A Virtual Power Plant Revenue Prediction and Optimization Method Based on Multi-Scenario Probability
By constructing the R-vine Copula multidimensional joint cumulative probability distribution and an improved particle swarm optimization algorithm, the accuracy and stability of virtual power plant revenue prediction and scheduling optimization are achieved. This solves the problems of low accuracy and unreasonable scheduling in existing technologies, and improves the operational stability and economic benefits of virtual power plants.
Patent Information
- Application Number
- CN202611254450.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-17
AI Technical Summary
Existing virtual power plant revenue forecasting and dispatch optimization suffer from low accuracy, unreasonable surplus power dispatch configuration, and insufficient power supply balance in the distribution network. They cannot adapt to the multidimensional uncertain fluctuation characteristics of the power system, and traditional optimization algorithms are prone to getting trapped in local optima.
A virtual power plant revenue prediction and optimization method based on multi-scenario probabilities is adopted. By constructing the R-Teng Copula multidimensional joint cumulative probability distribution, multidimensional probability feature vectors are generated. Combined with the improved particle swarm optimization algorithm, global surplus power weighted adaptive scenario clustering and secondary potential hierarchical clustering are performed to optimize the scheduling scheme.
It improves the accuracy of revenue forecasting and the rationality of surplus power dispatch, enhances the risk resistance and operational stability of virtual power plants in complex market environments, and improves the efficiency of optimization solutions and engineering operability.
Smart Images

Figure CN123121608A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of virtual power plant optimization, specifically a method for predicting and optimizing the revenue of virtual power plants based on multi-scenario probabilities. Background Technology
[0002] With the continuous increase in the penetration rate of renewable energy, virtual power plants have become an important technical path to improve the utilization efficiency of distributed energy by aggregating heterogeneous resources such as distributed power sources, energy storage and adjustable loads to participate in the operation of the electricity market.
[0003] Existing virtual power plants generally employ traditional deterministic single-point forecasting and static dispatch optimization methods, which have significant technical drawbacks. The core problems are low accuracy in revenue forecasting and dispatch optimization, and unreasonable allocation of surplus power. Current technologies often analyze single variables such as power sources and loads in isolation, failing to consider the coupled effects of multi-dimensional uncertainties. This makes it difficult to adapt to the dynamic fluctuations of power sources, loads, and markets, resulting in large prediction biases and distorted revenue assessments. Furthermore, traditional technologies use a unified resource dispatch model, failing to finely differentiate the adjustable potential of distributed resources, leading to poor dispatch matching and potential for surplus power waste and resource overload. Conventional optimization focuses only on economic indicators, neglecting distribution network balance constraints. Combined with traditional intelligent algorithms, this easily leads to local optima, further reducing dispatch optimization accuracy and making it difficult to balance the economic operation of virtual power plants with the safe and stable operation of the power grid. Therefore, technical optimization and improvement are urgently needed. Summary of the Invention
[0004] (a) Technical problems to be solved To address the aforementioned technical problems, the present invention aims to provide a virtual power plant revenue prediction and optimization method based on multi-scenario probabilities. This method solves the technical problems of existing virtual power plants relying on fixed single-point prediction, extensive unified scheduling, and single-objective optimization mechanisms, which cannot adapt to the multi-dimensional uncertain fluctuation characteristics of the power system, are difficult to adapt to the distributed resource regulation capabilities in a differentiated manner, and have limited optimization accuracy. Consequently, these methods result in poor revenue prediction accuracy, unreasonable surplus power scheduling configuration, and insufficient power supply balance in the distribution network.
[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities, the method comprising: Acquire basic data and grid operation constraints from the dimensions of distributed power sources, loads, and the electricity market; After preprocessing the basic data, a source-load-market coupled probability model is constructed, multi-scenario hierarchical surplus power prediction is performed, and the multi-scenario surplus power probability distribution and multi-dimensional probability feature vector corresponding to each device are output. Based on multidimensional probabilistic feature vectors, a first-level global residual power weighted adaptive scenario clustering is carried out to divide the first-level power consumption scenario clusters and generate the equivalent encapsulation parameters corresponding to each scenario. Based on the primary power consumption scenario cluster and equivalent encapsulation parameters, secondary potential hierarchical clustering within the scenario is used to generate multiple flexible level sub-clusters and sub-cluster equivalent scheduling parameters. An optimal fitness function is constructed with the objectives of maximizing dispatch revenue and minimizing power supply balance in the distribution network. Combined with power grid operation constraints, an improved particle swarm optimization algorithm is used to iteratively find the optimal surplus power dispatch scheme and power supply balance quantification index for each flexible level sub-cluster.
[0006] Preferably, the process for generating the multi-scenario residual power probability distribution and multi-dimensional probability feature vector of each device is as follows: The collected basic data is processed by filling in missing values, removing anomalies, and normalizing to obtain a standardized set of coupled variables and complete the labeling of working condition hierarchical labels. Marginal distributions were fitted to five types of variables: photovoltaic power output, wind power output, rigid load, flexible load, and market electricity price. A multidimensional joint cumulative probability distribution of R-Teng Copula was constructed to quantify the joint occurrence probability of the variables. The original time series scenarios are generated by Latin hypercube sampling based on the joint cumulative probability distribution. The scenarios are reduced by constructing a weighted objective function based on the adjustable residual power fluctuation variance and the estimated revenue, resulting in a set of typical scenarios with probability weights. A dedicated time-series forecasting network is selected to adapt to various equipment types, and outputs the power generation and load forecast values for each device under various typical scenarios; The expected value of adjustable surplus power for each device is calculated by subtracting rigid self-consumption load from the power generation forecast and adding the absorption margin released by the reduction of flexible load. Based on the historical prediction deviation of the equipment, a residual sequence is constructed, and a heavy-tailed distribution is used to fit the residual parameters. Combined with high and low confidence quantiles, a three-layer confidence adjustable residual current upper and lower boundaries are generated. By integrating adjustable residual power expectation, three-layer confidence boundary, and probability weights for typical scenarios, a multi-dimensional probabilistic feature vector matrix is constructed in the device-scenario dimension.
[0007] Preferably, the process of constructing the multidimensional joint cumulative probability distribution of R-vine Copula and quantifying the joint occurrence probability is as follows: The standardized time-series data of photovoltaic, wind power, rigid load, flexible load, and electricity price are respectively fitted with marginal distributions and transformed into a unified probability domain to generate cumulative probability sequences. The R-vine Copula decomposition of five types of variable relationships is adopted using a multi-layered tree structure, and the optimal binary Copula function is adaptively selected at each layer edge. By solving all Copula-related parameters through stepwise maximum likelihood estimation, a five-dimensional joint probability distribution that can characterize the nonlinearity and tail-free correlation of variables is constructed.
[0008] Preferably, the calculation process for the expected adjustable residual power of each device under various typical scenarios is as follows: The rigid self-consumption load that is tied to the equipment and cannot be reduced or transferred is taken as the mandatory consumption amount and deducted from the unit's power generation forecast; Extracting flexible loads can reduce the additional absorption margin corresponding to the capacity as a positive increment; The remaining output after deducting rigid loads is added to the absorption margin to obtain the expected value of adjustable surplus power of the equipment for the current period.
[0009] Preferably, the three-layer confidence-adjustable residual current boundary generation process is as follows: Extract the deviation between the historical actual output and the predicted output of the equipment, and construct a unique residual sample sequence for the equipment. A heavy-tailed probability distribution is used to fit the residuals, and the distribution location, scale, and degrees of freedom parameters are solved. The lower limit of adjustable residual power is obtained by superimposing the low confidence scale quantile offset on the benchmark value of adjustable residual power. The upper limit of adjustable residual power is obtained by superimposing the high-confidence scale quantile offset on the benchmark of the expected value of adjustable residual power. The expected value, lower limit, and upper limit together constitute the three-layer confidence adjustable residual current boundary of the device.
[0010] Preferably, the construction process of the primary power consumption scenario cluster and the equivalent encapsulation parameters of each scenario is as follows: The feature vectors of all typical scenarios of a single device are weighted and fused according to the probability of the corresponding scenario to obtain the global comprehensive feature vector of the device. All equipment is divided into geographical zones according to the distribution network feeder / transformer area, and equipment with no adjustable residual power potential is removed. The characteristics of the remaining equipment are normalized in each dimension to generate a zone-standardized comprehensive feature vector. Construct a weighted composite distance metric function oriented towards scheduling benefits to quantify scheduling differences between devices; Fuzzy iterative clustering is performed based on weighted composite distance, adaptively splitting and merging clusters, outputting the fuzzy membership degree of each device, and obtaining a first-level power consumption scenario cluster after convergence. Based on the fuzzy membership of the devices, the adjustable surplus power, unit revenue, adjustment cost, and response latency within the cluster are fuzzily weighted and summarized to generate equivalent encapsulation parameters for each scenario.
[0011] The preferred method for iterative calculation of the fuzzy membership degree between the primary power consumption scenario cluster and the equipment is as follows: The extreme values of the feature space modulus of the partition are selected as the initial cluster centers, and overlapping centers that are too close are removed to generate the initial clusters. Update the normalized fuzzy membership degree based on the weighted composite distance between the device and each cluster; The coordinates of each cluster center are updated based on a weighted average of membership degrees. Calculate the weighted loss function for this round of clustering; Perform cluster splitting and merging operations based on the dispersion threshold and the cluster spacing threshold, respectively; Repeat the iteration until the change in loss is less than the termination threshold, and output the membership degree of the stable scene cluster and device.
[0012] Preferably, the execution process of the joint strategy of secondary potential hierarchical clustering and two-layer coupling collaborative optimization within the scenario is as follows: The execution process of secondary potential hierarchical clustering within the scenario is as follows: Extract and normalize the three-dimensional adjustment potential features of equipment power ramp-up, energy storage charging and discharging range, and adjustment penalty cost within the primary cluster to generate a three-dimensional feature vector. Based on the three-dimensional potential characteristics, three types of sub-clusters with high, medium and low flexibility are divided. The upper and lower limits of adjustable power, comprehensive adjustment cost and average marginal benefit in each time period are aggregated to obtain the equivalent scheduling parameters of the sub-clusters.
[0013] Preferably, the process for generating equivalent scheduling parameters for each flexible level sub-cluster is as follows: Three types of extreme points—climbing rate, charge / discharge range, and adjustment cost—were selected as the initial centers of high / medium / low flexibility sub-clusters. The weighted Euclidean distance from the device to the three types of centers is calculated to complete the initial hard clustering. By summing the conservatively adjustable residual power of all devices in the same sub-cluster, the lower limit of the adjustable power of the sub-cluster is obtained. The upper limit of the adjustable power of the sub-cluster is obtained by summing the residual power of all devices in the sub-cluster. A quadratic curve is fitted to the adjustment cost of equipment within the sub-cluster to generate a comprehensive adjustment cost function. The average marginal revenue per unit of surplus electricity of all devices in the sub-cluster is calculated as the average marginal revenue. Adjustable power upper and lower limits, comprehensive adjustment cost function, and average marginal revenue are unified as equivalent scheduling parameters for sub-clusters.
[0014] Preferably, the improved particle swarm optimization algorithm specifically includes: constructing particle perturbation input based on a weighted set of typical scenarios output from the source-load-market model; constructing a comprehensive fitness based on scenario probability weighting; superimposing probability perturbation terms in particle position updates based on the coupling characteristics of multi-source variables; setting hard constraints based on a three-layer confidence-adjustable residual power boundary and configuring an out-of-bounds backoff mechanism; adding a robust penalty term based on the variance of the full-scenario revenue fluctuation; iteratively updating individual and globally optimal particles and selecting high-return, low-fluctuation, robust optimal solutions; terminating the iteration based on the convergence condition; and outputting the optimal residual power scheduling scheme for each flexible level sub-cluster that adapts to multiple stochastic operating conditions.
[0015] (III) Beneficial Effects Compared with the prior art, the beneficial effects of the present invention are: This invention overcomes the limitations of traditional independent prediction models by constructing a multidimensional joint cumulative probability distribution (R-Teng Copula) to deeply explore the nonlinearity and tail correlations among five variables: photovoltaic, wind power, rigid / flexible loads, and market electricity prices. It also introduces Latin hypercube sampling and weighted scenario reduction techniques to generate a set of typical scenarios with probability weights. Based on this, a dedicated time-series prediction network and heavy-tailed distribution residual fitting are combined to output a multidimensional probabilistic feature vector for each device, fusing "expected value + three-layer confidence boundary". This mechanism can accurately quantify the residual power fluctuation range under the coupling influence of multiple random sources, providing a highly reliable data foundation for subsequent optimized scheduling and effectively avoiding the risk of revenue deviation caused by prediction inaccuracies.
[0016] This invention employs a collaborative strategy of "first-level global weighted adaptive scenario clustering" and "second-level potential hierarchical clustering within the scenario." First-level clustering, based on a weighted composite distance metric, fuzzily aggregates devices within geographical partitions into several power consumption scenario clusters and generates equivalent encapsulation parameters, effectively reducing the dimensionality of scheduling variables. Second-level clustering delves deeper into the scenario, using three-dimensional potential features such as power ramp-up, energy storage charging / discharging range, and adjustment costs to finely divide high / medium / low flexibility level sub-clusters and extract equivalent scheduling parameters. This two-layer architecture avoids the curse of dimensionality problem of direct optimization of large-scale devices while fully preserving the differentiated adjustment capabilities of different clusters, greatly improving optimization efficiency and engineering operability.
[0017] This invention abandons single deterministic optimization, directly using a weighted set of typical scenarios as the perturbation input for the particle swarm optimization algorithm, and constructs a comprehensive fitness function based on scenario probability weighting. Simultaneously, it superimposes a probabilistic perturbation term during location updates, utilizes a three-layer confidence boundary to set hard constraints and an out-of-bounds backoff mechanism, ensuring that the solution always remains within the safe and feasible domain. Furthermore, it innovatively introduces a robust penalty term based on the variance of revenue fluctuations across all scenarios, guiding the algorithm to iteratively select robust optimal solutions with high returns and low volatility. The final output scheduling scheme can simultaneously satisfy the dual objectives of "maximizing scheduling revenue" and "minimizing distribution network power supply balance" under multiple stochastic operating conditions, significantly enhancing the virtual power plant's resilience and operational stability in the face of complex market environments and extreme scenario fluctuations. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0019] Figure 1 This is a schematic diagram of a virtual power plant revenue prediction and optimization method based on multi-scenario probabilities. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0022] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities. The method includes: Acquire basic data and operational constraints; in this embodiment, specifically: construct a multi-source heterogeneous raw dataset based on four types of raw data sources: distributed power source side environmental parameters and unit operation ledgers, user side graded load ledgers, electricity market full-category electricity price assessment data, and extreme environmental condition markings. Missing monitoring samples within the dataset are completed using a neighborhood time-series interpolation algorithm, and abnormal samples of output and load fluctuations caused by equipment failures are removed based on the 3σ anomaly discrimination criterion. All coupled variables are uniformly normalized to the same numerical range; Finally, the full dataset was labeled hierarchically according to equipment type, power grid operating conditions, and peak and off-peak electricity prices, and a standardized coupled time-series dataset with operating condition labels was output as subsequent input.
[0023] Based on basic data and operational constraints, multi-scenario hierarchical residual power prediction of source-load-market coupling probability is used to obtain the multi-scenario residual power probability distribution and multi-dimensional probability feature vector of each device. In this embodiment, the process of generating the multi-scenario residual power probability distribution and multi-dimensional probability feature vector of each device is as follows: The collected basic data is preprocessed to obtain a standardized set of coupled variables. and layered labels; Photovoltaic power output timing, Wind power output timing, Rigid load timing, Flexible load timing, Market electricity price time series; all variables are uniformly mapped to the [0,1] standard distribution domain for subsequent modeling.
[0024] Based on the set of coupled variables, the marginal distributions of five types of variables, namely photovoltaic power output, wind power output, rigid load, flexible load and market electricity price, are fitted respectively, and the R-Teng Copula multidimensional joint cumulative probability distribution is constructed to quantify the joint occurrence probability. Among them, the multidimensional joint cumulative probability distribution of R-vine Copula is constructed, and the joint occurrence probability is quantified to include: Based on five types of time-series data—standardized photovoltaic (PV) power output, wind power output, rigid load, flexible load, and market electricity price—marginal probability distributions of each variable are fitted, and each marginal distribution is converted into a cumulative probability sequence over a unified distribution domain. In this embodiment, maximum likelihood estimation (MLE) is used to fit the corresponding parameterized distributions for each of the five types of time-series data: for each type of time-series variable, multiple candidate parameter distributions are preset. Maximum likelihood estimation is used to solve for all parameters of the distribution, including location, scale, shape / degrees of freedom; the optimal marginal distribution is selected by the triple goodness-of-fit index of AIC, BIC, and KS tests; the original time series is mapped to the uniform probability domain of [0,1] by probability integral transformation (PIT) to generate a cumulative probability sequence.
[0025] On the unified distribution domain, the R-vine Copula is used to construct a joint correlation structure among the five types of variables through a multi-layered tree structure. Each edge of each layer of the tree adaptively selects the optimal fit from a variety of candidate binary Copulas. All correlation parameters are solved through stepwise maximum likelihood estimation, resulting in a five-dimensional joint cumulative probability distribution characterizing the nonlinear and asymmetric tail correlation between wind and solar power output, load, and electricity price. Specifically, the core decomposition formula of the R-vine Copula involves a 4-layer vine tree with 5 variables, and the joint distribution is as follows: ; The number of vine layers. Let C be the set of edges for each layer, and let C be a conditional binary Copula.
[0026] Based on the joint cumulative probability distribution, the original time series scenarios are generated by Latin hypercube sampling. A weighted objective function is constructed based on the adjustable residual power fluctuation variance and estimated revenue of each time period under each scenario. The original scenarios are reduced to obtain a set of typical scenarios and the occurrence probability weight of each typical scenario. Based on a set of typical scenarios, a suitable time-series prediction network is used to predict the power output and load of each device in each typical scenario and time period. Based on the predicted power output and load of each device, the rigid self-consumption load is deducted from the power output, and the absorption margin released by the reduction of flexible load is added as a positive increment to obtain the expected adjustable surplus power of each device in each typical scenario and time period. The specific process for generating the expected adjustable residual power of each device in each typical scenario and time period is as follows: Based on the predicted power generation output of each device in each typical scenario and time period, the rigid self-consumption load that is bound to the device and cannot be reduced or transferred is taken as the local consumption part that must be prioritized and deducted from the predicted power generation output. Based on the forecast results of flexible load, the additional absorption space released by the reduced electricity consumption of the flexible load bound to the equipment in the current period is obtained. The additional absorption space is used as the positive increment of adjustable surplus power and added to the remaining power generation output after deducting rigid self-consumption load. The accumulated result is taken as the expected adjustable residual power of the device during this typical scenario and time period. In this embodiment, the corresponding formula for calculating the adjustable residual power is: ; Forecast the unit's output. For rigid loads, This provides a flexible and releaseable absorption margin.
[0027] Based on the historical prediction residuals of each device, a residual sequence is constructed, the probability distribution of the residuals is fitted and the distribution parameters are solved. With the expected value of adjustable residual power as the benchmark, the product of the residual distribution quantiles and the scaling parameters at different confidence levels is superimposed to obtain the conservative lower limit of adjustable residual power and the aggressive upper limit of adjustable residual power, forming a three-layer confidence adjustable residual power boundary. The generation process of the three-layer confidence-adjustable residual current boundary is as follows: Using the deviation between the predicted and actual output values of each device in a historical period as residual samples, a device-specific historical residual sequence is constructed. The historical residual sequence is fitted with a heavy-tailed probability distribution, and the location parameters, scale parameters, and degrees of freedom of the distribution are solved. Using the adjustable residual charge expectation value as a benchmark, the scale parameter is multiplied by the quantile of the thick-tailed probability distribution at a preset low confidence level to obtain the downward offset. The benchmark value is then subtracted from the downward offset to obtain the lower limit of the adjustable residual charge. Multiply the scale parameter by the quantile of the thick-tailed probability distribution at a preset high confidence level to obtain the upward offset. Add the upward offset to the benchmark value to obtain the adjustable residual power upper limit. The baseline value, the lower limit of adjustable residual power, and the upper limit of adjustable residual power together constitute the three-layer confidence adjustable residual power boundary of each device in each typical scenario and time period.
[0028] Based on the adjustable residual power expectation, the three-layer confidence adjustable residual power boundary, and the occurrence probability weights of typical scenarios, a probabilistic feature vector is constructed for each device and each typical scenario, resulting in a multi-dimensional probabilistic feature vector matrix.
[0029] Based on the multidimensional probabilistic feature vector, a first-level global residual power weighted adaptive scenario clustering is performed to obtain a first-level power consumption scenario cluster and equivalent encapsulation parameters for each scenario. In this embodiment, the specific process for generating the primary power consumption scenario cluster and the equivalent encapsulation parameters for each scenario is as follows: Based on the multidimensional probability feature vectors of each device in each typical scenario and the occurrence probability weights of each typical scenario, the feature vectors of the same device in all typical scenarios are weighted and fused according to the corresponding scenario probabilities to obtain the global comprehensive feature vector of each device. The specific process for generating the standardized comprehensive feature vectors of devices within each partition is as follows: Based on the global comprehensive feature vector of each device, all devices are divided into corresponding geographical partitions according to the distribution network feeder or transformer area affiliation. Devices within the same geographical partition share the distribution network power flow transmission capacity constraints of that partition. Within each geographical region, based on the first dimension representing the average adjustable surplus power in the global comprehensive feature vector, devices with a value not greater than zero in the first dimension are removed, and devices with the potential for external dispatch are retained as the regional clustering objects. For the feature values of each retained device in each dimension, calculate the mean and standard deviation of each dimension in all partitioned devices. Subtract the mean of the corresponding dimension from the feature value of each device in each dimension and divide by the standard deviation of the corresponding dimension to obtain the standardized comprehensive feature vector of the devices in each partition.
[0030] Based on the global comprehensive feature vector of each device, all devices are grouped according to the geographical partition of the distribution network. Devices with no adjustable residual power potential are removed from each partition. The features of the remaining devices are normalized to obtain the standardized comprehensive feature vector of the devices in each partition. Based on the standardized comprehensive feature vector, a weighted composite distance metric function for optimizing surplus power dispatch revenue is constructed. The weighted composite distance measures the degree of comprehensive dispatch difference between devices by the square root of the sum of the weighted square differences between two devices on each feature dimension. Based on the weighted composite distance metric function, through fuzzy clustering iterative calculation, clusters are adaptively split or merged during the clustering process to automatically determine the number of scene clusters, and each device is assigned a fuzzy membership degree to each scene cluster until the clustering loss function converges, thus obtaining the first-level power consumption scene clusters and the fuzzy membership degree of each device in each scene cluster. The specific process for generating the fuzzy membership degree of each device in each primary power consumption scenario cluster is as follows: Based on the standardized comprehensive feature vectors of the devices in each partition, the extreme points of the feature vector magnitudes are selected from the device feature space of each partition as the initial candidate cluster centers, and overlapping candidate centers with a weighted composite distance less than a preset separation threshold are eliminated, and the initial number of clusters and the initial cluster centers of each partition are automatically generated. Based on the current weighted composite distance between each cluster center and each device, and in accordance with the constraint that the sum of the membership degrees of each device to all clusters is normalized to one, the fuzzy membership degree of each device to each cluster is updated. The closer the device is to a certain cluster, the higher its membership degree to that cluster. Based on the updated fuzzy membership degrees of each device, the position of each cluster center is updated by averaging the standardized comprehensive feature vector of each device according to the power of its membership degree to the cluster. Based on the updated weighted composite distance between the cluster center and each device and the updated fuzzy membership degree, calculate the weighted loss function value of the current clustering; After each iteration, the weighted dispersion of the feature values of the devices within each cluster relative to the cluster center is calculated. When the weighted dispersion of a cluster exceeds a preset splitting threshold, the cluster is split into two new clusters. The weighted composite distance between any two cluster centers is calculated. When the distance between the two cluster centers is less than a preset merging threshold, the two clusters are merged into one cluster. Repeat the above operation until the change in the loss function of two consecutive iterations is less than the preset termination threshold, and stop the iteration to obtain the final power consumption scenario cluster division of the current partition and the fuzzy membership degree of each device in each scenario cluster.
[0031] Based on the fuzzy membership of the first-level power consumption scenario clusters and each device, the adjustable surplus power, unit revenue, adjustment cost and response delay of the devices in each scenario cluster are fuzzily weighted and summarized to generate the equivalent encapsulation parameters of each scenario cluster.
[0032] Based on the primary power consumption scenario cluster and equivalent encapsulation parameters, secondary potential hierarchical clustering within the scenario is used to generate multiple flexible level sub-clusters and sub-cluster equivalent scheduling parameters. The execution process of the secondary potential hierarchical clustering within the scenario is as follows: Extract and normalize the three-dimensional adjustment potential features of equipment power ramp-up, energy storage charging and discharging range, and adjustment penalty cost within the primary cluster to generate a three-dimensional feature vector. Based on the three-dimensional potential characteristics, three types of sub-clusters with high, medium and low flexibility are divided. The upper and lower limits of adjustable power, comprehensive adjustment cost and average marginal benefit in each time period are aggregated to obtain the equivalent scheduling parameters of the sub-clusters.
[0033] The specific process for generating the equivalent scheduling parameters for each sub-cluster is as follows: Based on the three-dimensional adjustment potential feature vector of each device, three points are selected in the three-dimensional feature space: the maximum ramp rate, the middle of the feasible charging and discharging range, and the highest adjustment penalty cost. These points are used as the initial cluster centers of the high-flexibility sub-cluster, the medium-flexibility sub-cluster, and the low-flexibility sub-cluster, respectively. Calculate the weighted Euclidean distance from each device to the three initial cluster centers, assign each device to the sub-cluster corresponding to the initial cluster center with the smallest distance, complete the initial quadratic hard partition, and obtain the three initial sub-clusters and the set of devices contained therein; Based on the conservative adjustable residual power lower limit of each device in each sub-cluster under each typical scenario and time period, the conservative lower limits of all devices in the same sub-cluster are summed in the same time period to serve as the adjustable power lower limit of the sub-cluster in that time period. Based on the aggressive adjustable residual power limit of each device in each sub-cluster under each typical scenario and time period, the aggressive limits of all devices in the same sub-cluster are summed in the same time period to serve as the adjustable power limit of the sub-cluster in that time period. Based on the flexible adjustment cost of each device within each sub-cluster, a quadratic curve fitting is performed on the adjustment cost of each device to obtain the comprehensive adjustment cost function of the sub-cluster. Based on the unit surplus power marginal revenue of each device in each sub-cluster, the average marginal revenue of all devices in the same sub-cluster is calculated as the average marginal revenue of that sub-cluster. The adjustable power upper and lower limits, the comprehensive adjustment cost function, and the average marginal revenue of each sub-cluster are used together as the equivalent scheduling parameters of each sub-cluster.
[0034] An optimal fitness function is constructed with the objectives of maximizing dispatch revenue and minimizing power supply balance in the distribution network. Combined with power grid operation constraints, an improved particle swarm optimization algorithm is used to iteratively find the optimal surplus power dispatch scheme and power supply balance quantification index for each flexible level sub-cluster.
[0035] The improved particle swarm optimization algorithm specifically includes: constructing particle perturbation inputs based on a weighted set of typical scenarios output from the source-load-market model; constructing a comprehensive fitness based on scenario probability weighting; superimposing probability perturbation terms in particle position updates based on the coupling characteristics of multi-source variables; setting hard constraints and configuring an out-of-bounds backoff mechanism based on a three-layer confidence-adjustable residual power boundary; adding a robust penalty term based on the variance of full-scenario revenue fluctuations; iteratively updating individual and globally optimal particles and selecting high-return, low-fluctuation robust optimal solutions; terminating iteration based on convergence conditions; and outputting optimal residual power scheduling schemes for each flexible level sub-cluster adapted to multiple random operating conditions. Specifically, it reads all typical time-series scenarios generated by the source-load-market coupling probability model and the corresponding occurrence probability weights for each scenario; synchronously reads the three-layer confidence-adjustable residual power interval obtained by fitting the thick-tailed residuals for each device; and uses the above two patent-unique outputs as the basic input data for the particle swarm optimization; the particle encoding dimension corresponds one-to-one with the number of flexible sub-clusters and scheduling time periods under each first-level scenario cluster; and a single particle vector fully carries the power to be optimized for scheduling in all sub-clusters. Abandoning the traditional method of calculating deterministic returns in a single scenario, this paper uses the probability weights of each typical scenario as weighting coefficients, and superimposes the virtual power plant dispatch revenue and distribution network imbalance indicators under all scenarios. At the same time, a robust penalty term is introduced, the size of which is dynamically determined by the degree of fluctuation of the particle's dispatch revenue under all scenarios. The more violent the revenue fluctuation, the higher the penalty weight, thereby suppressing large fluctuations in revenue under extreme conditions.
[0036] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.
[0037] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0038] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.
[0039] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0040] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.
[0041] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0042] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0043] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities, characterized in that, The method includes: Acquire basic data and grid operation constraints from the dimensions of distributed power sources, loads, and the electricity market; After preprocessing the basic data, a source-load-market coupled probability model is constructed, multi-scenario hierarchical surplus power prediction is performed, and the multi-scenario surplus power probability distribution and multi-dimensional probability feature vector corresponding to each device are output. Based on multidimensional probabilistic feature vectors, a first-level global residual power weighted adaptive scenario clustering is carried out to divide the first-level power consumption scenario clusters and generate the equivalent encapsulation parameters corresponding to each scenario. Based on the primary power consumption scenario cluster and equivalent encapsulation parameters, secondary potential hierarchical clustering within the scenario is used to generate multiple flexible level sub-clusters and sub-cluster equivalent scheduling parameters. An optimal fitness function is constructed with the objectives of maximizing dispatch revenue and minimizing power supply balance in the distribution network. Combined with power grid operation constraints, an improved particle swarm optimization algorithm is used to iteratively find the optimal surplus power dispatch scheme and power supply balance quantification index for each flexible level sub-cluster.
2. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 1, characterized in that, The generation process of the multi-scenario residual power probability distribution and multi-dimensional probability feature vector of each device is as follows: The collected basic data is processed by filling in missing values, removing anomalies, and normalizing to obtain a standardized set of coupled variables and complete the labeling of working condition hierarchical labels. Marginal distributions were fitted to five types of variables: photovoltaic power output, wind power output, rigid load, flexible load, and market electricity price. A multidimensional joint cumulative probability distribution of R-Teng Copula was constructed to quantify the joint occurrence probability of the variables. The original time series scenarios are generated by Latin hypercube sampling based on the joint cumulative probability distribution. The scenarios are reduced by constructing a weighted objective function based on the adjustable residual power fluctuation variance and the estimated revenue, resulting in a set of typical scenarios with probability weights. A dedicated time-series forecasting network is selected to adapt to various equipment types, and outputs the power generation and load forecast values for each device under various typical scenarios; The expected value of adjustable surplus power for each device is calculated by subtracting rigid self-consumption load from the power generation forecast and adding the absorption margin released by the reduction of flexible load. Based on the historical prediction deviation of the equipment, a residual sequence is constructed, and a heavy-tailed distribution is used to fit the residual parameters. Combined with high and low confidence quantiles, a three-layer confidence adjustable residual current upper and lower boundaries are generated. By integrating adjustable residual power expectation, three-layer confidence boundary, and probability weights for typical scenarios, a multi-dimensional probabilistic feature vector matrix is constructed in the device-scenario dimension.
3. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 2, characterized in that, The process of constructing the R-vine Copula multidimensional joint cumulative probability distribution and quantifying the joint occurrence probability is as follows: The standardized time-series data of photovoltaic, wind power, rigid load, flexible load, and electricity price are respectively fitted with marginal distributions and transformed into a unified probability domain to generate cumulative probability sequences. The R-vine Copula decomposition of five types of variable relationships is adopted using a multi-layered tree structure, and the optimal binary Copula function is adaptively selected at each layer edge. By solving all Copula-related parameters through stepwise maximum likelihood estimation, a five-dimensional joint probability distribution that can characterize the nonlinearity and tail-free correlation of variables is constructed.
4. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 3, characterized in that, The calculation process for the expected adjustable residual power of each device under various typical scenarios is as follows: The rigid self-consumption load that is tied to the equipment and cannot be reduced or transferred is taken as the mandatory consumption amount and deducted from the unit's power generation forecast; Extracting flexible loads can reduce the additional absorption margin corresponding to the capacity as a positive increment; The remaining output after deducting rigid loads is added to the absorption margin to obtain the expected value of adjustable surplus power of the equipment for the current period.
5. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 4, characterized in that, The process for generating the three-layer confidence-adjustable residual current boundary is as follows: Extract the deviation between the historical actual output and the predicted output of the equipment, and construct a unique residual sample sequence for the equipment. A heavy-tailed probability distribution is used to fit the residuals, and the distribution location, scale, and degrees of freedom parameters are solved. The lower limit of adjustable residual power is obtained by superimposing the low confidence scale quantile offset on the benchmark value of adjustable residual power. The upper limit of adjustable residual power is obtained by superimposing the high-confidence scale quantile offset on the benchmark of the expected value of adjustable residual power. The expected value, lower limit, and upper limit together constitute the three-layer confidence adjustable residual current boundary of the device.
6. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 5, characterized in that, The process of constructing the primary power consumption scenario cluster and the equivalent encapsulation parameters for each scenario is as follows: The feature vectors of all typical scenarios of a single device are weighted and fused according to the probability of the corresponding scenario to obtain the global comprehensive feature vector of the device. All equipment is divided into geographical zones according to the distribution network feeder / transformer area, and equipment with no adjustable residual power potential is removed. The characteristics of the remaining equipment are normalized in each dimension to generate a zone-standardized comprehensive feature vector. Construct a weighted composite distance metric function oriented towards scheduling benefits to quantify scheduling differences between devices; Fuzzy iterative clustering is performed based on weighted composite distance, adaptively splitting and merging clusters, outputting the fuzzy membership degree of each device, and obtaining a first-level power consumption scenario cluster after convergence. Based on the fuzzy membership of the devices, the adjustable surplus power, unit revenue, adjustment cost, and response latency within the cluster are fuzzily weighted and summarized to generate equivalent encapsulation parameters for each scenario.
7. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 6, characterized in that, The iterative calculation process of fuzzy membership degree between the primary power consumption scenario cluster and the equipment is as follows: The extreme values of the feature space modulus of the partition are selected as the initial cluster centers, and overlapping centers that are too close are removed to generate the initial clusters. Update the normalized fuzzy membership degree based on the weighted composite distance between the device and each cluster; The coordinates of each cluster center are updated based on a weighted average of membership degrees. Calculate the weighted loss function for this round of clustering; Perform cluster splitting and merging operations based on the dispersion threshold and the cluster spacing threshold, respectively; Repeat the iteration until the change in loss is less than the termination threshold, and output the membership degree of the stable scene cluster and device.
8. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 7, characterized in that, The execution process of secondary potential hierarchical clustering within the scenario is as follows: Extract and normalize the three-dimensional adjustment potential features of equipment power ramp-up, energy storage charging and discharging range, and adjustment penalty cost within the primary cluster to generate a three-dimensional feature vector. Based on the three-dimensional potential characteristics, three types of sub-clusters with high, medium and low flexibility are divided. The upper and lower limits of adjustable power, comprehensive adjustment cost and average marginal benefit in each time period are aggregated to obtain the equivalent scheduling parameters of the sub-clusters.
9. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 8, characterized in that, The process of generating equivalent scheduling parameters for each flexible level sub-cluster is as follows: Three types of extreme points—climbing rate, charge / discharge range, and adjustment cost—were selected as the initial centers of high / medium / low flexibility sub-clusters. The weighted Euclidean distance from the device to the three types of centers is calculated to complete the initial hard clustering. By summing the conservatively adjustable residual power of all devices in the same sub-cluster, the lower limit of the adjustable power of the sub-cluster is obtained. The upper limit of the adjustable power of the sub-cluster is obtained by summing the residual power of all devices in the sub-cluster. A quadratic curve is fitted to the adjustment cost of equipment within the sub-cluster to generate a comprehensive adjustment cost function. The average marginal revenue per unit of surplus electricity of all devices in the sub-cluster is calculated as the average marginal revenue. Adjustable power upper and lower limits, comprehensive adjustment cost function, and average marginal revenue are unified as equivalent scheduling parameters for sub-clusters.
10. The method for predicting and optimizing the revenue of a virtual power plant based on multi-scenario probabilities according to claim 9, characterized in that, The improved particle swarm optimization algorithm specifically includes: constructing particle perturbation inputs based on a weighted set of typical scenarios output from the source-load-market model; constructing a comprehensive fitness based on scenario probability weighting; superimposing probability perturbation terms in particle position updates based on the coupling characteristics of multi-source variables; setting hard constraints and configuring an out-of-bounds backoff mechanism based on a three-layer confidence-adjustable residual power boundary; adding a robust penalty term based on the variance of the overall scenario revenue fluctuation; iteratively updating individual and globally optimal particles and selecting high-return, low-fluctuation, robust optimal solutions; terminating the iteration based on the convergence condition; and outputting optimal residual power scheduling schemes for each flexible level of sub-clusters that adapt to multiple stochastic operating conditions.
