Virtual power plant optimization management method based on conditional value-at-risk

Abstract

The invention discloses a virtual power plant optimization management method based on conditional value at risk, relates to the technical field of virtual power plant scheduling, and sequentially executes multi-mode safety pointer perception, risk quantile scene generation, three-layer CVaR coupling optimization, edge cloud double-threshold instruction issuing and real-time loss driven model adaptive updating. According to the method, instruction, market information and power data are fused in a millisecond level, distortion is marked, then a quantile GAN is used for generating a safety-power-price coupling scene, and then a unified CVaR target mixed integer second-order cone is constructed to obtain energy storage, load and bidding instructions. The method comprises the following steps: checking in a tide-market sandbox before issuing an instruction, sampling real loss by a sliding window after execution, and synchronously updating generator parameters and a CVaR confidence radius by using natural gradient and Bayesian conjugate to form a closed loop, thereby improving the safety and robust income of a virtual power plant in an extreme market.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of virtual power plant scheduling, in particular to a virtual power plant optimization management method based on conditional risk value. BACKGROUND

[0002] Under the background of rapid evolution of new power systems, distributed photovoltaic, wind power, energy storage and flexible load are massively connected to distribution networks, and three types of markets, namely power spot, auxiliary services and carbon emission quota, are iteratively matched at a frequency of seconds. As an aggregation subject, the virtual power plant needs to complete multi-objective scheduling optimization in the cloud and issue power instructions, load reduction and bidding curves to thousands of heterogeneous devices through edge controllers.

[0003] The general technology mainly adopts the mode of "static scene prediction + staged optimization": first, use offline methods to arrange a small number of typical scenes in advance; then, optimize according to the established risk preference; finally, broadcast the entire generation plan for the day in a single instruction package. Under the background of high penetration of renewable energy, high digitalization of terminal side and network security attack means, the significant defects are: ① the offline scene cannot predict extreme market fluctuations and cross-domain instantaneous distortion under instant occurrence; ② it cannot take into account the coupling loss of the 3D tail end of scheduling, and cannot calculate the error and price income while completing the safety risk; ③ the optimal solution in the cloud cannot be quickly traced back, and there is no timely feedback to the closed-loop optimization correction information, so it lacks model adaptive optimization capability. If the on-site situation does not meet the data requirements of the ground scene, it will bring errors in planning and execution, and further enlarge the gap in the final settlement period.

[0004] Under the scenario of high penetration of renewable energy virtual power plant, the tail end risks of safety layer, physical layer and financial layer caused by network delay injection or market tampering are discovered in milliseconds, and based on this, scheduling decision closed-loop input is generated, which needs to inject the above three aspects of tail end risks into the scheduling decision chain in a quantitative, controllable and feedback manner, for example: hackers insert tens of milliseconds of delay in the cloud-edge communication link, or tamper with a batch of key quotes, and the remote cloud scheduler will misjudge the remaining power and price signal of the power grid, and incorrectly issue control commands for excessive discharge and overload cutting. Once executed in the power shortage period, it will cause the "power penalty + transaction loss + safety loss" three-tail-end impact of bus voltage, spot settlement price and safety restart cost, and since there is no real-time sampling of actual loss and automatic adjustment function of model parameters at present, the three impacts will not be reflected in the next period, but will be superimposed on the next period and continuously accumulated and amplified with the extension of the continuous scheduling window, eventually causing the decline of regional power grid stability and the decline of virtual power plant overall credibility.

[0005] Virtual power plant safe operation and power grid safe operation are both facing great challenges, how to complete risk discovery to dispatch execution in the form of closed loop within seconds is a difficult problem. Therefore, how to establish the whole chain of detection, scenario generation, risk quantification, optimization decision, release verification and online update, so that all links can complete the closed loop completely under the rhythm of every second is a difficult problem, and is one of the important problems restricting the commercialization of virtual power plant and the safe operation of power grid. SUMMARY

[0006] (I) The technical problem solved In view of the current technical situation, a virtual power plant optimization management method based on conditional value at risk is invented in this paper, in which multi-modal safety pointer perception, risk quantile scenario generation, three-layer CVaR coupled optimization, edge cloud double threshold instruction release, real-time loss driven model adaptive update are realized in turn, the mixed integer second order cone of CVaR target is obtained, the virtual power plant energy storage, load and bidding instruction are obtained, and the instruction is verified in the tide-market sandbox before release, and the real loss is sampled in the form of sliding window after the instruction is released, the generator parameters and the CVaR confidence radius are updated synchronously by using natural gradient and Bayesian conjugate, and the cycle is repeated, so that the virtual power plant has stronger safety and robust revenue, and more revenue is obtained under the premise of standard operation in extreme market, solving the technical problems existing in the prior art.

[0007] (II) Technical scheme In order to achieve the above purpose, the following technical scheme is adopted: The virtual power plant optimization management method based on conditional value at risk comprises that the edge end synchronously collects instruction flow, market flow and power flow, generates a unified feature tensor by multi-view attention fusion and labels the distortion section, and determines the abnormal intensity according to the historical steady state distribution; The distortion section is input into the improved quantile generative adversarial network to generate a risk scenario set coupled with safety, power and price and attached with synchronous deviation quantile labels for subsequent optimization calling and processing; Taking the risk scenario set as input, a three-layer conditional value at risk coupled global target of safety, power and price is constructed, and energy storage, load and bidding variables are solved to obtain a small tail loss dispatch scheme; The dispatch scheme is subjected to consistency verification on the cloud-edge end, and the qualified device instruction set is generated and released after the local tide-market sandbox predicts the power and price deviation, and the release threshold is lower; Real-time sampling of execution loss and formation of observation matrix, updating the generation network parameters by using natural gradient, and shrinking the confidence radius by using Bayesian method and broadcasting the update package for next period calling and using continue.

[0008] Further, the original timestamps of the instruction flow, the market flow and the power flow are aligned by sliding index interpolation to obtain a unified millisecond-level common time axis. The aligned feature matrix is weighted and learned using a multi-view attention mechanism, and then a weighted feature tensor is generated through linear mapping. The multi-view attention mechanism simultaneously calculates the time sequence self-attention weight and the inter-modal interaction weight.

[0009] A variational Gaussian mixture steady-state model is established by using the weighted feature tensor to obtain a steady-state probability density, an anomaly score and a safety pointer. The continuous safety pointer is smoothed by bidirectional Bayesian filtering to generate a section-level anomaly label. The anomaly label is used as an input index for risk scenario generation.

[0010] Then, the quantile label is obtained by linear stretching and inverse cumulative distribution mapping of the safety pointer, and a scenario index matrix is established according to the quantile label. The scenario index and a random noise vector are combined as a generator input to drive the quantile generative adversarial network to output multiple scenario samples.

[0011] First, it is further clarified that the training target of the quantile generative adversarial network is coupled by the adversarial cross-entropy loss and the first-order Wasserstein distance of the quantile, and a balance coefficient is introduced to adjust the weight size of the two; second, the generated samples and the real samples are in the same quantile, and the gradient optimization between them can still be realized.

[0012] Further, the safety loss, the power loss and the price loss scalars are established respectively. The three types of losses are used to construct the conditional value at risk formula with shear variables and tail excess expectations; the three types of risk values are weighted to form the overall objective function.

[0013] Further, the overall objective function, the energy storage power vector, the load adjustment vector and the bidding adjustment vector jointly constitute a mixed integer second-order cone programming model, and the energy storage energy boundary, the bus power boundary and the market step declaration second-order cone constraint are set to solve the dispatching vector with the minimum comprehensive tail loss.

[0014] Further, the channel correction of the cloud energy storage power plan is performed using the communication delay mapping matrix, and the two-norm error is calculated with the edge power feedback of the previous period. The energy storage error, the load error and the bidding error are linearly combined into a consistency index according to the preset weight, and compared with the consistency threshold to determine the release of the unpacked instruction.

[0015] Further, the unpacked instruction is coupled and sandboxed in the power flow Jacobi fast iteration model and the local supply-demand price elasticity model at the edge end to calculate the power-price comprehensive deviation and the voltage deviation; the two are combined into a comprehensive risk index according to the weighting coefficient, and compared with the release threshold to determine the immediate release or repair of the instruction.

[0016] Furthermore, a sliding window stratified sampling mechanism is used to extract samples in each loss category according to power weighting to form a loss subset; The contents of different subsets are spliced ​​together to obtain the loss observation matrix, which is then used to calculate the real-time empirical quantile and category entropy for subsequent incremental model updates.

[0017] While further using the loss observation matrix to realize the natural gradient update of the generator parameters, the gamma conjugate Bayes method is used to complete the update of the posterior shape and scale parameters of the confidence radius, and then the generator parameters, confidence radius, and category entropy to be used in the final stage of model training are obtained, forming a broadcast package and sending it to the risk scenario generation and scheduling optimization module for call in the next cycle.

[0018] (3) Beneficial effects The present invention provides a virtual power plant optimization management method based on conditional value at risk, which has the following beneficial effects: The multimodal safety indicator perception and anomaly labeling method constructs a three-line flow field integrating instruction flow, market flow, and power flow. It establishes a mechanism for mapping safety indicators at the same scale for network distortion, market jump, and power deviation, and can accurately locate risk fragments in milliseconds. Based on this, subsequent links can obtain high-confidence risk labels without manually setting thresholds, greatly improving the sensitivity to tail-end events.

[0019] An improved quantile generative adversarial network is used to map the discretized risk label values ​​to a risk scenario set consisting of different risk scenarios in the security layer, physical layer, and financial layer. The synchronization deviation is used to quantify the degree of deviation in power rate and electricity price, thereby forming a system with multidimensional scenarios, high tail accuracy, and strong risk controllability. This link provides subsequent scheduling optimization with more input samples of different resolutions and higher scenario diversity, breaking through the process limitation of traditional scheduling that can only make scheduling decisions based on a single pipeline prediction scenario.

[0020] Safety loss, power loss, and price loss are integrated into the same optimizer, and the minimum comprehensive tail loss is obtained using mixed integer second-order cone programming. The decision vector includes energy storage power, load regulation, and bidding adjustment. The three types of losses are all in the same metric space for game play, minimizing risks while ensuring maximum benefits.

[0021] A dual-threshold circuit breaker mechanism is constructed before the issuance of instructions using weighted consistency indicators and comprehensive risk indicators, and a multi-resolution trend-market coupling sandbox is used to perform second-level prediction, evaluation, and issuance. The mathematically optimal solution has been physically and market-tested before the instruction is issued, eliminating the situation where the risk is highly measurable but the risk is high. Only when it is predicted that the risk is controllable will the instruction be allowed to enter the execution stage, and once issued, it will immediately begin to operate seamlessly in the model and reality, greatly reducing the plan-execution deviation.

[0022] The generator parameters and CVaR confidence radius are cyclically updated by combining three incremental learning methods: sliding window stratified sampling, natural gradient correction, and Bayesian shrinkage. Then, broadcast packets are used to continuously push the latest model status to scenario generation, risk measurement, and release thresholds, enabling the system to make adjustments at any time to the evolving attack strategies and the ever-changing financial market environment, keeping the system in an effective state. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of the virtual power plant optimization management method based on conditional risk value of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] See also Figure 1 The present invention provides a virtual power plant optimization management method based on conditional risk value, including: As a cloud-based dispatching platform, in scenarios involving high-penetration renewable energy virtual power plants, dispatching instructions to hundreds or thousands of edge controllers should be performed within seconds, taking into account spot electricity prices, carbon quota pricing, and ancillary service quotes. Simultaneously, upstream, power measurements from each distributed photovoltaic, wind, and energy storage unit should be received at the same granularity. This alone makes the entire business chain highly digitized. However, due to this high degree of digitization, traditional monitoring methods that rely solely on single indicator thresholds and offline alerts are no longer able to detect hackers who introduce subtle delays or distortions into data streams.

[0026] Although the initial value of this type of distortion is small, during periods of resource shortage or sharp price fluctuations, it may still be misinterpreted by the cloud as normal distortion and sent to the edge, causing overcharging, overdischarging, or significant load adjustments. This can also further increase the power-price synchronization error.

[0027] In the virtual power plant, the differentiated deviation fee needs to be paid for the instantaneous power difference, and the loss caused by the extreme settlement price jump is also borne, thereby causing the decline of grid frequency, voltage and other indicators. Therefore, the distorted business decision executed by the cloud due to judgment error can be identified in the front end of the dispatch link, and the specific decision risk degree is judged according to the different scale of the business, which is a necessary basic link in the risk control strategy.

[0028] Step one, first use the cross-domain time flow fine-grained alignment and multi-view attention fusion method to model a unified multi-modal feature tensor, and then automatically generate a potential distortion interval based on the historical steady-state distribution, and form a high-credibility data basis for subsequent risk scenario generation.

[0029] The three types of information of instructions, market conditions and power in the virtual power plant dispatch link belong to different systems respectively, but they are closely related in time and business. If there is any deviation in time, order or semantics between the data of each flow, it may accumulate as a tail risk.

[0030] Therefore, on the one hand, millisecond-level synchronization needs to be ensured, and on the other hand, cross-domain association needs to be captured, so a fusion mechanism is needed. It is difficult to reflect the dynamic resonance characteristics of the three flows and impossible to find the weak distortion existing in them by using a static threshold or simple sliding window mean method. Therefore, we start from the multi-modal safety pointer, introduce attention-based feature fusion and steady-state distribution self-supervised labeling, and improve the ability of abnormal transaction detection on the basis of ensuring millisecond-level synchronization.

[0031] Step 101, cross-domain time flow alignment and feature merging of different types of nodes: The instruction vector collected from the edge controller , the price vector and the power measurement vector all have their own independent millisecond timestamp offset, so the access method is to concatenate all the data together, that is, to splice according to the time dimension, through the sliding index-linear interpolation-morphology maintaining time sequence alignment algorithm, and then build a common time axis , so the spliced tensor is the preliminary spliced tensor to be processed:

[0032] Among them: the active / reactive instruction sequence of the instruction vector at time step , the value range is normalized to , for characterizing the planned side control intention; price vector , the spot price, carbon price and reserve service offer triplet at time step , value range is Min-Max normalized to , for mapping the financial side incentive; power measurement vector , the instantaneous output, power factor and bus voltage segment at time step , value range is standardized to , for measuring the physical response; common time axis , the unified millisecond-level time index after interpolation alignment, for ensuring three-flow synchronization; preliminary splicing tensor , the feature matrix obtained by column-level splicing of instruction, price and power three vectors, for providing original perspective for subsequent attention calculation.

[0033] The time series alignment algorithm is an algorithm for aligning multiple source time series with different sampling rates, timestamps, and data missing attributes to the same reference time series according to certain regulations. The following three steps are completed in turn: ① Project each original sequence to the target reference time axis, and construct a sliding window for each time point to quickly locate the nearest valid sampling point from the current anchor point; ② When there are blank, displacement and other missing sample points in the valid data of the sequence, complete the missing sample point filling through linear interpolation method, zero-order hold, local polynomial fitting method, etc., while retaining the original waveform change, so that the sequence meets the length equalization requirement; ③ After completing the three steps, all the filled sequences are numbered in the order of unified millisecond-level index and arranged into the corresponding matrix, and the new matrix with consistent length and time synchronization is output. Using the time series alignment algorithm, the time series of the target station management and control subsystem, the market subsystem and the field acquisition subsystem are processed for cross-space and time synchronization in milliseconds, and these synchronized multi-modal data are used as input to the relevant feature fusion, anomaly detection, risk scenario generation and dispatch optimization links, completely eliminating the influence of time error.

[0034] A unified time anchor is established, so that the data of the three flows can be aligned at any millisecond level, and the influence of time series drift on anomaly detection is eliminated, thereby providing an unbiased baseline for subsequent high-order feature extraction.

[0035] After obtaining the aligned tensor , the interpretable attention mechanism is used to generate a weighted feature tensor :

[0036] wherein the weighted feature tensor , attention fusion output, used to represent the integrated state after three-flow interaction; weight matrix , learned linear mapping matrix, size , value range usually , used for dimension reduction and feature rotation; , number of rows, represents the dimension of the output or projected features; , number of columns, represents the dimension of the input features.

[0037] attention weight matrix , weight obtained by self-attention calculation, both rows and columns are time steps, element range , used to measure the correlation strength of different time slices; activation function , GaussianErrorLinearUnit is adopted, which can improve the nonlinear expression ability of multi-modal features; Hadamard multiplication , element-wise multiplication operator, which couples linear projection and attention weight together; The attention mechanism explicitly learns the importance of different time slices and different modalities, so the platform can bring easily ignored distortion information points into the model, and through statistical modeling, the fusion representation can achieve the purpose of being dense but interpretable.

[0038] Step 102, build a steady-state baseline, and mark the abnormal interval: After unsupervised fitting of the historical steady-state window of the weighted feature tensor using the variational Gaussian mixture model, the baseline density can be obtained:

[0039] abnormal score is defined as:

[0040] where: baseline density , represents the probability density of the feature tensor value under steady state, used to distinguish normal and abnormal; mixing coefficient , the weight of the th Gaussian component, satisfies , value range ; mean vector , the center position of the th component, the dimension is consistent with the feature tensor, and the unit is consistent with the input. covariance matrix , covariance matrix of the same dimension, representing the intra-component covariance, needs to maintain positive definite property; abnormal score , a rarity indicator obtained using the right-tail cumulative distribution, whose range , the closer to 1, the rarer, directly as a risk measure.

[0041] By modeling the mixed density of the historical steady-state period, the tail distance of the feature tensor at any moment can be quantified, providing a probabilistic basis for subsequent abnormal segment determination. Based on the abnormal score construct a safety pointer :

[0042] And use bidirectional Bayesian filtering to smooth the prior-posterior of discrete safety pointers, get the segment-level abnormal label set .

[0043] Where: safety pointer , immediate safety warning quantity, activated when the abnormal score exceeds the threshold , range , used to pass the abnormal intensity to the downstream step; threshold , quantile threshold automatically set according to baseline density, usually take the right tail 95% quantile, used to ensure that only high-risk fragments trigger the pointer; indicator function , logical switch, return 1 if the condition is met, otherwise 0; segment label set , after the safety pointer is smoothed by bidirectional filtering, use the serial number to represent the time point of the occurrence and end of the anomaly, for scene generation.

[0044] The safety pointer converts real-time anomaly measurement into readable operation labels, and removes incidental noise through filtering, making it both sensitive and stable. In quantile-GAN, it is used to refer to the true situation of the labeled sample; From the original multi-modal data to the accurate mapping of the segment-level abnormal label, and the generated weighted feature tensor and segment label set together ensure that the segment-level label input of the generator is highly reliable and can fully cover the risk scenarios and maintain probability consistency.

[0045] From the foregoing process description, it can be seen that the multi-modal security pointer perception method solves the problem of data heterogeneity and time asynchrony. At the same time, through the combination of attention fusion and mixed density estimation, it realizes the chain enhancement of information screening, probability empowerment and abnormality measurement. The continuous ordering of these processes can make the real-time business flow formed by them have high robustness to various weak distortions, and can symbolize the business flow weak distortion into a retrievable label, meeting the requirements of high signal-to-noise ratio and high time sequence alignment of the data training and inference model for subsequent attack risk quantile mapping and scene construction.

[0046] With the generation of risk scenarios and the coupling optimization of three layers of CVaR expanding layer by layer in steps two and three, the above-mentioned parameters will form a transverse needle thread, organically combining the time sequence information of the security layer, the physical layer and the financial layer, and fully reflecting in the overall process of the virtual power plant optimization scheduling method, ensuring that the final output scheduling strategy can maximize the benefits while minimizing the tail loss in extreme cases.

[0047] The biggest source of uncertainty in real-time scheduling of virtual power plants is not the regular fluctuation, but the cross-domain amplification of micro-distortion after the security layer is triggered, causing a chain reaction of tail impact, and realizing high-confidence positioning labeling of possible abnormal places.

[0048] However, if these discrete labels cannot be converted into probability scenarios across the security-physical-financial three perspectives, the decision layer cannot use models to quantify the tail risk.

[0049] Step two, according to the abnormal section label, construct the cross-domain quantile prior, based on the improved quantile-GAN, get the high-confidence risk scenario set of security-physical-financial three-dimensional coupling, as the quantitative tail input of subsequent three-layer CVaR optimization.

[0050] Step 201, abnormal label embedding and quantile prior construction: The distortion section has been obtained through the section label set However, directly feeding the generation model will have the problem of sample sparsity and tail dispersion, and it is difficult to provide sufficient support for each quantile. Therefore, it is necessary to first embed into the continuous quantile space to construct a differentiable and sampleable prior distribution, so as to stabilize the training generator.

[0051] The distortion section has been obtained through the section label set However, directly feeding the generation model will have the problem of sample sparsity and tail dispersion, and it is difficult to provide sufficient support for each quantile. Therefore, it is necessary to first embed

[0052] To avoid the static metric value, set the abnormal intensity mapping :

[0053] And construct quantile label accordingly :

[0054] Wherein: Inverse cumulative distribution function of abnormal intensity, inverse cumulative distribution function (also known as quantile function) is a mathematical tool in probability statistics to map probability value to random variable value, ensuring .

[0055] Abnormal intensity mapping , the safety pointer Linearly mapped to , used to measure the degree of distortion; threshold , the lower limit of the safety pointer trigger, the value comes from the setting of quantile in the previous step, which is constant; quantile label , used to guide the generation network to generate the corresponding intensity sample after mapping at the end of the network.

[0056] Further use of mapping-inverse distribution process, the discrete safety pointer is converted into continuous quantile number through smooth transition, so as to provide a more detailed tail control signal for GAN network; with the quantile label , the scene index matrix :

[0057] And the scene index matrix Concatenated with Gaussian noise vector Get input vector :

[0058] In the formula: noise vector , zero mean unit variance Gaussian vector, used to provide randomness; concatenation vector , as the final input of GAN generator, combines randomness and quantile information.

[0059] Scene index matrix , quantile label sequence arranged in chronological order, used to drive the generator to output different risk intensity; Concatenate the quantile label to the noise vector, so that the generator can feel random jitter and tail end level on the sampling space at the same time, which can avoid the shortcomings of traditional conditional GAN that is difficult to depict continuous tail end intensity gradient.

[0060] Step 202, improved quantile-GAN training and risk scenario generation: Traditional GANs only align the overall distribution and cannot guarantee the quality of generated samples at the tail end. Therefore, we use a dual-objective training method combining adversarial loss and quantile-Wasserstein constraints to align the mainstream distribution and accurately align the tail-end quantile points, and design a corresponding generator and discriminator and optimize its objective function:

[0061] In the formula: the first and second terms of the adversarial loss are the conventional GAN cross-entropy, used for overall distribution alignment; the generator is a deep network with a set of parameters used to generate three-flow scenarios; the discriminator is a deep network with a set of parameters used to distinguish between real and generated samples; the real distribution is the distribution of measured three-flow data corresponding to the abnormal section label; The first-order Wasserstein distance is used to measure the difference between the conditional quantile layer distribution, and the balance coefficient is used to adjust the weights of the adversarial loss term and the quantile loss term, with a value controlling the tail fitting accuracy; the conditional symbol indicates that the distance is calculated under the given quantile condition.

[0062] The goal is not only to ensure the fitting of the mainstream distribution while aligning the quantiles, but also to achieve a high fitting of the tail quantiles without causing significant damage to the overall realism. After training converges, input a set of quantile sequences and random noise to generate a risk scenario set :

[0063] Simultaneously record the synchronization deviation scalar of each scenario :

[0064] In the formula: the risk scenario set consists of three-flow samples output by the generator, used for subsequent optimization; the number of scenarios is adaptively set according to the optimization complexity, taking a positive integer value; generate the instruction vector and the price vector , generating power vectors , scenarios corresponding three-flow data; synchronization deviation scalar , by mapping function L1 distance between the price estimate after mapping the power vector as the price estimate and the generated price, used to quantify the power-price mismatch; mapping function , price-power nonlinear mapping obtained by modeling the day-ahead market experience, used to convert physical quantities into financial quantities; three-dimensional coupled scenario set that retains both the physical constraint texture of the instruction and power and the financial characteristics of the market price; deviation scalar can directly provide comprehensive tail measurement for the subsequent three-layer CVaR optimization, breaking the quantitative shackles of the security-physical-financial three-dimensional space.

[0065] label mapping of abnormal segments to quantile sequences , generating a risk scenario set using an improved quantile-GAN and introducing synchronization deviation , which is the high flexibility input interface of the three-layer CVaR coupled optimization in the third step with quantile-sample-deviation as the three elements, for decision-making risk budget segmentation, providing random scenario samples, can directly enter the tail loss function.

[0066] Through the steps centered on quantile embedding and improved quantile-GAN, discrete and sparse abnormal labels are converted into continuous and sampleable risk scenario sets, realizing three-dimensional coupled expression that organically combines across the security-physical-financial three aspects.

[0067] In this process, abnormal intensity mapping realizes unobstructed docking of security pointers and quantile labels; double-target adversarial training ensures high-precision fitting of tail quantiles; and synchronization deviation can directly introduce the risk scenario set into the subsequent CVaR objective function. Compared with traditional single-layer risk simulation, after replacing the original tail quantile fitting strategy with double-target adversarial training, this method can dynamically adjust the scenario distribution according to the risk budget, and has greater extreme scenario coverage than the result obtained by directly drawing scenario trees from third-party vendors, and can provide flexible and programmable input granularity for three-layer CVaR optimization. When the risk scenario set is added in step three, the virtual power plant can truly realize coordinated decision-making between the security layer, the physical layer, and the financial layer according to the unified risk measure, and can guarantee excellent performance of both power and revenue indicators regardless of the severity of the situation.

[0068] According to the detection-generation link obtained by the virtual power plant, a risk scenario set with quantile scale can be obtained and the synchronization deviation scalar corresponding to different scenarios However, it is difficult to ensure that the grid safety, power balance and price income are met at the same time in the worst case by adjusting the energy storage charging and discharging, controllable load and bidding curve only by intuition or based on static experience. The traditional method is to independently optimize market income when solving power tracking error, and at the end, external risk exposure due to future network attack behavior is not considered, which will lead to triple losses of power penalty + transaction loss + safety recovery cost of dispatching results in a settlement period when facing attack and serious market fluctuation.

[0069] Step three, based on the risk scenario set, a CVaR coupled optimization model is constructed, and the energy storage-load-bidding linkage decision vector is obtained through the process by using the iterative risk budgeting method to achieve comprehensive tail loss minimization.

[0070] Step 301, three-dimensional tail loss mapping and CVaR hierarchical expression: In order to measure the safety, power and price three types of tail risks in the same optimization framework, the original quantity in the risk scenario set needs to be converted into a loss scalar with consistent numerical value, and a CVaR analytical expression is constructed accordingly. Thus, the mapping relationship between the decision variable and the tail loss is kept differentiable and coupled.

[0071] Wind-solar-storage coupling scenario The three types of losses under the wind-solar-storage coupling scenario are defined as:

[0072] In the formula: the safety loss scalar is amplified by the synchronization deviation amplification coefficient and the synchronization deviation Linear mapping is used to measure the abnormal safety reset cost; The power loss scalar is measured by the two-norm distance between the scheduling instruction and the actual power The amplification coefficient reflects the penalty unit price; the price loss scalar is measured by the one-norm distance between the generated price and the reference offer The amplification coefficient reflects the degree of income loss; the reference offer is composed of the segment transaction average price announced by the day-ahead market, which is used to measure the deviation of strategy pricing.

[0073] Thus, the three-dimensional risk is mapped into a unified loss space in the form of dimensionless scalar, ensuring the tail risks with different dimensions can be directly weighted and compared in the subsequent CVaR calculation. The CVaR expression is constructed for each type of loss:

[0074] wherein: tail level represents the confidence section reserved in the tail distribution, and the higher the value represents the more sensitive to extreme scenarios; reduction variable , represents the real number field, and the auxiliary variable introduced by optimization is used to locate the CVaR clipping point; positive part operator ensures that only the tail part exceeding is counted in the expectation; expectation operator takes the expectation of the scene index according to the scenario set probability distribution; category index corresponds to the three types of tail losses of safety s, power p, and price c.

[0075] By introducing the reduction variable , the non-convex tail risk is converted into a convex expectation, which is compatible with linear / quadratic cone constraints, thereby ensuring the solvability of optimization under large-scale scenarios.

[0076] Step 302, mixed integer quadratic cone programming and iterative risk budgeting: After obtaining the three types of analytical expressions, the decision variables and physical / market constraints need to be included in the solution, and a risk budget feedback mechanism needs to be designed to make the CVaR weight shrink in real time according to the operating state. Only in this way, can the dynamic optimal balance between income and safety be ensured in the scheduling result.

[0077] Define the decision vector :

[0078] Objective function:

[0079] The constraints include:

[0080] wherein: energy storage power vector , the charge and discharge power in each period of the decision cycle, normalized to according to the rated power of the device; load adjustment vector , , refers to the set of non-negative integers, and the reduction amount of the stoppable and movable load in each period is a non-negative integer; bid adjustment vector The increment / decrement declaration quantity submitted at different price levels is subject to the market mixed integer second-order cone constraint set ; The initial energy state vector at the beginning of the current scheduling period is derived from the real-time SoC measurement of the previous control period; the energy storage state matrix The lower triangular matrix that maps the integral of charge and discharge power to energy increment; the power coupling matrix The mapping matrix that applies the instantaneous injection of energy storage power to system flow; Upper and lower energy bounds , Physical boundaries of SoC of energy storage; upper and lower power bounds , Indicate the safety limits of inverters and busbars; weights The priority coefficient set by the operation and maintenance party according to the risk tolerance; Mixed integer second-order cone set The bidding quantity must fall within the feasible region defined by the minimum declaration quantity, price ladder and second-order cone relationship.

[0081] Unify physical, market, integer decision and CVaR target in the same MISOCP structure, so that the problem can be converged by commercial solvers within a feasible time, and weights Leave adjustment space for subsequent risk budget feedback.

[0082] The optimal value of the objective function is :

[0083] If the measured loss After execution, it meets:

[0084] Then update the weight:

[0085] In the formula: the optimal value of the target The comprehensive tail loss obtained by optimization solution; the measured loss According to the real-time measurement and market settlement calculation after execution; the tolerance coefficient Indicates the multiple threshold value that allows the measured loss to exceed the model estimate; risk gain rate When the measured loss exceeds the threshold, increase the weight in proportion; update operation Indicates that the new value is written back to the risk control dictionary for solving in the next period.

[0086] Based on the closed-loop contrastive actual measurement loss and theoretical CVaR, the proportion of excessive category weight greater than 1 is dynamically improved, the adaptive contraction of risk budget for the remaining capacity is realized, the model pays more attention to the tail categories with poor performance in the next period, the cumulative risk exposure is achieved, and the punitive learning effect is achieved. The upper star in the figure indicates the minimum tail expectation obtained by the reduced variable solution in the optimization process.

[0087] Step 301 unifies the three types of tail losses into the analytical CVaR expression, realizes the three-dimensional mapping of risk-safety, power and price, and compares them in the same measurement space; in step 302, the decision variables, physical / market constraints and three-layer CVaR target are coupled into the MISOCP framework, and the closed-loop learning is formed by adaptive contraction of risk budget.

[0088] Based on the three-dimensional tail loss mapping and CVaR hierarchical expression in step three, combined with mixed integer second-order cone programming and adaptive contraction of risk budget, the mapping process from generating risk scenarios to forming execution instructions is established, and the simultaneous game of safety risk, power risk and price risk in this step is realized based on a mixed integer second-order cone programming optimizer in a single solving process, obtaining the global optimal scheduling instruction of one-time optimization. At the same time, the closed-loop risk budget of this step enables the model to actively warn when dealing with continuous attacks or market black swans, iteratively contract tail losses, and does not require additional human intervention; based on this, in step four, the consistency verification of edge-cloud is realized and the rapid test run in the sandbox environment is completed, and after confirming the mathematical optimality and engineering feasibility, the virtual power plant is truly realized in the scene of the most uncontrollable edge side through cloud-edge collaboration to achieve the extreme value of safety-income-steady-state trinity.

[0089] After the cloud CVaR three-layer coupled optimization, the optimal decision vector and the optimal energy storage power vector , load adjustment vector and bidding adjustment vector are generated; since there may be differences in microseconds between the assumed network delay, device transient, market matching in the cloud model and the actual situation, if the entire instruction package is directly issued to each gateway for execution, it is easy to cause the loss of instruction integrity or the failure to capture the local constraints of part of the instruction caused by communication jitter, resulting in errors in plan execution. Once this plan execution error triggers a secondary event, it is easy to cause a secondary safety production event, and has a superimposed impact on the entire production safety layer, physical layer and financial layer, and makes the previously obtained optimal solution fail. Therefore, an overall process of instruction collaboration verification→rapid sandbox simulation→threshold judgment→immediate release should be added between the cloud and the edge to ensure the consistency of the mathematical model and the physical site of the platform and not to delay the speed of second-level scheduling.

[0090] Step four, consistency check and rapid sandbox double-chain verification mechanism between cloud output and edge physical site is established, and threshold strategy is set to ensure that the instruction set with acceptable risk can be published in real time.

[0091] Step 401, weighted consistency check and instruction unpacking: Cloud output decision vector Including multi-dimensional continuous variables and discrete variables, when converted into corresponding terminal-side micro-operation commands, it needs to be further split and refined according to the characteristics of each terminal-side device itself.

[0092] Considering the communication link delay and the precision problem of each terminal-side device execution, the consistency error vector calculation is carried out in the cloud-edge interaction window, and each instruction output to the terminal side is determined according to a certain weighting method.

[0093]

[0094] Among them:

[0095] In the formula: communication delay mapping matrix Map the cloud-delivered charge and discharge plan to the arrival power of different delay channels, the element range ; Edge feedback power The actual received energy storage power vector in the last period, the range ; Edge feedback load Actual load reduction, non-negative integer; Edge feedback bid The amount of bid that has taken effect in the last period, floating point; The norm operation remains consistent with the previous step to ensure dimensional consistency.

[0096] Then calculate the weighted consistency index : ,

[0097] The weight is set according to the three-dimensional tail risk sensitivity, the value , and satisfies .

[0098] Linearly combine different types of errors into a single index So that the subsequent decision can be released by threshold judgment, avoiding the delay caused by manual item-by-item review. If the consistency index Satisfies:

[0099] Then execute instruction unpacking:

[0100] The execution amplitude :

[0101] Where: consistency threshold , the operation and maintenance party can dynamically adjust the scope according to the network quality ; Device Collection , including all device indexes of energy storage, load and bidding channels; time slice collection , the same time scale as the optimization model; the device type set , , , subsets distinguished by function; Energy storage power decision-making quantity , at the moment of optimization solution Energy storage charge / discharge power (positive value for discharge, negative value for charge); Load regulation decision , the flexible load obtained in step 3 At time The reduction in power or displacement; Bid adjustment decision volume , the bidding channel obtained in step 3 At time The increase or decrease in the reported amount represents the increase or decrease in the declared power at the corresponding price level; Execution amplitude , the power, reduction amount or declared price difference that the corresponding equipment needs to execute at the corresponding time.

[0102] Once the consistency index Qualified means that cloud plans can be quickly split according to device types without the need for loop judgment; at the same time, the device label and time are kept synchronized to provide clear input for sandbox simulation.

[0103] Step 402: Rapid sandbox simulation and threshold release: Edge physical nodes can simultaneously receive real-time measurement values ​​and unpacking instructions from the cloud. Without affecting the decision-making rhythm, they can analyze the risks after the instructions are issued and executed, build a multi-resolution trend-market coupling sandbox, make millisecond-level power and price linkage predictions, and use comprehensive risk indicators to determine whether they can be issued immediately.

[0104] The core of the sandbox consists of two layers of models: electrical layer: using reduced-order power flow Jacobian The voltage increment is obtained by fast iteration method :

[0105] Market layer: local supply-demand elasticity vector Estimate instantaneous price:

[0106] Where:

[0107] Thus, the prediction-execution difference is obtained :

[0108] In the formula: power flow Jacobian , which is obtained by online linearization of the edge measurement in the last step, and is positive definite; power injection vector , the injected power calculated according to the unpacking instruction and the node position; reference power , the power corrected by real-time measurement; Voltage increment , predicted voltage offset, unit ; price elasticity vector , estimated according to the local supply-demand elasticity curve, floating point; instantaneous price prediction , predicted market price output by the sandbox; reference price , reference price broadcast by the market; Predicted power , predicted output after instruction execution; , the energy storage charge-discharge power decision quantity obtained by three-layer CVaR optimization in the last cycle, positive for discharge and negative for charge; , the flexible load reduction quantity vector obtained by optimization, corresponding to the power reduction value of each adjustable load in the time slice; , price deviation amplification coefficient, converting the dimension of currency to the comprehensive weight of the same scale as the power deviation, the value of which is set by the operator according to the risk preference; Real-time power is the current power generation output detected by the edge; prediction-execution difference is the quantitative sandbox prediction and the actual situation. Through the sandbox, the power-price dual-quantity linkage prediction is realized, and the prediction-execution difference is used as an immediate risk probe for reference when publishing decisions, and a comprehensive risk index :

[0109] If the following conditions are met:

[0110] Immediately call the edge executor to publish , and sign the publication upload to the cloud, then continue to the next process; if the repair flow is triggered, the consistency index , risk index upload to the cloud, and call MISOCP again through the cloud.

[0111] In the formula: comprehensive risk index , the maximum value of the linear combination of power-price difference and voltage offset; weighting coefficient , , set by the operation and maintenance party according to the security level; release threshold , below which the risk is considered controllable; release signature , a hash generated by the blockchain trusted execution environment, which realizes the backtracking function.

[0112] Use the maximum time risk as the release standard, and repair it as long as the risk of any time slice exceeds the allowed value; prevent tampering or denial of instructions through hash signature.

[0113] is the consistency index derived from the cloud optimal decision and the edge report information of the last time in the current period , if qualified, it is disassembled into edge instruction set ; step 402 is to complete the power-market coupling prediction under the local sandbox , and give the comprehensive risk index , if , release the instruction and generate the signature , otherwise repair it, and trigger the cloud to recalculate. Through the threshold control the model-reality difference within a quantifiable range, and obtain high-confidence execution-prediction comparison data that can be used for real-time loss sampling and model parameter updating in the next step.

[0114] Step four is to convert the optimization output obtained by high-level abstraction in the cloud into micro-instructions that can be directly run on the edge, and to pre-predict the impact (millisecond level) on the physical-financial-security brought by the release of this optimization scheme, thereby reducing the probability of failure. At the same time, the weighted consistency measure is used to solve the instruction drift problem caused by communication delay and historical execution deviation, and the comprehensive risk measure is used to realize one-stop evaluation of power-price deviation and voltage safety.

[0115] Use a double-threshold mechanism The second step is to implement a reliable real-time fuse entry under extreme scenarios under the second step of the scheduling rhythm, to ensure that all instructions are only executed after safe and reliable judgment, and all risky instructions are rejected. For high-risk indicators, MISOCP recalculation and weight are triggered after automatic backwriting to the cloud Adjustment, risk budget shrinkage, and closed loop. At this point, the virtual power plant completes the full-stack end-of-process risk closed-loop management of the detection-generation-optimization-verification process, completes real-time loss sampling and model parameter updating, and feeds back to the virtual power plant optimization scheme stage of step three of the next iteration cycle.

[0116] After the fourth step of double threshold release, the edge-cloud has entered the benign operation track of second-level rhythm, but the risk scenario set and the three-layer CVaR model are still in the pre-event and event assumption.

[0117] That is, if we keep our generator parameters and confidence radius long Time represents the tail risk of the system, but the means of hackers, the behavior of load parties, and the mechanism of the market are constantly changing. If we do not map the real loss of the three lines in execution-measurement-settlement to our model parameters in a timely manner, our generator will not be able to cover the tail, and the cut point of the three-layer CVaR will deviate from the tail distribution under the latest mode.

[0118] Step five: Update the generator parameters and CVaR confidence radius by driving the joint natural gradient update method with real-time loss observation, so that the risk model can adapt to the dynamic environment.

[0119] Step 501, streaming loss sampling and dynamic statistical aggregation: When the edge executor releases the instruction set and generates a release signature , the three-stream data instruction-execution-market begins to refresh at the millisecond level. If these high-frequency feedbacks are not structured sampled, data flooding and model starvation will coexist: useful tail samples are submerged, and the generator cannot learn the latest anomalies. Therefore, a loss sampling mechanism is needed to accumulate the three types of losses: safe loss scalar , power loss scalar , and price loss scalar , and extract key statistics to provide low-latency, high-entropy input for subsequent incremental learning.

[0120] First, define the sliding window index set :

[0121] Construct a hierarchical sampling probability :

[0122] and the hierarchical sampling probability random sampling samples, forming a loss subset :

[0123] wherein: the sliding window index set : the most recent set of time points, , is a positive natural number set, set by the operation side; is a loss category label: represents a security loss, represents a power loss, represents a price loss, , is a scenario index within the sliding window, used to enumerate the loss value of each tail-end scenario participating in the probability calculation; hierarchical sampling probability : the samples within the sliding window are weighted according to the loss power , with the tail-end samples having a larger weight; the power factor : adjusts the degree of tail-end amplification, with a value range of ; the number of samples : the number of samples retained for each type of loss, determined according to the computing power and real-time constraints.

[0124] Through power-weighted sampling, the tail-end samples obtain a higher retention rate under the same communication overhead, ensuring the learning density of the generator on rare patterns. The three types of subsets are concatenated according to the time index to obtain the loss observation matrix :

[0125] wherein .

[0126] security loss subset , a row vector composed of security loss samples extracted according to the power-weighted sampling within the sliding window; power loss subset , a row vector composed of power loss samples extracted within the sliding window; price loss subset , a row vector composed of price loss samples extracted within the sliding window; is a 3-row column matrix space in the real number field; wherein: the loss observation matrix : the real-time observation stacked according to the loss category, used for online estimation of quantiles and gradients; Total sample number : The sum of three sample sizes, taking values .

[0127] The matrix form facilitates the one-time calculation of new quantile statistics, gradient estimates, and KL divergence, creating a data interface for subsequent natural gradient updates. The loss observation matrix The empirical quantile is calculated column by column :

[0128] And calculate the category entropy :

[0129] In the formula: empirical quantile : the quantile estimate of the loss in the sliding window; category entropy : measures the diversity of samples, the larger the value, the more uniform the distribution; , loss category index, taking values representing the safety loss, representing the power loss, representing the price loss; , the category sampled loss sample value, , the confidence level of the category , equal to the probability mass required to be retained by the conditional value at risk, usually taking values close to 1.

[0130] Further, the empirical quantile provides a real-time benchmark for CVaR clipping, and the category entropy serves as an entropy measure, injecting a diversity penalty into the generator update to prevent pattern collapse.

[0131] Step 502, generator-CVaR joint natural gradient update: The existing loss observation matrix , empirical quantile and observation information are all injected into the generator parameters and the confidence radius . Based on the current problem that traditional stochastic gradient relies too much on tail samples and has distribution bias, it is difficult to converge, here we choose to use the natural gradient-Bayes mapping dual-channel scheme which is more adaptive: use natural gradient to update ; use conjugate Bayes to update ; finally, the two results are fused on the cloud, and unified communication broadcasting is carried out on the cloud.

[0132] Set generator output distribution Estimate target distribution with loss observation matrix . Minimize KL divergence with natural gradient:

[0133] Where: Fisher information matrix, with current parameter Fisher information matrix calculated, describes local curvature of parameter space. Step size : Adjust convergence speed, learning rate; Positive scalar less than 1, used to control parameter update step size; Model output distribution defined by current generator parameter Natural gradient : Steepest descent direction in the sense of information geometry, at time Natural gradient correction of generator parameter vector , used for online update in this period. Target data distribution estimated by loss observation matrix in real time, reflecting the statistical characteristics of the latest tail samples; Gradient operator Further, the natural gradient has curvature correction, which can avoid oscillation caused by too large step size when the tail samples surge. Set confidence radius vector Prior distribution is Gamma:

[0134] Update posterior: ,

[0135] In the formula: shape parameter : Cumulative sample size, integer, Corresponding to the next time; Scale parameter : Cumulative excess loss of hyperquantile, Corresponding to the next time; Gamma distribution : Conjugate prior ensures analytical update, Indicates the category Empirical quantile benchmark at the current time .

[0136] Further, Bayesian update makes confidence radius automatically shrink with increasing samples, and reversely enlarge if tail excess is frequent, realizing adaptive risk budget. Synthesize joint update:

[0137] And generate broadcast packet​ :

[0138] Step two and step three are pushed to the next round of scenario generation and used as CVaR solution.

[0139] In the formula: broadcast packet , carrying the updated generator parameters, confidence radius, entropy measure.

[0140] Entropy measure , as a reference index of sample diversity, for the generator to avoid falling into pattern collapse again in the next round. The joint update formula makes both effective and gives consideration to the development speed of both, when neither can fall behind the other, both can steadily move forward. Single packet broadcast can avoid multi-channel synchronization delay problems and ensure that all global nodes can directly use the latest risk scale in the next period.

[0141] The high-frequency loss flow is compressed into a statistical packet with optimal information density by using sliding window hierarchical sampling-observation matrix aggregation-quantile+entropy double index three-section type; The natural gradient-Bayesian conjugate double-path update mode is adopted, and the iteration process changes the generator parameters and confidence radius in the model, and makes a broadcast . This step drives new quantile-GAN training; the second step drives CVaR cut point update; the third step is used as the basis for dynamic adjustment of the issuance threshold; the fourth step is from feedback->risk model flow, which plays a role of a needle in a chain on the closed loop.

[0142] Step five maps the real-time loss flow into the power needed for model evolution, establishes a 5-level adaptive link composed of sampling-statistics-natural gradient-Bayesian contraction-broadcast, so that the virtual power plant risk model has the ability of self-evolution.

[0143] The application completes the optimization parameter refresh based on the CVaR framework in a few minutes without reducing services, without offline retraining, and without batch historical data, reduces the operation and maintenance cost to the minimum, and controls the running risk exposure in the shortest time window; combined with abnormal detection, the risk control self-healing closed loop under the transaction frequency abnormal transaction scenario is connected, which can guarantee the normal operation of the whole system in parallel when high-frequency network attacks, large amplitude market jumping, equipment degradation, and security-income-stability triple optimal target realization.

[0144] Those skilled in the art can clearly understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0145] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0146] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0147] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0148] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A virtual power plant optimization management method based on conditional value at risk, characterized by: include, The edge synchronously collects instruction streams, market flow, and power flow, generates a unified feature tensor through multi-view attention fusion, marks the distorted segments, and determines the abnormal intensity based on the historical steady-state distribution; The distorted segments are fed into an improved quantile generative adversarial network to generate a set of risk scenarios coupling safety, power, and price, along with synchronization deviation quantile labels for subsequent optimization calls and processing. Taking the risk scenario set as input, a global target of risk value coupling with three layers of safety, power, and price is constructed. Energy storage, load, and bidding variables are combined to solve the small tail loss scheduling solution. The scheduling plan is checked for consistency between the cloud and the edge. If it passes the check, it is unpacked and generated into a device instruction set. The power and price deviation is predicted in the local power flow-market sandbox and issued if it falls below the release threshold. The execution loss is sampled in real time and the observation matrix is ​​formed. The network parameters are updated using natural gradients. The confidence radius is shrunk using the Bayesian method and the update package is broadcast for the next cycle to continue.

2. The virtual power plant optimization management method based on conditional value at risk according to claim 1, characterized in that: Perform sliding index interpolation alignment on the original timestamps of the instruction stream, market flow, and power flow to obtain a unified millisecond-level public timeline; A multi-view attention mechanism is used to learn the weights of the aligned feature matrix and linearly map it to generate a weighted feature tensor. The multi-view attention mechanism simultaneously calculates the temporal self-attention weights and the inter-modal interaction weights.

3. The virtual power plant optimization management method based on conditional value at risk according to claim 2, characterized in that: A variational Gaussian mixture steady-state model is established based on the weighted feature tensor to obtain the steady-state probability density; the anomaly score is calculated and a safety pointer is generated; the continuous safety pointer is smoothed using a bidirectional Bayesian filter to generate segment-level anomaly labels, which are used as input indexes for risk scenario generation.

4. The virtual power plant optimization management method based on conditional value at risk according to claim 3, characterized in that: The security pointer is linearly stretched and mapped with the inverse cumulative distribution to obtain the quantile label and construct the scene index matrix; The scene index is then concatenated with a random noise vector to form the generator input, driving the quantile generative adversarial network to output multi-scene samples with quantile control capabilities.

5. The virtual power plant optimization management method based on conditional value at risk according to claim 4, characterized in that: The training objective of the quantile generative adversarial network is composed of the coupling of adversarial cross entropy loss and quantile first-order Wasserstein distance, and a balance coefficient is set to adjust the weights of the two types of losses; the distribution difference between the generated samples and the real samples under the same quantile conditions is gradient optimized.

6. The virtual power plant optimization management method based on conditional value at risk according to claim 5, characterized in that: Establish safety loss, power loss and price loss scalars respectively; The three types of losses all form conditional risk value analytical expressions based on shear variables and tail excess expectations; the three types of risk values ​​are weighted by weight coefficients to form the overall objective function.

7. The virtual power plant optimization management method based on conditional value at risk according to claim 6, characterized in that: The overall objective function, energy storage power vector, load regulation vector and bidding adjustment vector together constitute a mixed integer second-order cone programming model, and set energy storage energy boundary, bus power boundary and market ladder declaration second-order cone constraints to solve the dispatch vector with the minimum comprehensive tail loss.

8. The virtual power plant optimization management method based on conditional value at risk according to claim 7, characterized in that: The communication delay mapping matrix is ​​used to perform channel correction on the cloud energy storage power plan and calculate the quadratic error with the edge power feedback of the previous cycle; The energy storage error, load error and bidding error are linearly synthesized into a consistency index according to the preset weights, and compared with the consistency threshold to decide whether to unpack and release the instruction.

9. The virtual power plant optimization management method based on conditional value at risk according to claim 8, characterized in that: The edge performs coupled sandbox simulation on the unpacking instructions in the power flow Jacobian fast iteration model and the local supply and demand price elasticity model to calculate the power-price comprehensive deviation and voltage offset; the two are combined into a comprehensive risk indicator according to the weighted coefficient, and after comparing with the release threshold, it is decided whether the instruction should be immediately released or repaired.

10. The virtual power plant optimization management method based on conditional value at risk according to claim 9, characterized in that: Through the sliding window stratified sampling mechanism, samples are extracted in each type of loss according to the power weight to form a loss subset; The three subsets are spliced ​​into a loss observation matrix, and the real-time empirical quantile and category entropy are calculated based on the matrix for subsequent incremental model updates.

11. The virtual power plant optimization management method based on conditional value at risk according to claim 10, characterized in that: Based on the loss observation matrix, the generator parameters are updated with natural gradients, and the posterior shape and scale parameters of the confidence radius are updated using the gamma conjugate Bayesian method. Finally, the updated generator parameters, confidence radius, and category entropy are encapsulated as a broadcast package and sent to the risk scenario generation and scheduling optimization module for the next cycle.

Images