

[0071] The function employs an improved reinforcement learning algorithm with a perturbation index R. d and recovery time Tr The path is scored for the objective function, and the scoring formula is as follows:

[0072]

[0073] β = 0.4

[0074] α = 0.6

[0075] Among them, R d T represents the cumulative value of the residuals of abnormal disturbances under the current path. r The time taken for the abnormal state to recover to a steady state is expressed in seconds. The path with the highest score is selected to form the optimal treatment path sequence.

[0076] Based on the selected path, construct a digital twin simulation environment:

[0077] During the simulation, the response time, suppression magnitude, and system stabilization time for each path are recorded, and the current execution path is determined based on the comprehensive simulation score.

[0078] Historical anomaly levels, trigger frequencies, and structural locations in the propagation graphs of each device node are collected, normalized, and then a priority scoring model is constructed using a multi-factor weighted function.

[0079] P i =γ1·Freq i +γ2·RiskLevel i +γ3·GraphDepth i γ i ∈[0,1];

[0080] With P i The value prioritizes all paths to be processed, so that high-risk nodes are scheduled first, while low-priority paths are included in the buffer queue.

[0081] After the disposal path is executed, the process disturbance residuals and simulation deviations are recorded. The strategy scoring model and threshold function parameters are updated using the minimum mean square error (MSE) function. The path scoring strategy is adjusted synchronously according to changes in external meteorological disturbances, such as humidity changes exceeding 15% or wind speed fluctuations exceeding 20%, in order to achieve adaptive learning and long-term optimization.

[0082] When the generated anomaly handling path contains multiple critical response nodes, a policy path mapping function is constructed based on the current edge node topology and response latency characteristics. Each key response action in the anomaly handling path is mapped to the corresponding edge node, and a linkage isolation control sequence is generated based on the execution coupling relationship and priority weight between nodes to form a distributed edge execution path;

[0083] The linkage isolation control sequence is dynamically trimmed based on environmental disturbance factors, its own state residuals, and task conflict factors to ensure the executability and stability of the path under edge conditions.

[0084] Based on the generated anomaly handling path sequence, a control mapping relationship is constructed for the edge execution nodes in multiple virtual power plant power resource scheduling systems, including the following specific steps:

[0085] S411: Based on the physical connection structure and control response hierarchy of equipment in the power system, establish an edge node topology graph G = (V, E), where node V i This represents each execution unit, such as energy storage node devices, load node devices, power supply node devices, etc., with edge E. ij This indicates the response dependency or control latency channel between nodes;

[0086] Each node is allocated a processing latency coefficient δ i With resource state vector R i (t), used to represent the load feasibility of the current node;

[0087] S412: Process each step of the strategy path a k Mapped to edge control action sequence φ(a) k )={v i ,op i}, where op i Indicates the type of physical operation to be performed on the corresponding node;

[0088] The sequence of control actions must satisfy the temporal dependencies and safety constraints between operations. Candidate paths are filtered by judging the following mapping feasibility function:

[0089]

[0090] S413: For all action sequences filtered by mapping, establish the edge node action scheduling matrix:

[0091]

[0092] Where t start With t end Calculated based on system latency and priority;

[0093] If a task conflict is detected, such as two tasks being scheduled to the same node at the same time, the scheduling will be restructured using the conflict penalty function.

[0094]

[0095] S414: If multiple resource conflicting nodes or resource overload occurs, a scoring function is used to prune candidate paths.

[0096] Score(a k ) = α·contribution to abnormal suppression + β·resource consumption efficiency - γ·P confilct ;

[0097] High-scoring paths are retained, while paths with high resource costs and large response delays are discarded to generate the final executable linkage control sequence;

[0098] S415: Send the actions in the final scheduling matrix to each edge node in a timely manner for execution by the local controller;

[0099] During execution, each node sends back its response time, action completion status, and local disturbance impact value, and updates the path execution log.

[0100] If an execution failure or abnormal feedback exceeds the threshold, the backup control path switching mechanism is triggered.

[0101] The control execution mechanism based on the anomaly handling path at the edge nodes further includes:

[0102] In a multi-edge node environment, a dynamic graph of path execution residuals is constructed. The graph is based on the execution residual sequence, resource occupancy status and task response stability index formed by different edge nodes in multiple historical feedback cycles. The residual trend tensor is extracted and a node behavior stability prediction model is constructed.

[0103] Based on the stability prediction model, a migration scheduling mapping relationship for candidate disposal paths is established, and the current preferred path is mapped to the edge node with sufficient remaining resources and high predicted stability according to the strategy scoring results.

[0104] During execution, the path migration decision is dynamically adjusted based on changes in node resource load and response delay after migration. A feedback update mechanism is constructed to realize dynamic scheduling of the handling path among edge nodes, execution residual suppression and collaborative stability control, so as to improve the overall robustness and response consistency of the abnormal handling path.

[0105] In this embodiment, to enhance the stability and adaptability of the exception handling path during execution at edge nodes, the following implementation steps are adopted:

[0106] First, for each edge node E i During the execution of the power dispatch anomaly handling path across multiple feedback cycles N, the response index residuals after each round of path execution are collected, and an execution residual sequence tensor is constructed:

[0107]

[0108] in This represents the response error value at time t during the nth round of path execution, such as disturbance suppression residual, recovery offset residual, etc.

[0109] Next, for each node E i Extract the maximum offset from the residual sequence with N periods.

[0110] residual mean square deviation With stability index function S(E) i The execution stability score of the computing node:

[0111]

[0112] By inputting the stability scores of each edge node into the scheduling mapping function Ψ, a node migration priority vector is generated:

[0113] Ψ:{S(E1),S(E2),...,S(E m )}→{P(E1),P(E2),...,P(E m )};

[0114] Where P(E) i ) represents node E i The path migration priority is determined by the score; the lower the score, the higher the priority.

[0115] Subsequently, the task execution of the processing path of the node with the lowest current score is moved to the adjacent node with the higher score. For example, if P(E3) > P(E7), the E3 node path is moved to E7 for execution. At the same time, new residual data is collected after execution to update the tensor. This information is then fed back to the mapping function Ψ for dynamic reallocation, resulting in the following closed-loop path migration graph:

[0116]

[0117] The above methods can achieve dynamic scheduling, execution optimization, and adaptive stability improvement of anomaly handling paths without significantly increasing edge load, and can be integrated into the main controller strategy evolution module to participate in the next round of path optimization strategy selection.

[0118] In this embodiment of the invention, the dynamic scheduling mechanism for anomaly handling paths among edge nodes includes the following:

[0119] After completing the construction of the path execution residual dynamic graph, the path segments corresponding to the high-frequency offset nodes in the graph are extracted, and the resource consumption weight and response stability index of the path segment in multiple feedback cycles are calculated.

[0120] Construct a path segment pruning priority function Used to dynamically prune each path segment based on resource constraints, response bias, and residual weights;

[0121] The remaining path segments are constructed with residual minimization and stability maximization as the joint optimization objective function. The disposal path is reorganized based on the graph structure shortest path algorithm, and the modified disposal path is executed according to the new node scheduling priority sequence to improve the overall execution efficiency and multi-node response stability.

[0122] In this embodiment, in order to verify the effectiveness of the edge node dynamic scheduling strategy based on the abnormal handling path pruning and reconstruction mechanism in the abnormal handling scenario of power dispatching system equipment, a set of simulation experiments were designed.

[0123] The experimental environment uses a virtual edge computing network simulation platform to simulate six heterogeneous edge nodes, and deploy historical task datasets and different resource load models respectively.

[0124] The simulated task is the execution scheduling task after an abnormal alarm of power system equipment. This task has moderate resource consumption volatility and real-time requirements, and is suitable for verification in this embodiment.

[0125] First, four different anomaly handling paths were constructed as a comparison group: existing path A (traditional scheduling strategy), path B (residual optimization strategy), path C (node ​​priority scheduling mechanism), and path D (the pruning and reconstruction joint mechanism proposed in this invention).

[0126] Each path is simulated for 20 rounds of scheduling tasks under the same period. During the process, the resource utilization rate, execution residual (expressed as the standard deviation between path feedback and predicted response), path switching delay time (ms), stability score (comprehensively considering node response jitter, feedback stability and residual gradient) and task completion efficiency improvement ratio (based on path A) of each node are recorded in real time.

[0127] In this process, path D constructs a high-frequency offset node feature map based on the execution residual map, extracts resource consumption weights and response stability scores for multi-cycle path segments, and jointly constructs a pruning priority function. And execute the strategy compression;

[0128] The path segments are then regrouped and sorted based on the joint score of residuals and resource indicators, and mapped to the node with the best stability prediction score.

[0129] During path execution, a migration feedback correction mechanism is constructed to automatically suppress residual jumps and dynamically schedule path segments, thereby achieving pruning and reconstruction between edge nodes. See the table for details.

[0130] Experimental data table of abnormal handling path pruning and reconstruction

[0131]

[0132] According to the table above, by constructing a dynamic graph with path residual as the driving parameter, the system can identify abnormal fluctuation trends in the execution path in advance, and on this basis guide the pruning and optimization process to converge towards the dual indicators of resource consumption rate and path deviation amplitude.

[0133] The pruning process is not carried out in isolation, but is further coupled with the historical stability score results of the execution nodes, so that the optimal path not only has the ability to compress resources, but also shows strong adaptability in task response.

[0134] Subsequently, the system uses the execution deviation trend function derived from the prediction model to reconstruct and adjust the candidate path set, forming a more balanced execution distribution pattern.

[0135] The refactored structure continuously receives feedback signals during execution and dynamically fine-tunes itself based on real-time load conditions to ensure minimal path execution latency and a stable improvement in task scheduling success rate.

[0136] Verification showed that the selected path D outperformed paths A, B, and C in all five key performance indicators: its resource utilization was reduced to 59.7%, the path residual was controlled within 0.7, the response latency was reduced to 89ms, the path stability score reached 9.3 points, and the task completion rate was increased to 19.4%.

[0137] The above performance clearly demonstrates that the path optimization mechanism proposed in this embodiment can exhibit strong adaptability, high stability, and significant resource efficiency improvement in scenarios involving multi-task collaboration, high-pressure load on edge nodes, and sudden abnormal disturbances.

[0138] If the preset scoring threshold is not reached within two or more consecutive scheduling cycles of the feedback scoring function, the path scoring residual sequence is extracted based on the path execution residual map, and a residual trend tensor for trend analysis is constructed.

[0139] Based on the score decay pattern identified in the residual trend tensor, a preset path structure template library is retrieved, and candidate path templates with similarity to the current path structure features above a threshold value are selected.

[0140] Fine-tune the key node sequence in the candidate path template to form alternative path candidates that can adapt to the current edge node state characteristics;

[0141] Alternative path candidates are incorporated into the scheduling path scoring system and evaluated in real time through a feedback scoring function. Once the continuous scoring meets the stability requirements, they are solidified as the main execution path for the current scheduling cycle to replace the original path in execution.

[0142] In this embodiment, it is necessary to further explain that the feedback scoring function refers to the function structure that dynamically adjusts the scoring parameters according to the path execution results during the task scheduling process. It has the characteristics of periodic feedback, adaptive adjustment and multi-dimensional scoring. This function not only considers static indicators such as path length or time delay, but also introduces dynamic evolution factors such as scoring residual fluctuation and stability change rate to guide path evolution.

[0143] The path execution residual graph is a multi-dimensional residual representation structure built around the difference between the path execution result and the expected score. The graph consists of a tensor graph with time and score dimensions to capture the performance degradation trend and local fluctuation pattern of the scheduling path in multiple cycles, supporting subsequent trend identification and template matching.

[0144] The residual trend tensor refers to a high-dimensional tensor structure extracted and constructed from the score residual map for trend modeling. Its dimensions can cover time windows, score factors, node positions, etc., to model the dynamic evolution of path scores in time series. This tensor supports tools such as singular value decomposition and Fourier analysis to perform pattern recognition of trends.

[0145] The path structure template library is a knowledge base containing predefined path structure information. It summarizes the structure based on historical execution data, expert experience and typical scenarios. Its role is to provide structural reference samples, support the efficient search for alternative paths in similar structural domains after the scheduling path fails, and has the ability to quickly match structures and abstract features.

[0146] Key node fine-tuning refers to making minor adjustments to the position, weight, or order of some key scheduling nodes in the candidate path template to better adapt them to the current edge node status (such as load, resource utilization, and latency sensitivity). This strategy significantly reduces scheduling costs and improves matching accuracy by dynamically adjusting the local structure rather than reconstructing the entire path.

[0147] Therefore, by constructing a score residual trend tensor, we can proactively identify the risk of score decay and sudden changes, which strengthens the feedforward control capability of the scheduling system. Furthermore, by using a structural template library and a trend alignment screening mechanism, we can avoid blindness and resource waste in the path replacement process.

[0148] Furthermore, by fine-tuning the key node strategy, the path structure can be reconstructed in a refined manner, significantly reducing scheduling costs and enhancing adaptability to edge environments. Finally, the path is solidified by a dynamic scoring function, ensuring the dual stability of the path in terms of local scoring and periodic performance, thus avoiding the problem of high scores and low reliability that is common in traditional methods.

[0149] After the adaptive alternative path is loaded into the task scheduling sequence of the current edge node, a score residual trend tensor is constructed based on the score residual change trend and score fluctuation characteristics of the alternative path in the historical period.

[0150] Based on the scoring residual trend tensor, the residual attribution sequence of each scoring factor in the scoring function is extracted, and the penalty parameter and stability evaluation criterion of the feedback scoring function are dynamically adjusted accordingly.

[0151] When the deviation of the scoring standard exceeds the preset tolerance threshold within two consecutive scheduling cycles, the evolution mechanism of the scoring function structure is triggered, the evaluation dimension of the scoring function is reconstructed, and a scoring adjustment factor based on the residual fluctuation frequency is introduced.

[0152] The scoring adjustment factor is applied to the current scoring function to re-evaluate the adaptive alternative path. When the scoring trend stabilizes and the scoring confidence interval converges to the target interval, the alternative path is marked as a long-term scheduling candidate path, and a stability index is generated for reference in subsequent scheduling cycles.

[0153] In this embodiment, to verify the effectiveness of the scoring function reconstruction and long-term candidate path identification method based on the scoring residual trend tensor proposed in the above embodiments, a set of simulated edge node scheduling scenarios were constructed to test the impact of different scoring function configurations on the stability judgment and scoring trend of alternative paths.

[0154] The experimental scenario is set as a task scheduling node in a remote intelligent detection system for chemical equipment. The goal is to select the most stable alternative path to include in the long-term scheduling candidate in 30 consecutive scheduling cycles.

[0155] The experimental data includes four types of scoring function configurations:

[0156] Class A uses a traditional static scoring function with no scoring structure evolution mechanism;

[0157] Class B is the scoring function of this invention, which has a scoring residual trend tensor construction and function reconstruction mechanism;

[0158] Type C is a multi-factor weighted scoring function in the existing technology, which has a simple factor adjustment function;

[0159] Class D is a random rating function simulating a scenario where the rating trend is unstable;

[0160] Each type of function receives real-time path score input during the scheduling simulation process, and records the score trend changes, the convergence speed of the score confidence interval, the frequency of offset tolerance exceeding the threshold, and the final stability index.

[0161] The scoring residual trend tensor is constructed based on the periodic fluctuations in the historical scoring sequence of the path, with the periodic average difference as the tensor's basic dimension; the scoring adjustment factor is dynamically introduced based on the frequency domain analysis results to adjust the sensitivity coefficient of the feedback function and the weight of the evaluation factor; when the scoring trend deviates beyond the preset threshold (set as ±15% of the scoring mean) within two consecutive periods, the scoring function dimension is automatically reconstructed, and a new round of attribution factor feedback mechanism is introduced.

[0162] The experimental environment was built using a Python simulation platform, and a fixed path scoring perturbation model was used for uniform perturbation input to ensure the fairness and reproducibility of the experiment.

[0163] The experimental results are shown in the table below:

[0164] Comparison Table of Experimental Results on Trend Regulation of Scoring Function

[0165]

[0166] From the perspective of the stability of the scoring trend (measured by the standard deviation of the scoring), the scoring function B of this invention has the lowest fluctuation value during the scheduling period, which is only 0.8σ. This is significantly better than the traditional scoring function A (2.3σ) and the weighted function C (1.7σ), and much lower than the random function D (3.2σ) which has significant scoring fluctuation. This indicates that the B scheme has better trend stability.

[0167] In terms of the convergence time of the score confidence interval, Scheme B only requires 6 cycles to achieve score interval convergence, which is better than A (15 cycles) and C (12 cycles), demonstrating its faster adaptability to the changing trend of path scores.

[0168] Scheme D failed to stabilize after more than 20 cycles, indicating that the scoring function lacked effective control.

[0169] The frequency of deviation tolerance exceeding the threshold is a key indicator for measuring the deviation of the score from the target standard. Option B has 0 times, which shows that the method of the present invention has a strong error control capability in residual trend judgment and score function structure evolution.

[0170] In comparison, A and C have 4 and 2 offsets respectively, while D frequently crosses the boundary 6 times, making it the least reliable.

[0171] In terms of the final stability index, the solution of this invention scored 0.93, which is much higher than A (0.52), C (0.68) and D (0.41), indicating that the method has higher accuracy in identifying long-term candidate paths in scheduling strategy optimization.

[0172] In summary, this embodiment effectively overcomes the shortcomings of traditional scoring mechanisms, such as scoring lag and inability to adapt to changing scoring trends, by constructing a dynamic evolution mechanism for the scoring residual trend tensor and the scoring function. It demonstrates significant technological progress and innovative effects in long-term stable scheduling path identification, especially in the evolution mechanism of the scoring function, the rapid convergence capability of the confidence interval, and the fine adjustment capability of residual attribution.

[0173] After the adaptive alternative path is loaded into the scheduling sequence of the edge node, a scoring residual trend tensor is constructed based on the scoring residual change trend and scoring fluctuation frequency of the alternative path in continuous historical periods, in order to identify the feedback lag of the current scoring mechanism to the path stability.

[0174] Based on the scoring residual trend tensor, the residual attribution sequence of each scoring factor in the scoring function is extracted, its offset magnitude and contribution ratio are calculated, and the penalty parameter and path stability evaluation dimension in the scoring function are dynamically adjusted according to the attribution results.

[0175] When the deviation of the stability judgment result of the scoring function exceeds the preset tolerance threshold in two consecutive scheduling cycles, the evolution mechanism of the scoring function structure is triggered, the evaluation sub-dimension of the scoring function is reconstructed, and a scoring adjustment factor based on the fluctuation frequency of the scoring residual is introduced to enhance the sensitivity of the scoring function to path behavior anomalies.

[0176] The scoring adjustment factor is modeled by weighting the residual fluctuation period, maximum amplitude and delay trend of each scoring factor, and is applied to the current scoring function to re-score the alternative path;

[0177] When the re-scoring result converges to the target interval within a continuous period and the scoring trend meets the stability criterion, the alternative path is marked as a long-term scheduling candidate path, and a stability index parameter is generated for reference in subsequent task scheduling cycles.

[0178] In this embodiment, to achieve dynamic tracking and stability evaluation of the scoring effect of adaptive alternative paths, firstly, after the alternative path is loaded into the task scheduling sequence of the current edge node, its scoring results over multiple historical scheduling cycles are collected. Then, a three-dimensional scoring residual trend tensor T is constructed using scoring factors, such as response time, resource consumption, and anomaly suppression effect, as dimensions. ijk , where i represents the scoring factor number, j represents the scheduling cycle number, and k represents the edge node number;

[0179] The scoring residual is defined as the difference between the current period's score and the baseline expected score, and its tensor element form is as follows:

[0180]

[0181] The scoring results are processed using a normalization method, with a value range of [-1, 1]. The scoring residual is recorded once for each scheduling cycle, and the sample time window is 5 to 10 consecutive cycles.

[0182] Next, statistical analysis was performed on the tensor to extract the residual fluctuation standard deviation corresponding to each scoring factor. Compared with the reference expected mean Constructing a rating moderating factor γ i This is used to correct the response strength of the current scoring function under different factor dimensions:

[0183]

[0184] Where ∈ represents the smallest positive value to prevent division by zero;

[0185] Regulatory factor γ i It is used to amplify or attenuate the factor weights in the current scoring function, thereby constructing a scoring function F that strengthens the residual attribution mechanism. adj ;

[0186] If the total score of the scoring function is lower than the set scoring threshold (e.g., 0.6) for two or more consecutive scheduling cycles, and the offset of each scoring factor exceeds the preset tolerance threshold δ = 0.1, then the evolution mechanism of the scoring function will be triggered.

[0187] This mechanism retrieves the residual attribution sequence along the rating factor dimension and, based on the residual fluctuation frequency, automatically removes unstable factors or introduces new factors, such as "rating trend variance," and reconstructs the evaluation function structure to a new version F. t+1 The specific evolutionary rules are shown in the table below:

[0188] Example table of evolution rules

[0189]

[0190]

[0191] Subsequently, the confidence interval of the newly constructed scoring function was evaluated;

[0192] With a confidence level of 95%, the following confidence interval is calculated for each rating factor:

[0193]

[0194] in σ is the historical average score. i Here, n represents the historical standard deviation, and n is the number of sampling periods. If the score values ​​are all within the confidence interval for three consecutive periods, and the interval width is less than 0.08, the scoring results are considered to have stabilized.

[0195] Once the above stability requirements are met, the alternative path is marked as a long-term scheduling candidate path. A stability index I is then generated by combining the path's score trend volatility, mean stability, and score deviation over historical periods. stab The calculation method is as follows:

[0196]

[0197] Where Var(St) represents the variance of the rating fluctuation. μ is the average score for the current period. target The center value of the target scoring interval;

[0198] The stability index is used for path selection, priority setting, and dynamic scheduling judgment in subsequent scheduling strategies.

[0199] Through the above mechanism, this invention not only realizes the dynamic evolution and adaptive adjustment of the scoring mechanism, but also constructs a complete evaluation and decision-making link in multiple dimensions, such as scoring factors, trend changes, statistical confidence, and path labels, ensuring the high stability, high responsiveness, and continuous optimization capability of the anomaly handling path in different edge node environments.

[0200] See attached document Figure 2 This embodiment provides a virtual power plant optimized scheduling system, including the following:

[0201] The multimodal data acquisition module collects real-time multimodal status data from each node through smart meters and sensors. The multimodal status data includes, but is not limited to, electrical characteristic data, equipment status data, environmental status data, and operation control data. It also constructs a unified multimodal feature input tensor based on time series.

[0202] The residual generation and trend recognition module is used to construct a prediction tensor based on the power system resource status and historical operation data, and perform differential processing with the input tensor to form a residual tensor, further constructing a scoring residual trend tensor and identifying scoring fluctuation characteristics.

[0203] The anomaly propagation graph construction module is used to construct an anomaly propagation path graph based on the multidimensional distribution and temporal characteristics of the residual tensor, and to identify key response nodes and linkage path structures within it.

[0204] The strategy path generation module is used to construct a feedback-driven strategy path candidate set based on the state sequence and historical handling paths of key response nodes, and to select the optimal anomaly handling path by minimizing perturbation and restoring stability objective functions.

[0205] The virtual simulation and scoring evaluation module is used to simulate candidate paths in a simulation environment containing multiple potential fault evolution characteristics, collect feedback indicators of the paths in terms of response speed, anomaly suppression efficiency and stable recovery time, and calibrate the current optimal handling path based on the comprehensive scoring function.

[0206] The edge node scheduling and linkage module is used to map the optimal anomaly handling path to the edge node, generate linkage isolation control sequence based on topology, resource status and priority relationship, and perform path migration, task pruning and residual suppression scheduling between nodes.

[0207] The scoring function evolution and path stability evaluation module is used to trigger the scoring function structure reconstruction mechanism when the scoring deviation exceeds the threshold, adjust the scoring factor weights and penalty parameters based on the scoring residual trend tensor, and solidify the alternative path as a long-term scheduling candidate path when the scoring trend of the alternative path stabilizes.

[0208] The feedback and parameter adaptive update module is used to synchronously update the path selection parameters, level discrimination threshold and scoring function structure based on the disturbance residuals, score fluctuations and node state change trends generated during the execution process, so as to realize the periodic reconstruction and evolution optimization of the anomaly handling strategy.

[0209] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0210] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0211] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0212] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0213] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A virtual power plant optimized scheduling method, characterized in that, Includes the following steps: Step S1. Collect real-time multimodal status data of each node through smart meters and sensors. The multimodal status data includes electrical characteristic data, equipment status data, environmental status data, and operation control data, and construct a time series input tensor. Step S2. Generate a prediction tensor based on the power system resource status and historical operating data, and identify anomalies by performing a difference operation with the real-time tensor; Step S3. Construct an anomaly propagation graph and identify key linkage nodes; Step S4. Generate candidate disposal paths and edge mapping strategies based on key node data; Step S5. Test the response metrics of each path in a virtual environment and collect residual data to select the optimal path; Step S6. Schedule the processing order based on node priority; Step S7. Dynamically update the model parameters and scoring function structure based on the execution feedback.

2. The virtual power plant optimization scheduling method according to claim 1, characterized in that, The edge mapping strategy includes the following: Based on the current topology and response latency characteristics of the edge nodes, a policy path mapping function is constructed. Each key response action in the anomaly handling path is mapped to the corresponding edge node, and a linkage isolation control sequence is generated based on the execution coupling relationship and priority weight between nodes to form a distributed edge execution path; The linkage isolation control sequence is dynamically trimmed based on environmental disturbance factors, its own state residuals, and task conflict factors.

3. The virtual power plant optimization scheduling method according to claim 2, characterized in that, Based on the control and execution mechanism of the abnormal handling path on the edge node, a dynamic graph of path execution residual is constructed in a multi-edge node environment. The dynamic graph of path execution residual is based on the execution residual sequence, resource occupation status and task response stability index formed by different edge nodes in multiple historical feedback cycles. The residual trend tensor is extracted and a node behavior stability prediction model is constructed. Based on the stability prediction model, a migration scheduling mapping relationship for candidate disposal paths is established, and the current preferred path is mapped to an edge node with sufficient remaining resources and high predicted stability according to the strategy scoring result. During execution, path migration decisions are dynamically adjusted based on changes in node resource load and response delays after migration. A feedback update mechanism is constructed to achieve dynamic scheduling of disposal paths among edge nodes, execution residual suppression, and collaborative stability control.

4. The virtual power plant optimization scheduling method according to claim 3, characterized in that, In the dynamic scheduling mechanism between edge nodes for abnormal handling paths, after completing the construction of the dynamic graph of path execution residuals, the path segments corresponding to high-frequency offset nodes in the graph are extracted, and the resource consumption weight and response stability index of the path segment in multiple feedback cycles are calculated. Construct the path segment pruning priority function Φ(S) k This is used to dynamically prune each path segment based on resource constraints, response bias, and residual weights. The remaining path segments are constructed with residual minimization and stability maximization as joint optimization objective functions. The disposal paths are reorganized based on the graph structure shortest path algorithm, and the corrected disposal paths are executed according to the new node scheduling priority sequence to improve overall execution efficiency and multi-node response stability.

5. The virtual power plant optimization scheduling method according to claim 4, characterized in that, When the output value of the node behavior stability prediction model is lower than the preset stability threshold for two consecutive scheduling cycles, the following steps are executed: Based on the constructed path execution residual dynamic graph, the path execution residual sequence is extracted; A residual trend tensor is constructed based on the residual sequence to identify stability decay patterns. The retrieval system uses a pre-stored path structure template library to filter candidate templates with a feature similarity of more than 85% to the current path. The key node sequences in the candidate template are fine-tuned in terms of position and weight to generate alternative paths that adapt to the edge node states; The alternative path is input into the stability prediction model for evaluation. Once the model meets the criteria for three consecutive evaluations, the original path is replaced and executed.

6. The virtual power plant optimization scheduling method according to claim 5, characterized in that, After the adaptive alternative path is loaded into the task scheduling sequence of the current edge node, a scoring residual trend tensor is constructed based on the scoring residual change trend and scoring fluctuation characteristics of the alternative path in the historical period. Based on the aforementioned scoring residual trend tensor, the residual attribution sequence of each scoring factor in the scoring function is extracted, and the penalty parameter and stability evaluation criterion of the feedback scoring function are dynamically adjusted accordingly. When the deviation of the scoring standard exceeds the preset tolerance threshold within two consecutive scheduling cycles, the evolution mechanism of the scoring function structure is triggered, the evaluation dimension of the scoring function is reconstructed, and a scoring adjustment factor based on the residual fluctuation frequency is introduced. The scoring adjustment factor is applied to the current scoring function to re-evaluate the adaptive alternative path. When the scoring trend stabilizes and the scoring confidence interval converges to the target interval, the alternative path is marked as a long-term scheduling candidate path, and a stability index is generated for reference in subsequent scheduling cycles.

7. The virtual power plant optimization scheduling method according to claim 6, characterized in that, After the adaptive alternative path is loaded into the scheduling sequence of the edge node, a scoring residual trend tensor is constructed based on the scoring residual change trend and scoring fluctuation frequency of the alternative path in continuous historical periods, in order to identify the feedback lag of the current scoring mechanism to the path stability. Based on the scoring residual trend tensor, the residual attribution sequence of each scoring factor in the scoring function is extracted, its offset magnitude and contribution ratio are calculated, and the penalty parameter and path stability evaluation dimension in the scoring function are dynamically adjusted according to the attribution results. When the deviation of the stability judgment result of the scoring function exceeds the preset tolerance threshold in two consecutive scheduling cycles, the evolution mechanism of the scoring function structure is triggered, the evaluation sub-dimension of the scoring function is reconstructed, and a scoring adjustment factor based on the fluctuation frequency of the scoring residual is introduced to enhance the sensitivity of the scoring function to path behavior anomalies. The scoring adjustment factor is modeled by weighting the residual fluctuation period, maximum amplitude and delay trend of each scoring factor, and is applied to the current scoring function to re-score the alternative path; When the re-scoring result converges to the target interval within a continuous period and the scoring trend meets the stability criterion, the alternative path is marked as a long-term scheduling candidate path, and a stability index parameter is generated for reference in subsequent task scheduling cycles.

8. A virtual power plant optimization scheduling system, applied to the virtual power plant optimization scheduling method of claim 1, characterized in that, Includes the following: The multimodal data acquisition module collects real-time multimodal status data from each node through smart meters and sensors. The multimodal status data includes, but is not limited to, electrical characteristic data, equipment status data, environmental status data, and operation control data. It also constructs a unified multimodal feature input tensor based on time series. The residual generation and trend recognition module is used to construct a prediction tensor based on the power system resource status and historical operation data, and perform differential processing with the input tensor to form a residual tensor, further constructing a scoring residual trend tensor and identifying scoring fluctuation characteristics. The anomaly propagation graph construction module is used to construct an anomaly propagation path graph based on the multidimensional distribution and temporal characteristics of the residual tensor, and to identify key response nodes and linkage path structures within it. The strategy path generation module is used to construct a feedback-driven strategy path candidate set based on the state sequence and historical handling paths of key response nodes, and to select the optimal anomaly handling path by minimizing perturbation and restoring stability objective functions. The virtual simulation and scoring evaluation module is used to simulate candidate paths in a simulation environment containing multiple potential fault evolution characteristics, collect feedback indicators of path response speed, anomaly suppression efficiency and stable recovery time, and calibrate the current optimal handling path based on the comprehensive scoring function. The edge node scheduling and linkage module is used to map the optimal anomaly handling path to the edge node, generate linkage isolation control sequence based on topology, resource status and priority relationship, and perform path migration, task pruning and residual suppression scheduling between nodes. The scoring function evolution and path stability evaluation module is used to trigger the scoring function structure reconstruction mechanism when the scoring deviation exceeds the threshold, adjust the scoring factor weights and penalty parameters based on the scoring residual trend tensor, and solidify the alternative path as a long-term scheduling candidate path when the scoring trend of the alternative path stabilizes. The feedback and parameter adaptive update module is used to synchronously update the path selection parameters, level discrimination threshold and scoring function structure based on the disturbance residuals, score fluctuations and node state change trends generated during the execution process, so as to realize the periodic reconstruction and evolution optimization of the anomaly handling strategy.

Images