A data analysis method in a virtual power plant

Abstract

The present application relates to the technical field of virtual power plant intelligent scheduling and data analysis, and discloses a data analysis method in a virtual power plant, comprising: identifying an activation section based on a residual window sliding and an event-driven mechanism; constructing a response disturbance index and selecting an interpolation path to reconstruct a high-resolution response curve; analyzing a deviation rate through multi-window time registration and similarity evaluation; and executing nonlinear attribution and outputting a causal chain set in combination with a Bayesian causal graph. Compared with the prior art, under the condition that part of the access equipment has low-frequency sampling or incomplete data, the technical problems that the behavior curve of scheduling cannot be effectively restored and the interpretable causal tracing of the yield fluctuation of the virtual power plant multi-station cannot be performed are solved. Since the dynamic disturbance judgment mechanism and the branch interpolation model are constructed, the residual counterfactual prediction model and the causal chain generation are introduced, the problems of misjudgment and attribution error of abnormal response in the traditional method are avoided, and the scheduling accuracy and yield stability are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent scheduling and data analysis technology for virtual power plants, and in particular to a data analysis method for virtual power plants. Background Technology

[0002] Currently, Virtual Power Plants (VPPs), as intelligent dispatching platforms aggregating distributed power sources, loads, and energy storage devices, are widely used in ancillary service markets, electricity spot markets, and renewable energy consumption scenarios. The operational effectiveness of VPPs highly depends on the accurate perception and analysis of the status data and control execution behavior of each access resource. However, in existing technologies, many edge access devices (such as building air conditioning group control, industrial and commercial load controllers, and electric vehicle charging piles) often have long data upload cycles (e.g., every 15 to 30 minutes) due to communication link bandwidth limitations or insufficient terminal processing capabilities. This results in significant information sparsity at the actual dispatching cycle scale (e.g., 4 seconds to 1 minute), making it difficult to reconstruct a complete power response curve, thus affecting the accurate assessment of the authenticity, effectiveness, and compliance of control behavior. Furthermore, the economic returns during VPP operation are often affected by multiple coupled factors, including real-time electricity market price changes, local critical resource disconnection or response failure, and accumulated task execution deviations. Existing solutions mostly employ static weighting or linear regression methods to assess the revenue contribution of stations, lacking causal identification and explainable attribution mechanisms for abnormal revenue fluctuations. This makes it difficult to support the intelligent implementation of scheduling optimization, accountability, and a closed-loop reward and punishment system. For example, if a sudden decrease or overflow in total revenue occurs within a certain ancillary service settlement cycle, it is difficult to distinguish whether it is caused by market signal disturbances, abnormal equipment behavior, or interruption of scheduling command transmission or execution links. This can lead to delays in adjusting control strategies, increased volatility in economic returns, and even compliance risks.

[0003] Therefore, there is an urgent need for a data analysis method with the ability to reconstruct response behavior and multi-factor nonlinear attribution, so that even under conditions of long sampling period and incomplete control data, the control process can still be recovered with high fidelity and an explainable causal chain for abnormal returns can be output, thereby improving the cognitive ability and control accuracy of virtual power plants in complex environments. Summary of the Invention

[0004] To address the aforementioned technical shortcomings, the present invention aims to propose a data analysis method for virtual power plants, which solves the technical problems in the prior art where some access devices have low-frequency sampling or incomplete data, making it impossible to effectively reconstruct the control behavior curve and to provide an explainable causal attribution for the revenue fluctuations of multiple virtual power plants.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a data analysis method in a virtual power plant.

[0006] The data analysis methods in the virtual power plant include:

[0007] Step S10: Obtain the control command sequence from the virtual power plant dispatching backend, and simultaneously receive the power reporting data sequence from the access device; based on the residual window sliding alignment and event-driven enhanced recognition mechanism, at a preset sampling interval... Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment ;

[0008] Step S20: For the identified active segments Perform the tasks of constructing the average slope index and determining the interpolation path, and output the interpolation results for the active segment; among them, the task of constructing the average slope index outputs the response perturbation index. When the response disturbance index Greater than the preset disturbance index threshold When the model path determination task selects the first interpolation execution path, and the response perturbation index... Less than or equal to the preset disturbance index threshold At that time, the model path determination task selects the second interpolation execution path;

[0009] Step S30: Construct a set of high-resolution response curves based on the interpolation results of the activation segments. And combined with a pre-set standard template library A deviation rate analysis is performed using a multi-window time registration and disturbance response similarity evaluation mechanism, and the corresponding segments with abnormal returns are output based on the analysis results.

[0010] Step S40: Based on the abnormal profit corresponding segment, a nonlinear attribution mechanism combining perturbation propagation fuzzy clustering and residual-driven counterfactual simulation is used to perform factor correlation identification processing, and output nonlinear attribution results;

[0011] Step S50: Construct a Bayesian causal structure graph G based on the nonlinear attribution results, calculate the maximum a posteriori path, and output the reference causal chain set.

[0012] Preferably, in step S10, the control command sequence includes the target power, control rate, and target duration; based on the residual window sliding alignment and event-driven enhanced recognition mechanism, within a preset sampling interval... Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment The steps specifically include:

[0013] For power reporting data sequences Perform sliding window processing, where, Let k be the sampling time point; For access devices at a given time point The reported power value; Let be the total number of sampling points; define the window length. In window length The power fitting value is calculated using the local first-order polynomial least squares method. And based on the power fitting value and reported power values The residual vector is calculated using the absolute residual method. And based on the residual vector Constructing residual sequences As an indicator of the non-stationarity of power series changes;

[0014] Let the control command sequence be ,in, Let be the start time of the j-th control instruction; For each sampling point, the corresponding control parameter vector contains information on the commanded target power, control rate, and target duration. Define event label function , ;in, The "pre-triggered window" indicates the time tolerance at which control commands may affect device behavior in advance. The "later-triggered window" indicates the tolerance time required for the device to respond to control lag.

[0015] Set residual threshold , when the residual sequence The residual term is greater than or equal to the residual threshold. and When 1, mark the corresponding segment as the active segment; when the residual sequence The residual term is less than the residual threshold. and When the time interval is 0, the corresponding segment is marked as a stationary segment; the final output is the stationary segment marked as a set of stationary time intervals. and the output is marked as the activation segment of the activation time interval set. .

[0016] Preferably, in step S20, the average slope index is used to construct the task output response perturbation index. The specific steps include: constructing a set of power change slopes for the active region; calculating the mean and variance of the set of power change slopes; and defining the response perturbation index based on the mean and variance of the set of power change slopes. This is a normalized estimate of the slope change amplitude per unit time; the first interpolation execution path includes directly reconstructing the missing points in the active segment based on the linear interpolation strategy, and the output active segment interpolation result is a low dynamic steady-state interpolation result, which is used for subsequent response quality assessment and reward / penalty feedback generation.

[0017] Preferably, in step S20, the second interpolation execution path includes three stages:

[0018] Modal decomposition stage: Empirical mode decomposition is performed on the original power curve in the activation segment to decompose the original power curve into several intrinsic mode components with different frequency distribution characteristics; all intrinsic mode components are linearly combined according to a preset weighting factor to obtain an initial interpolation response curve with dynamic feature preservation capability, which serves as the input reference for subsequent nonlinear modeling.

[0019] Nonlinear modeling stage: The initial interpolation response curve obtained from mode decomposition is used as time series input data and input into a recurrent neural network structure based on a gating mechanism; the short-term dynamic change trend of the power sequence is modeled and predicted by the recurrent neural network structure, and a fitted interpolation response curve is output to characterize the fast response behavior features in the activation segment.

[0020] Error rollback phase: Determine the fitting error between the interpolated response curve output from the nonlinear modeling phase and the initial interpolated response curve from the mode decomposition phase; if the error exceeds a preset threshold, it is considered that there is an atypical perturbation mode in the current active segment, and the response interpolation curve is reverted to be constructed using a physical modeling strategy based on an exponential decay function.

[0021] Preferably, in step S30, a set of high-resolution response curves is constructed based on the interpolation results of the activated segments. And combined with a pre-set standard template library The steps for performing deviation rate analysis using a multi-window time registration and disturbance response similarity evaluation mechanism, and outputting the corresponding segments of abnormal returns based on the analysis results, specifically include:

[0022] Based on the interpolation results of the activation segment, multiple time-aligned windows with different proportions and positions are constructed on the time axis to capture the behavioral characteristics of the device at different stages of response startup, reaching the maximum deviation, and recovering to steady state.

[0023] Feature points characterizing disturbance behavior are extracted within each time alignment window, including the time point of the initial response, the time point corresponding to the maximum disturbance amplitude, and the steady-state recovery time point; these feature points are then calculated relative to a pre-defined standard template library. The time difference and amplitude difference are used as quantitative indicators to describe the disturbance behavior.

[0024] The deviation rate index is calculated based on the quantitative description index of the disturbance behavior. The segments where the deviation rate index exceeds the deviation rate index threshold are marked and output as the corresponding segments of abnormal returns.

[0025] Preferably, in step S30, the step of extracting feature points representing perturbation behavior in each time alignment window is based on a perturbation feature recognition method that combines piecewise derivative mutation detection and dynamic threshold adjustment, and includes, in sequence, a piecewise derivative mutation detection step, a dynamic threshold adjustment mechanism step, a perturbation behavior key point localization step, and a robustness verification mechanism step.

[0026] Preferably, step S40, which involves performing factor correlation identification processing based on a nonlinear attribution mechanism combining perturbation-propagation fuzzy clustering and residual-driven counterfactual simulation for the corresponding segments of abnormal returns, and outputting the nonlinear attribution result, specifically includes:

[0027] Step S401: Disturbance Propagation Fuzzy Clustering Stage: Based on the response behavior of multiple stations involved in the abnormal revenue section, the three-element index of response time difference, amplitude coordination degree and load coupling degree between different resources is defined as the input of fuzzy membership function. Fuzzy C-means clustering analysis is performed to extract the resource subset with causal relationship.

[0028] Step S402: Residual-driven counterfactual simulation stage: For each resource subset in the causal relationship, a residual-enhanced recurrent network modeler based on time-gated units is introduced. The current interpolated response data under each resource subset and the preset market price sequence are used as inputs to output the theoretical profit prediction value; the actual profit data is obtained, and the residual between the theoretical profit prediction value and the actual profit data is calculated to output the residual index value of all resource subsets.

[0029] Step S403: Nonlinear contribution decomposition stage: The recursive feature ablation method is used to gradually eliminate the resource subsets with the greatest impact on abnormal returns for all resource subset residual index values, and finally outputs the nonlinear attribution results.

[0030] The present invention also provides a data analysis system for a virtual power plant, comprising:

[0031] The instruction recognition module is used to obtain the control instruction sequence from the virtual power plant dispatching backend, and simultaneously receive the power reporting data sequence from the access device; based on the residual window sliding alignment and event-driven enhanced recognition mechanism, it operates at a preset sampling interval. Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment ;

[0032] The interpolation path determination module is used to determine the identified active segments. Perform the tasks of constructing the average slope index and determining the interpolation path, and output the interpolation results for the active segment; among them, the task of constructing the average slope index outputs the response perturbation index. When the response disturbance index Greater than the preset disturbance index threshold When the model path determination task selects the first interpolation execution path, and the response perturbation index... Less than or equal to the preset disturbance index threshold At that time, the model path determination task selects the second interpolation execution path;

[0033] The deviation analysis module is used to construct a set of high-resolution response curves based on the interpolation results of the activation segments. And combined with a pre-set standard template library A deviation rate analysis is performed using a multi-window time registration and disturbance response similarity evaluation mechanism, and the corresponding segments with abnormal returns are output based on the analysis results.

[0034] The nonlinear attribution module is used to perform factor correlation identification processing based on the nonlinear attribution mechanism that combines perturbation propagation fuzzy clustering and residual-driven counterfactual simulation based on the corresponding segments of abnormal returns, and outputs nonlinear attribution results.

[0035] The causal reasoning module is used to construct a Bayesian causal structure graph G based on nonlinear attribution results, calculate the maximum posterior path, and output a set of reference causal chains.

[0036] The present invention also provides a data analysis device in a virtual power plant, comprising: a memory, a processor, and a data analysis program in a virtual power plant stored in the memory and executable on the processor, wherein the data analysis program in a virtual power plant implements a data analysis method in a virtual power plant when executed by the processor.

[0037] The present invention also provides a computer program product, including a data analysis program in a virtual power plant, wherein the data analysis program in the virtual power plant implements the data analysis method in the virtual power plant when executed by a processor.

[0038] The beneficial effects of this invention are as follows: Compared with the technical problems in the prior art where some access devices have low-frequency sampling or incomplete data, making it impossible to effectively restore the control behavior curve and to explain the causal causes of revenue fluctuations in multiple virtual power plants, this invention avoids the problems of misjudgment and attribution errors of traditional methods for abnormal responses by constructing a dynamic disturbance judgment mechanism and a branch interpolation model, introducing a residual counterfactual prediction model and causal chain generation, thereby improving scheduling accuracy and revenue stability. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the first embodiment of a data analysis method in a virtual power plant according to the present invention.

[0041] Figure 2 This is a schematic diagram of the equipment used in a data analysis method for a virtual power plant according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the data analysis method in the virtual power plant of the present invention, which presents the first embodiment of the data analysis method in the virtual power plant of the present invention.

[0044] In the first embodiment, the data analysis method in the virtual power plant includes:

[0045] Step S10: Obtain the control command sequence from the virtual power plant dispatching backend, and simultaneously receive the power reporting data sequence from the access device; based on the residual window sliding alignment and event-driven enhanced recognition mechanism, at a preset sampling interval... Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment ;

[0046] It should be noted that the "residual window sliding alignment mechanism" refers to calculating the dynamic residual value (such as mean absolute residual, standard deviation, or quadratic residual) between the power reported data of the access device and the control target power within a given sliding time window, and combining it with the fitting slope trend information within the window to evaluate the strength of the response behavior. This mechanism specifically includes three parts: (1) dynamic monitoring of the mean residual value within the window; (2) power slope mutation detection based on the fitting curve; and (3) assignment of local response labels triggered by the residual threshold. The "event-driven enhanced recognition mechanism" refers to automatically activating a set of recognition processes with slope change rate, variance mutation point, and number of high-frequency change segments as triggering factors when a control command is detected (such as load aggregation scheduling or automatically generated frequency control commands), thereby enhancing the sensitivity of capturing the activated response area. This mechanism is particularly suitable for assisting in the identification of the initial response time point and the end time point under operating conditions with long data upload cycles and unclear response boundaries.

[0047] Understandably, the dual-mechanism joint judgment method proposed in this step can effectively divide the power sequence into "response segment" (active segment) and "steady-state segment" (stable segment) when the device data upload cycle cannot cover the control cycle, and provide high-confidence data segment labels for subsequent interpolation modeling and deviation assessment. This division not only improves the local fitting quality of the reconstructed curve, but also significantly reduces the interference effect introduced by the non-response segment.

[0048] It should be understood that traditional solutions typically determine whether a response segment has been entered by setting fixed slope thresholds, power change thresholds, or moving average differences. While this works reasonably well under high-frequency sampling conditions, it is prone to misjudgment in low-frequency reporting scenarios of 15-30 minutes, leading to missed detections or false alarms in response determination. The residual and slope combined sliding window method introduced in this invention can more precisely characterize "whether there is an effective response intention relative to the control curve," and the event-driven enhancement mechanism triggers the recognition process at the control point, further improving the boundary accuracy and completeness of activation segment extraction, and possessing higher adaptability and accuracy.

[0049] For example, in a frequency regulation task, the control objective is to rapidly increase the power from 0kW to 100kW and maintain it for 5 minutes. However, a building's air conditioning controller only uploads power data at 10:00 and 10:15. Traditional solutions can only determine that the device is still at 0kW at 10:00 and has reached 90kW at 10:15, but cannot identify when its response curve begins to change. Using the residual window mechanism of this invention (window width Δt = 2 minutes, step size = 30 seconds), a significant decrease in the mean power residual and a sudden increase in the fitting slope are detected between 10:02 and 10:06, which aligns with the time of the control signal. Therefore, 10:02 is determined as the response starting point and 10:07 as the stable response interval. This segment is then used as the activation segment and accurately incorporated into subsequent interpolation modeling, improving the goodness of fit of the reconstructed response from 0.68 to 0.91, and reducing the deviation error by nearly 34%.

[0050] Step S20: For the identified active segments Perform the tasks of constructing the average slope index and determining the interpolation path, and output the interpolation results for the active segment; among them, the task of constructing the average slope index outputs the response perturbation index. When the response disturbance index Greater than the preset disturbance index threshold When the model path determination task selects the first interpolation execution path, and the response perturbation index... Less than or equal to the preset disturbance index threshold At that time, the model path determination task selects the second interpolation execution path;

[0051] It should be noted that the "average slope index construction task" refers to constructing a continuous difference slope sequence within each response segment based on the active segments identified in step S10, and calculating the statistical expectation (i.e., average slope) of this sequence and the deviation from the maximum gradient slope, defining a combined function as the response perturbation index. This index is used to measure the drastic change in response within the active segment, reflecting the nonlinear strength and temporal instability tendency of the device's response to the control task. The "interpolation path determination task" refers to automatically selecting different interpolation modeling paths to generate a complete response curve based on the magnitude of the response perturbation index value relative to a preset threshold. When the response perturbation index is large, it indicates that the response has obvious nonlinearity and abrupt changes, and the "first interpolation execution path" is selected, using a combination of mode decomposition and neural network nonlinear modeling for fitting; when the response perturbation index is small, it indicates that the response is relatively stable, and the "second interpolation execution path" is selected, using trend piecewise linear fitting + error compensation correction as the main methods to construct the interpolation curve.

[0052] Understandably, by introducing the response perturbation index as the basis for interpolation path decision-making, an adaptive dynamic interpolation path selection mechanism is achieved, significantly improving the adaptability and accuracy of the interpolation algorithm in complex response behavior scenarios. This mechanism avoids the overfitting or underfitting problem of the unified modeling path in all response scenarios, resulting in stronger structural consistency and prediction accuracy of the reconstructed response curve between local fluctuation segments and stable sustained segments.

[0053] It should be understood that, compared to traditional methods that often use statically selected interpolation models (such as cubic spline interpolation, linear interpolation, or ARIMA models) to perform unified reconstruction, this invention dynamically constructs a response perturbation index and divides the interpolation path based on this index, enabling the interpolation process to have response intensity sensing capabilities. This mechanism not only improves interpolation accuracy but also enhances the ability to detect abnormal responses, exhibiting outstanding robustness and generalization, especially in scenarios with blurred response boundaries or extremely sparse sampling.

[0054] Step S30: Construct a set of high-resolution response curves based on the interpolation results of the activation segments. And combined with a pre-set standard template library A deviation rate analysis is performed using a multi-window time registration and disturbance response similarity evaluation mechanism, and the corresponding segments with abnormal returns are output based on the analysis results.

[0055] It should be noted that the "high-resolution response curve set" in this step refers to the response power time series matrix constructed based on the interpolation results of the activation segment output from step S20, for all resources with matching control commands, using isochronous sampling (e.g., 1-minute intervals) to reconstruct the control behavior of each resource at a fine-grained time scale. The "preset standard template library" refers to a pre-constructed set of response pattern samples, containing standard response forms of typical stations under different control task types (e.g., frequency regulation, load reduction, emergency start / stop), including indicators such as standard power trajectory, rise / fall rate, duration, and delay interval, used as a benchmark for behavior alignment. The "multi-window time registration and disturbance response similarity evaluation mechanism" involves dividing the high-resolution response curve into overlapping time windows (e.g., 3-minute sliding window, 6-minute sliding window, 9-minute fixed window, etc.), aligning the actual response with the template using the Dynamic Time Warping (DTW) algorithm within each window, and introducing three joint similarity indicators: disturbance start alignment factor, response amplitude deviation rate, and fluctuation frequency distance, to construct a comprehensive deviation rate function to measure the degree of behavioral deviation of the response trajectory.

[0056] It should be understood that this step significantly enhances the stability and resolution of response behavior deviation analysis through the fusion of "time-scale sliding window and multi-dimensional behavioral similarity indicators." Compared to the traditional full-segment Euclidean distance matching method, this method can dynamically align scenarios with misaligned response starting points, identify short-term abrupt change segments or local abnormal tail segments, thereby achieving accurate behavioral deviation rate identification on both the time axis and the response intensity axis. By setting a deviation rate threshold, time segments with significant deviations can be marked as "abnormal return response segments," providing precise input boundaries for subsequent causal attribution and economic analysis.

[0057] Step S40: Based on the abnormal profit corresponding segment, a nonlinear attribution mechanism combining perturbation propagation fuzzy clustering and residual-driven counterfactual simulation is used to perform factor correlation identification processing, and output nonlinear attribution results;

[0058] It should be noted that the "perturbation propagation fuzzy clustering" mentioned in this step refers to constructing perturbation collaborative behavior indicators for multiple resources within the abnormal return response segment identified in step S30, and embedding them as ternary input variables into the fuzzy C-means (FCM) algorithm to automatically segment resource subsets with highly correlated response behaviors, thus initially identifying potential influencing factor groups. "Residual-driven counterfactual simulation" refers to constructing a prediction network based on a gated recurrent unit (GRU) nested temporal embedding mechanism for each resource within the identified resource subset. The interpolated response sequence of the resource and the market price signal sequence are input, and the theoretical return prediction value is output. The prediction value is compared with the actual settlement return, and the residual is calculated. A high residual indicates that the resource contributes more to the abnormal return under the current strategy. Finally, based on the residual indicators of all resource subsets, a recursive feature ablation method is used to gradually eliminate high-contribution resources, iterating repeatedly until the residual error no longer decreases significantly, and the retained set is output as the nonlinear attribution result.

[0059] Understandably, this step constructs a nonlinear attribution framework that integrates fuzzy clustering and dynamic time prediction, which has the following technical effects: On the one hand, it determines the coupling relationship between resources in the fuzzy space through three-dimensional behavioral synergy indicators, making up for the misidentification problem of traditional linear correlation coefficients or static threshold methods; on the other hand, it introduces counterfactual simulation to construct a theoretical benefit scenario of "if this resource is not available, the response remains unchanged", which is used to perform "hypothesis-reality" comparative analysis on the actual benefit error, making the attribution results verifiable and explanatory.

[0060] It should be understood that, compared to common linear regression attribution or single-point residual-based attribution methods in existing technologies, this invention introduces a two-stage mechanism of perturbation propagation clustering and temporal counterfactual simulation, which has stronger nonlinear adaptability and attribution signal-to-noise ratio improvement capabilities in the identification process. Especially in virtual power plant scenarios involving multiple resources, overlapping behaviors, and uncertain control chains, this method can effectively reduce the risk of false attributions and missed attributions, and improve the transparency and operability of rewards and penalties.

[0061] For example, in a certain auxiliary frequency modulation task, the total revenue decreased by 17.6%. Step S30 identified the abnormal revenue segment as minutes 13-16. In S40, fuzzy clustering grouped resources A, B, and C into the same perturbation subset, with highly consistent response starting points and coordinated curve shapes. Subsequently, counterfactual simulation revealed that the theoretical revenue prediction value for resource B was 138 yuan, while the actual revenue was 112 yuan, with a residual of -26 yuan (the largest among all resources). Finally, RFE removed resources A and C, retaining only resource B as the main cause of the revenue anomaly. This resource was then marked as an abnormal unit by the scheduler and a policy adjustment was performed. Experimental statistics show that in multiple real-world scenarios, this attribution method improves accuracy by more than 12% compared to the traditional linear weight method, significantly enhancing decision-making accuracy.

[0062] Step S50: Construct a Bayesian causal structure graph G based on the nonlinear attribution results, calculate the maximum a posteriori path, and output the reference causal chain set.

[0063] It should be noted that the "Bayesian causal structure graph G" refers to a directed acyclic graph structure used in this invention to model the potential causal relationships between various resource subsets. This graph uses the anomalous resource subsets identified in the nonlinear attribution stage (step S40) as graph nodes, and uses the disturbance synergy, response order, and residual change trend between subsets as the basis for graph construction to establish causal paths between resources. Specifically, the construction of edge weights does not rely on a single linear indicator, but comprehensively considers the following three types of information: the order of response time (i.e., which resource subset's behavior occurs first); the synergy of residual trends (e.g., whether they rise or fall together); and the upstream and downstream logic in the control instruction chain (e.g., a clear leading relationship exists in load relay response). The Bayesian graph constructed using the above information can more realistically express the possible paths of disturbance propagation among multiple resources.

[0064] It should be understood that, unlike traditional schemes that often employ static weighted factors or simple regression analysis, the Bayesian causal structure diagram constructed in this invention possesses dynamic modeling capabilities and the ability to express uncertainty. It is particularly suitable for the typical scenario in the current power sector where resource response behavior is highly heterogeneous and control chains are complex and intersecting. Traditional methods struggle to handle issues such as causal ambiguity, overlapping effects, or temporal coupling among resources, easily leading to "misattribution of responsibility" or "ineffective intervention." In contrast, this invention, by introducing a Bayesian posterior inference process and automatically calculating the maximum posterior probability path, can identify the most reliable causal propagation chain, avoiding the problem of subjectively setting paths in causal mapping, and improving the scientific rigor and operability of diagnosis and intervention.

[0065] For example, taking a scenario of "frequency modulation response + abnormal revenue" as an example, suppose a sudden drop in scheduling revenue is detected between 11:15 and 11:45, and three abnormal resource subsets are identified during this period: User A, User B, and User C. Among them, User A's response behavior is ahead of time, and its residual volatility is the highest; User B's response change follows closely; User C's power output is abnormally delayed but with a large amplitude. After constructing a Bayesian causal structure graph, the maximum posterior path "User A → User B → User C" is obtained through inference and output as a set of reference causal chains. This chain path reveals that this revenue anomaly is not caused by a single site, but by User A's inaccurate response causing B's collaborative deviation, which ultimately leads to C's revenue imbalance. Experimental results show that in 20 real revenue anomaly scenarios, this causal graph construction algorithm improves the root cause matching rate by 27% compared to traditional weighted average or residual sorting methods, and correctly identifies the causal propagation chain in 14 scenarios, demonstrating higher causal reliability and diagnostic accuracy.

[0066] Example 2: Furthermore, the data analysis system for a virtual power plant provided by the present invention employs a data analysis method for a virtual power plant as described in the above embodiments, and can solve the technical problem of data analysis in a virtual power plant. Compared with the prior art, the beneficial effects of the data analysis system for a virtual power plant provided by the present invention are the same as the beneficial effects of the data analysis method for a virtual power plant provided in the above embodiments, and other technical features of the data analysis system for a virtual power plant are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0067] Example 3: This invention provides a data analysis device for a virtual power plant. Please refer to... Figure 2A data analysis device in a virtual power plant includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a data analysis method in a virtual power plant as described in Embodiment 1 above. The data analysis device in a virtual power plant according to embodiments of the present invention may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. The data analysis device in a virtual power plant is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention. The data analysis device in a virtual power plant may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. Random access memory 1004 also stores various programs and data required for the operation of a data analysis device in a virtual power plant. Processing unit 1001, read-only memory 1002, and random access memory 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the data analysis device in a virtual power plant to communicate wirelessly or wiredly with other devices to exchange data. Although a data analysis device in a virtual power plant with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented or possessed alternatively.

[0068] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the data analysis method in a virtual power plant as described above. The computer program product provided by this invention can solve the technical problem of data analysis in a virtual power plant. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as the beneficial effects of the data analysis method in a virtual power plant provided in the above embodiments, and will not be repeated here.

[0069] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0070] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0071] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A data analysis method in a virtual power plant, characterized in that, The methods include: Step S10: Obtain the control command sequence from the virtual power plant dispatching backend, and simultaneously receive the power reporting data sequence from the access device; based on the residual window sliding alignment and event-driven enhanced recognition mechanism, at a preset sampling interval... Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment ; Step S20: For the identified active segments Perform the tasks of constructing the average slope index and determining the interpolation path, and output the interpolation results for the active segment; among them, the task of constructing the average slope index outputs the response perturbation index. When the response disturbance index Greater than the preset disturbance index threshold When the model path determination task selects the first interpolation execution path, and the response perturbation index... Less than or equal to the preset disturbance index threshold At that time, the model path determination task selects the second interpolation execution path; Among them, the average slope index constructs the task output response perturbation index. The specific steps include: constructing a set of power change slopes for the active region; calculating the mean and variance of the set of power change slopes; and defining the response perturbation index based on the mean and variance of the set of power change slopes. This is a normalized estimate of the slope change amplitude per unit time; the first interpolation execution path includes directly reconstructing the missing points in the active segment based on the linear interpolation strategy, and the output active segment interpolation result is a low dynamic steady-state interpolation result, which is used for subsequent response quality assessment and reward / penalty feedback generation; The second interpolation execution path includes three stages: Mode decomposition stage: Empirical mode decomposition is performed on the original power curve within the activation segment, decomposing the original power curve into several intrinsic mode components with different frequency distribution characteristics; all intrinsic mode components are linearly combined according to a preset weighting factor to obtain an initial interpolation response curve with dynamic feature preservation capability, which serves as the input reference for subsequent nonlinear modeling; Nonlinear modeling stage: the initial interpolation response curve obtained from mode decomposition is used as time series input data and input into a recurrent neural network structure based on a gating mechanism; the short-term dynamic change trend of the power series is modeled and predicted through the recurrent neural network structure, and a fitted interpolation response curve is output to characterize the fast response behavior characteristics within the activation segment; Error backoff stage: the fitting error between the interpolation response curve output by the nonlinear modeling stage and the initial interpolation response curve of the mode decomposition stage is judged; if the error exceeds a preset threshold, it is considered that there is an atypical disturbance mode in the current activation segment, and the response interpolation curve is constructed by using a physical modeling strategy based on an exponential decay function; Step S30: Construct a set of high-resolution response curves based on the interpolation results of the activation segments. And combined with a pre-set standard template library A deviation rate analysis is performed using a multi-window time registration and disturbance response similarity evaluation mechanism, and the corresponding segments with abnormal returns are output based on the analysis results. Step S40: Based on the abnormal profit corresponding segment, a nonlinear attribution mechanism combining perturbation propagation fuzzy clustering and residual-driven counterfactual simulation is used to perform factor correlation identification processing, and output nonlinear attribution results; The steps involved in performing factor correlation identification processing based on a nonlinear attribution mechanism combining perturbation-propagation fuzzy clustering and residual-driven counterfactual simulation for the corresponding segments of abnormal returns, and outputting nonlinear attribution results, specifically include: Disturbance propagation fuzzy clustering stage: Based on the response behavior of multiple stations involved in the abnormal revenue segment, the three-element index of response time difference, amplitude coordination degree and load coupling degree between different resources is defined as the input of fuzzy membership function, and fuzzy C-means clustering analysis is performed to extract the resource subset with causal relationship; Residual-driven counterfactual simulation stage: For each resource subset in the causal relationship, a residual-enhanced recurrent network modeler based on time-gated units is introduced. The current interpolation response data under each resource subset and the preset market price sequence are used as inputs to output the theoretical profit prediction value; the actual profit data is obtained, and the residual between the theoretical profit prediction value and the actual profit data is calculated to output the residual index value of all resource subsets. Nonlinear contribution decomposition stage: The recursive feature ablation method is used to gradually eliminate the resource subsets with the greatest impact on abnormal returns for all resource subset residual index values, and finally outputs the nonlinear attribution results. Step S50: Construct a Bayesian causal structure graph G based on the nonlinear attribution results, calculate the maximum a posteriori path, and output the reference causal chain set.

2. The data analysis method in a virtual power plant as described in claim 1, characterized in that, In step S10, the control command sequence includes the target power, control rate, and target duration; Based on residual window sliding alignment and event-driven enhanced recognition mechanism at a preset sampling interval Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment The steps specifically include: For power reporting data sequences Perform sliding window processing, where, This refers to the k-th sampling time point; For access devices at a given time point The reported power value; Define the total number of sampling points; define the window length. In window length The power fitting value is calculated using the local first-order polynomial least squares method. And based on the power fitting value and reported power values The residual vector is calculated using the absolute residual method. And based on the residual vector Constructing residual sequences As an indicator of the non-stationarity of power series changes; Let the control command sequence be ,in, Let j be the start time of the control command; The corresponding control parameter vector contains information on the commanded target power, control rate, and target duration; at each sampling point Define event label function , ;in, The "pre-triggered window" indicates the time tolerance at which control commands may affect device behavior in advance. The "later-triggered window" indicates the tolerance time required for the device to respond to control lag. Set residual threshold , when the residual sequence The residual term is greater than or equal to the residual threshold. and When 1, mark the corresponding segment as the active segment; when the residual sequence The residual term is less than the residual threshold. and At time 0, the corresponding segment is marked as a stationary segment; the final output is the stationary segment marked as a set of stationary time intervals. and the output is marked as the activation segment of the activation time interval set. .

3. The data analysis method in a virtual power plant as described in claim 1, characterized in that, In step S30, a set of high-resolution response curves is constructed based on the interpolation results of the activation segments. And combined with a pre-set standard template library The steps for performing deviation rate analysis using a multi-window time registration and disturbance response similarity evaluation mechanism, and outputting the corresponding segments of abnormal returns based on the analysis results, specifically include: Based on the interpolation results of the activation segment, multiple time-aligned windows with different proportions and positions are constructed on the time axis to capture the behavioral characteristics of the device at different stages of response startup, reaching the maximum deviation, and recovering to steady state. Feature points characterizing disturbance behavior are extracted within each time alignment window, including the time point of the initial response, the time point corresponding to the maximum disturbance amplitude, and the steady-state recovery time point; these feature points are then calculated relative to a pre-defined standard template library. The time difference and amplitude difference are used as quantitative indicators to describe the disturbance behavior. The deviation rate index is calculated based on the quantitative description index of the disturbance behavior. The segments where the deviation rate index exceeds the deviation rate index threshold are marked and output as the corresponding segments of abnormal returns.

4. The data analysis method in a virtual power plant as described in claim 3, characterized in that, In step S30, the step of extracting feature points representing perturbation behavior in each time alignment window is based on a perturbation feature recognition method that combines piecewise derivative mutation detection and dynamic threshold adjustment. It includes, in sequence, a piecewise derivative mutation detection step, a dynamic threshold adjustment mechanism step, a perturbation behavior key point localization step, and a robustness verification mechanism step.

5. A data analysis system for a virtual power plant, applied to the data analysis method for a virtual power plant as described in any one of claims 1 to 4, characterized in that, The data analysis system in the virtual power plant includes: The instruction recognition module is used to obtain the control instruction sequence from the virtual power plant dispatching backend, and simultaneously receive the power reporting data sequence from the access device; based on the residual window sliding alignment and event-driven enhanced recognition mechanism, it operates at a preset sampling interval. Determine whether the power reporting data sequence covers the control command sequence, and divide the power reporting data sequence into stable sections based on the determination result. and activation segment ; The interpolation path determination module is used to determine the identified active segments. Perform the tasks of constructing the average slope index and determining the interpolation path, and output the interpolation results for the active segment; among them, the task of constructing the average slope index outputs the response perturbation index. When the response disturbance index Greater than the preset disturbance index threshold When the model path determination task selects the first interpolation execution path, and the response perturbation index... Less than or equal to the preset disturbance index threshold At that time, the model path determination task selects the second interpolation execution path; Among them, the average slope index constructs the task output response perturbation index. The specific steps include: constructing a set of power change slopes for the active region; calculating the mean and variance of the set of power change slopes; and defining the response perturbation index based on the mean and variance of the set of power change slopes. This is a normalized estimate of the slope change amplitude per unit time; the first interpolation execution path includes directly reconstructing the missing points in the active segment based on the linear interpolation strategy, and the output active segment interpolation result is a low dynamic steady-state interpolation result, which is used for subsequent response quality assessment and reward / penalty feedback generation; The second interpolation execution path includes three stages: Mode decomposition stage: Empirical mode decomposition is performed on the original power curve within the activation segment, decomposing the original power curve into several intrinsic mode components with different frequency distribution characteristics; all intrinsic mode components are linearly combined according to a preset weighting factor to obtain an initial interpolation response curve with dynamic feature preservation capability, which serves as the input reference for subsequent nonlinear modeling; Nonlinear modeling stage: the initial interpolation response curve obtained from mode decomposition is used as time series input data and input into a recurrent neural network structure based on a gating mechanism; the short-term dynamic change trend of the power series is modeled and predicted through the recurrent neural network structure, and a fitted interpolation response curve is output to characterize the fast response behavior characteristics within the activation segment; Error backoff stage: the fitting error between the interpolation response curve output by the nonlinear modeling stage and the initial interpolation response curve of the mode decomposition stage is judged; if the error exceeds a preset threshold, it is considered that there is an atypical disturbance mode in the current activation segment, and the response interpolation curve is constructed by using a physical modeling strategy based on an exponential decay function; The deviation analysis module is used to construct a set of high-resolution response curves based on the interpolation results of the activation segments. And combined with a pre-set standard template library A deviation rate analysis is performed using a multi-window time registration and disturbance response similarity evaluation mechanism, and the corresponding segments with abnormal returns are output based on the analysis results. The nonlinear attribution module is used to perform factor correlation identification processing based on the nonlinear attribution mechanism that combines perturbation propagation fuzzy clustering and residual-driven counterfactual simulation based on the corresponding segments of abnormal returns, and outputs nonlinear attribution results. The steps involved in performing factor correlation identification processing based on a nonlinear attribution mechanism combining perturbation-propagation fuzzy clustering and residual-driven counterfactual simulation for the corresponding segments of abnormal returns, and outputting nonlinear attribution results, specifically include: Disturbance propagation fuzzy clustering stage: Based on the response behavior of multiple stations involved in the abnormal revenue segment, the three-element index of response time difference, amplitude coordination degree and load coupling degree between different resources is defined as the input of fuzzy membership function, and fuzzy C-means clustering analysis is performed to extract the resource subset with causal relationship; Residual-driven counterfactual simulation stage: For each resource subset in the causal relationship, a residual-enhanced recurrent network modeler based on time-gated units is introduced. The current interpolation response data under each resource subset and the preset market price sequence are used as inputs to output the theoretical profit prediction value; the actual profit data is obtained, and the residual between the theoretical profit prediction value and the actual profit data is calculated to output the residual index value of all resource subsets. Nonlinear contribution decomposition stage: The recursive feature ablation method is used to gradually eliminate the resource subsets with the greatest impact on abnormal returns for all resource subset residual index values, and finally outputs the nonlinear attribution results. The causal reasoning module is used to construct a Bayesian causal structure graph G based on nonlinear attribution results, calculate the maximum posterior path, and output a set of reference causal chains.

6. A data analysis device for a virtual power plant, characterized in that, The data analysis device in the virtual power plant includes: a memory, a processor, and a data analysis program for the virtual power plant stored in the memory and executable on the processor. When the data analysis program for the virtual power plant is executed by the processor, it implements a data analysis method for the virtual power plant according to any one of claims 1 to 4.

7. A computer program product, characterized in that, The computer program product includes a data analysis program in a virtual power plant, which, when executed by a processor, implements a data analysis method in a virtual power plant according to any one of claims 1 to 4.

Images