A stgnn prediction-based robust optimization scheduling method for distribution of optical storage and charging virtual power plant

Abstract

The application discloses a kind of light storage fills virtual power plant distribution robust optimization scheduling method based on STGNN prediction, by constructing virtual power plant physical topology adjacency matrix, utilize space-time graph neural network to carry out joint prediction to each node future multi-time period photovoltaic output and load demand;Based on the prediction result and historical real operation data, a transformer station area node multi-scenario prediction error experience sample library is constructed, and an experience distribution, a net disturbance support set and a corresponding Wasserstein fuzzy set are constructed during rolling scheduling;Taking energy storage charging and discharging power, electric vehicle V2G charging and discharging power and frequency modulation market declaration capacity as control variables, a virtual power plant operation model is established with the maximum comprehensive net profit as the target;Combined with CVaR risk constraint, a distribution robust model predictive control model is constructed and solved, realizing the rolling optimization scheduling of virtual power plant multi-resource.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant optimization scheduling technology, and particularly relates to a method for optimizing the scheduling of a photovoltaic-storage-charging virtual power plant based on STGNN prediction. Background Technology

[0002] With the construction of new power systems and the large-scale integration of distributed energy sources such as distributed photovoltaics, energy storage devices, and electric vehicles, the safe and stable operation of distribution network areas has been greatly challenged. Virtual power plants, as a key technology capable of aggregating and coordinating various resources such as distributed energy, energy storage devices, flexible loads, and electric vehicles, can achieve resource aggregation for participation in the energy market, ancillary service market, and distribution network operation regulation through unified modeling, centralized optimization, and collaborative control. This is of great significance for improving the flexibility of the power system, promoting the consumption of new energy sources, and enhancing the safe operation level of the distribution network.

[0003] However, due to the highly random and volatile nature of source and load power, accurately predicting and managing the internal state of virtual power plants remains a challenge. Currently, traditional time-series forecasting methods often treat each node within the distribution network in isolation, neglecting the spatiotemporal coupling effect based on physical topology between photovoltaic, energy storage, and charging networks, thus limiting prediction accuracy. Furthermore, in dispatching decisions dealing with uncertainty, the Wasserstein radius of traditional distributed bar optimization is fixed, leading to overly conservative dispatching results that severely sacrifice the system's economic benefits. Moreover, the large-scale integration of distributed energy sources poses challenges to the safe operation of distribution areas. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned technical problems by providing a distributed bar optimization scheduling method for virtual power plants based on STGNN prediction. This method considers the coordination between distributed assets and the main grid operation boundary, establishes a virtual power plant optimization scheduling model that considers multi-market participation, and achieves physical coupling between massive multi-source heterogeneous assets and the distribution network through transformer substation nodes. By transforming the residuals predicted by the spatiotemporal graph neural network multi-node cluster into multi-scenario empirical samples, and combining the measure concentration theorem to dynamically calculate the Wasserstein radius and construct conditional risk value constraints, this method addresses the risk of transformer substation exceeding limits caused by uncertainties in the source-load environment. Through this method, this invention not only overcomes the conservatism of traditional distributed bar optimization scheduling but also strictly prevents transformer exceeding-limit risks, ensuring the safe operation of the system.

[0005] Technical Solution: To achieve the above-mentioned objectives, this invention proposes a method for optimizing the scheduling of virtual power plants based on STGNN prediction using a distributed bloater system. This method includes the following steps:

[0006] Step 1: Obtain the network topology parameters and historical operation scenario data of the virtual power plant; the network topology parameters include transformer capacity, photovoltaic-storage-charging station node connection relationship, distributed photovoltaic capacity, energy storage equipment parameters, and electric vehicle charging pile configuration; the historical operation scenario data includes the historical load demand, historical photovoltaic output, grid time-of-use electricity price, and frequency regulation market price data of each node.

[0007] Step 2: Construct a virtual power plant physical topology adjacency matrix based on the node connection relationship of the photovoltaic-storage-charging station. Using the historical load demand and historical photovoltaic output data of each node as input, extract the spatiotemporal coupling features using a spatiotemporal graph neural network STGNN containing three spatiotemporal blocks of time-space-time, and output the photovoltaic output prediction sequence and load demand prediction sequence of each photovoltaic-storage-charging station node in the future multiple time steps.

[0008] Step 3: Compare the predicted sequence with historical real operating data, extract the real prediction error of photovoltaic and load and convert it to the corresponding transformer substation node, and build a multi-scenario prediction error experience sample library for transformer substation node;

[0009] Step 4: Based on the sample matrix and preset safety confidence level parameters obtained from the sample library in Step 3 at each rolling scheduling time, construct the empirical distribution and net disturbance error support set of the transformer substation nodes in the current prediction time domain, and dynamically calculate the Wasserstein radius using the measure set theorem, thereby generating the time series Wasserstein fuzzy set corresponding to each transformer substation node.

[0010] Step 5: Based on the parameters and data obtained in Step 1, with energy storage charging and discharging power, electric vehicle V2G charging and discharging power and frequency regulation market application capacity as control variables, and taking into account the operation constraints of photovoltaic-storage charging stations and transformer capacity constraints, and with the objective function of maximizing the comprehensive profit of the virtual power plant in the energy market and frequency regulation market, establish a deterministic predictive control model for the virtual power plant.

[0011] Step 6: Introduce the prediction error empirical sample library from Step 3 and the dynamic Wasserstein fuzzy set from Step 4 into the deterministic model prediction and control model from Step 5, and construct CVaR risk constraints to establish a sub-Bruker model prediction and control model with a safety risk budget.

[0012] Step 7: Collect the initial state of charge of each energy storage device and electric vehicle in the virtual power plant in real time, and combine it with the real-time updated prediction sequence, price window and dynamic Wasserstein radius as input parameters for the sub-Bruker bar model predictive control model in Step 6 to obtain the optimal scheduling strategy set within the future time window; execute the charging, discharging and frequency regulation instructions of the first time step of the strategy set, update the state, and roll to the next scheduling time to solve, so as to realize the optimal scheduling of the virtual power plant.

[0013] Furthermore, in step 2, the spatiotemporal graph neural network model containing three spatiotemporal blocks (time-space-time) is used as follows:

[0014] (1) Initial feature mapping and sine and cosine position coding

[0015] (A-1)

[0016] (A-2)

[0017] (A-3)

[0018] (A-4)

[0019] In the formula, This represents the original input tensor composed of the historical load demand and photovoltaic output of each node; This represents the feature tensor mapped to the hidden layer dimension; This represents the linear projection weight matrix of the input layer; This represents the bias vector of the input layer; Represents the absolute position encoding matrix; Indicates the absolute position encoding matrix at the th The first time step Values ​​in each feature dimension; This indicates that the absolute position encoding matrix is ​​in the th... The first time step Values ​​in each feature dimension; This represents the dimension of the hidden layer features within the model network; Indicates the index of the time step in the sequence; Indicates the feature dimension index; This represents the initial spatiotemporal feature tensor of each node, which incorporates location encoding information.

[0020] (2) First-layer temporal feature extraction

[0021] (A-5)

[0022] (A-6)

[0023] (A-7)

[0024] (A-8)

[0025] (A-9)

[0026] (A-10)

[0027] (A-11)

[0028] (A-12)

[0029] In the formula, , and They represent the first The query vector, key vector, and value vector of each attention head; , , They represent the first The learnable weight matrix used to generate queries, keys, and values ​​in each attention head; Indicates the total number of nodes; Indicates that for a given first... The corresponding time series extracted from each node. ; Indicates the attention head index; Indicates the first Local temporal feature representation of the attention head output; This represents the exponential normalization function; Represents the feature dimensions of each attention head; This indicates a high level of attention from multiple parties. This indicates matrix concatenation along the feature dimension; This represents the total number of heads receiving multi-head attention. This represents the output weights of the multi-head attention concatenation; This represents the intermediate feature tensor after multi-head self-attention and the first residual connection; Indicates the layer normalization function; This represents a feedforward neural network; Represents a non-linear activation function; and This is the weight vector; and For bias terms; This indicates the node after the first layer of time features has been extracted. The characteristic output;

[0030] (3) Spatial feature aggregation

[0031] (A-13)

[0032] (A-14)

[0033] (A-15)

[0034] In the formula, The original adjacency matrix representing the physical connection relationships of virtual power plant nodes; It is the identity matrix; This represents the adjacency matrix after introducing self-loops; Represents the symmetric normalized adjacency matrix; It is a diagonal degree matrix; This represents the shared weight matrix for feature propagation between nodes in a graph convolutional layer; This represents the bias vector of the graph convolutional layer; Indicates the regularization method; Indicates time step The feature input of all nodes after the first layer of temporal feature extraction at any given time; Indicates time step The node characteristics after spatial topological aggregation at any given moment;

[0035] (4) Second-layer temporal feature extraction and global prediction

[0036] (A-16)

[0037] (A-17)

[0038] In the formula, This indicates the output obtained by extracting temporal features again using the same multi-head self-attention and feedforward network structure as the first layer; This represents the fusion feature of the final output of the entire spatiotemporal block; Indicates the length of the input history sequence; This represents the comprehensive spatiotemporal characteristic state of all nodes at the last moment of the current historical sequence; and This represents the weight matrix of the output mapping layer; and This represents the bias vector of the output mapping layer; This represents the predicted values ​​of the output photovoltaic power and the load.

[0039] Furthermore, in step 3, the method for constructing a multi-scenario prediction error experience sample library for transformer substation nodes is as follows:

[0040] (A-18)

[0041] (A-19)

[0042] (A-20)

[0043] (A-21)

[0044] In the formula, Indicates an index of historical experience error scenarios; Indicates the number of valid historical error samples; Indicates the index of photovoltaic and energy storage charging stations; Indicates the first In the scenario, the first The actual load of each site; Indicates the first In the scenario, the first Forecast load for each site; Indicates the first In the scenario, the first The actual photovoltaic output of each site; Indicates the first In the scenario, the first Predicted photovoltaic output for each site; Indicates the first In the scenario, the first Load forecast residuals for each site; Indicates the first In the scenario, the first Photovoltaic prediction residuals at each site; Indicates the transformer substation index; This indicates that it belongs to the transformer substation area. A collection of power stations; Indicates transformer substation area In the Net prediction error samples in each scenario; Indicates transformer substation area A multi-scenario prediction error empirical sample library.

[0045] Furthermore, the method for step 4 is as follows:

[0046] (A-22)

[0047] (A-23)

[0048] (A-24)

[0049] (A-25)

[0050] (A-26)

[0051] (A-27)

[0052] (A-28)

[0053] (A-29)

[0054] (A-30)

[0055] (A-31)

[0056] (A-32)

[0057] (A-33)

[0058] (A-34)

[0059] (A-35)

[0060] (A-36)

[0061] In the formula, Indicates the current rolling scheduling time; This represents the total number of time steps in the look-ahead prediction time domain; This represents the time step index within the look-ahead prediction time domain; Indicates the current rolling scheduling time Starting from the first The scheduling period corresponding to each prediction step; To extract from the historical error sample database Extracted corresponding transformer substations At the current rolling scheduling moment Corresponding future Subsample matrix of each prediction step ; Indicates transformer substation area In the scene Next The net perturbation error sample corresponding to each prediction step; Indicates transformer substation area In the Empirical distribution under each prediction step; Indicates the location of the net disturbance error sample Dirac measure at the location; Indicates the first The first prediction step Load forecast values ​​for each site; Indicates the first The first prediction step The historical load limit for each site; Indicates the first The first prediction step Lower bound of load residual for each site; Indicates the first The first prediction step Upper bound of load residuals for each site; Indicates transformer substation node Lower bound of the aggregated load residual; Indicates transformer substation node Upper bound of the aggregated load residual; Indicates the first The first prediction step Photovoltaic forecast values ​​for each site; Indicates the first The first prediction step The upper limit of historical photovoltaic experience for each site; Indicates the first The first prediction step Lower bound of photovoltaic residuals at each site; Indicates the first The first prediction step Upper bound of photovoltaic residuals at each site; Indicates transformer substation node Lower bound of the photovoltaic residual after aggregation; Indicates transformer substation node Upper bound of the photovoltaic residual after aggregation; This represents the lower bound of the transformer net disturbance support set after considering the combined load and photovoltaic uncertainties. This represents the upper bound of the transformer net disturbance support set after considering the combined load and photovoltaic uncertainties. Indicates the diameter of the support set for net disturbance error; This represents the preset significance level parameter for the sub-bars; Indicates the dynamic Wasserstein radius; Represents the true probability distribution; Indicated by Wasserstein distance A fuzzy set constructed for the radius; Represents the Wasserstein distance operator; The support set represents the probability distribution.

[0062] Furthermore, the deterministic predictive control model for the virtual power plant in step 5 is as follows:

[0063] (A-37)

[0064] (A-38)

[0065] (A-39)

[0066] (A-40)

[0067] (A-41)

[0068] (A-42)

[0069] (A-43)

[0070] (A-44)

[0071] (A-45)

[0072] (A-46)

[0073] (A-47)

[0074] (A-48)

[0075] (A-49)

[0076] (A-50)

[0077] (A-51)

[0078] (A-52)

[0079] (A-53)

[0080] (A-54)

[0081] (A-55)

[0082] (A-56)

[0083] (A-57)

[0084] (A-58)

[0085] (A-59)

[0086] (A-60)

[0087] (A-61)

[0088] (A-62)

[0089] in, Indicates an index for electric vehicles; Indicates the first A collection of electric vehicles within a single station; Indicates the charging station index; Indicates the first A collection of charging stations within a single site; Indicates the first The site is The power that is constantly supplied to the external power grid; Indicates the first The site is The power that is constantly purchased from the external power grid; Indicates the first The site is Solar power output at all times; Indicates the first The site is Base load at any given time; Indicates the first The site is Net energy storage power at any given time; Indicates the first The site is The net power of all electric vehicles at any given moment; Indicates the first Energy storage in individual power stations The state of charge at any given moment; Indicates the first The rated capacity of energy storage in each power station; Indicates the charging efficiency of energy storage; Indicates the discharge efficiency of energy storage; Indicates a time period; Indicates the first The minimum state-of-charge capacity of energy storage in a power station; Indicates the first The upper limit of the energy storage state of charge of a power station; Indicates the first Energy storage in individual power stations The charging power at any given time; Indicates the first The maximum charging power of energy storage in each power station; Indicates the first Energy storage in individual power stations Discharge power at any given moment; Indicates the first The maximum discharge power of the energy storage of each power station; Represents the energy storage charging state variables; Represents the state variables of energy storage discharge; Indicates the first Each power station The first moment Is the electric vehicle connected to the first...? One charging station; Indicates the first The electric car reached the first The moment of each power station; Indicates the first The electric car left the first The moment of each power station; Indicates the first The first power station Are the electric vehicles at the station? Indicates the first The first power station The charging power allowed by the electric vehicle itself; Indicates the first The first power station The maximum charging power of each charging station; This represents the first physical extreme value combining the vehicle and the pile. The maximum charging power of an electric vehicle; Indicates the first The first power station The permissible discharge power of an electric vehicle itself; Indicates the first The first power station The maximum discharge power of each charging pile; This represents the first physical extreme value combining the vehicle and the pile. The maximum discharge power of an electric vehicle; Indicates the first The first power station A number of electric vehicles The charging power at any given time; Indicates the first The first power station A number of electric vehicles Discharge power at any given moment; Indicates the first The charging state variables of an electric vehicle; Indicates the first One electric vehicle discharge state variable; Indicates the first The first power station A number of electric vehicles The state of charge at any given moment; Indicates the first The first power station The charging efficiency of an electric vehicle; Indicates the first The first power station The discharge efficiency of an electric vehicle; Indicates the first The first power station The rated capacity of the battery of an electric vehicle; Indicates the first The first power station The target state of charge of the electric vehicle when it leaves the station; Indicates the first The electric vehicles at the departure time The state of charge; Indicates the first The first power station The state of charge limit of an electric vehicle; Indicates the first The first power station The upper limit of the state of charge of an electric vehicle; Indicates transformer substation area exist Net power at any given time; Indicates the first Each power station The frequency regulation reserve capacity to be declared at any time; Indicates the first Each power station The frequency regulation reserve capacity to be declared at all times; Indicates transformer substation area The transformer capacity; express One step forward in predicting energy market electricity prices; express Price of frequency regulation reserve capacity for each predicted step; Indicates the first The degradation cost of energy storage batteries in a single power station; This indicates the charging service fee; This indicates the V2G compensation unit price.

[0090] Furthermore, in step 6, the construction of CVaR risk constraints specifically involves transforming the transformer capacity over-limit risk into the following equivalent solvable convex optimization constraint set using duality theory:

[0091] (A-63)

[0092] (A-64)

[0093] (A-65)

[0094] (A-66)

[0095] (A-67)

[0096] (A-68)

[0097] (A-69)

[0098] (A-70)

[0099] (A-71)

[0100] (A-72)

[0101] (A-73)

[0102] (A-74)

[0103] (A-75)

[0104] (A-76)

[0105] (A-77)

[0106] In the formula, This represents the risk tolerance ratio coefficient; Indicates a security risk budget; This represents the transformer capacity over-limit loss function; Indicates transformer substation area In the prediction step Net power below; This indicates the set conditional risk value-safety confidence level; Describes the supremum operator; Indicate the infimum operator; Indicates transformer substation area exist Value at Risk (VaR) auxiliary optimization variables for each prediction step; Indicates transformer substation area exist The net perturbation error variable for each prediction step; Indicating targeting The first prediction step The local dual supremum function defined for each historical error scenario; The first norm of a vector; Indicates distribution The expected value of the following; Indicates correspondence The first prediction step Relaxed auxiliary variables for each error scenario; Indicates the dual multiplier variable; The above equivalent solvable convex optimization constraint set is introduced as a new safety risk constraint condition into the deterministic operation model described in step 5. Together with the original operation constraint equations (A-37) to (A-61) and objective function equation (A-62), it constitutes the sub-Blu-shaped bar model predictive control model with safety risk budget.

[0107] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0108] (1) Compared with the traditional virtual power plant optimization scheduling scheme, the present invention, based on the multi-node physical coupling prediction of spatiotemporal graph neural network, dynamically adjusts the Wasserstein radius by utilizing the measurement concentration characteristics of historical error samples, and guides the coordinated operation of heterogeneous photovoltaic, energy storage, and charging assets through the predictive control of the sub-Bruker model. This scheme helps to accurately quantify the safety risk boundary under uncertain conditions on both the source and load sides, and facilitates the scheduling of virtual power plants.

[0109] (2) The method proposed in this invention can effectively eliminate the risk of transformer capacity exceeding the limit, and greatly improve the system's operational safety and real-time response capability. The method proposed in this invention can also effectively characterize the uncertainty and spatiotemporal coupling characteristics of the virtual power plant's source and load sides, reduce the conservatism of fixed radius distributed bar optimization, and improve the safety and economy of multi-market collaborative scheduling of the virtual power plant. Attached Figure Description

[0110] Figure 1 This is a flowchart of the method of the present invention;

[0111] Figure 2 This is a graph showing the photovoltaic prediction results for a typical day as predicted by different deep learning models;

[0112] Figure 3 This is a graph showing the load forecast results for a typical day predicted by different deep learning models;

[0113] Figure 4 This is a diagram showing the decomposition results of the aggregated scheduling plan at the virtual power plant level;

[0114] Figure 5 It is the intraday variation curve of the dynamic Wasserstein radius of a typical transformer area. Detailed Implementation

[0115] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0116] like Figure 1 As shown, this invention proposes a blobs-based optimization scheduling method for virtual power plants using photovoltaic-storage-charging systems based on STGNN prediction. This method includes the following steps:

[0117] Step 1: Obtain the network topology parameters and historical operation scenario data of the virtual power plant; the network topology parameters include transformer capacity, photovoltaic-storage-charging station node connection relationship, distributed photovoltaic capacity, energy storage equipment parameters, and electric vehicle charging pile configuration; the historical operation scenario data includes the historical load demand, historical photovoltaic output, grid time-of-use electricity price, and frequency regulation market price data of each node.

[0118] Step 2: Construct a virtual power plant physical topology adjacency matrix based on the node connection relationship of the photovoltaic-storage-charging station. Using the historical load demand and historical photovoltaic output data of each node as input, extract the spatiotemporal coupling features using a spatiotemporal graph neural network STGNN containing three spatiotemporal blocks of time-space-time, and output the photovoltaic output prediction sequence and load demand prediction sequence of each photovoltaic-storage-charging station node in the future multiple time steps.

[0119] Step 3: Compare the predicted sequence with historical real operating data, extract the real prediction error of photovoltaic and load and convert it to the corresponding transformer substation node, and build a multi-scenario prediction error experience sample library for transformer substation node;

[0120] Step 4: Based on the sample matrix and preset safety confidence level parameters obtained from the sample library in Step 3 at each rolling scheduling time, construct the empirical distribution and net disturbance error support set of the transformer substation nodes in the current prediction time domain, and dynamically calculate the Wasserstein radius using the measure set theorem, thereby generating the time series Wasserstein fuzzy set corresponding to each transformer substation node.

[0121] Step 5: Based on the parameters and data obtained in Step 1, with energy storage charging and discharging power, electric vehicle V2G charging and discharging power and frequency regulation market application capacity as control variables, and taking into account the operation constraints of photovoltaic-storage charging stations and transformer capacity constraints, and with the objective function of maximizing the comprehensive profit of the virtual power plant in the energy market and frequency regulation market, establish a deterministic predictive control model for the virtual power plant.

[0122] Step 6: Introduce the prediction error empirical sample library from Step 3 and the dynamic Wasserstein fuzzy set from Step 4 into the deterministic model prediction and control model from Step 5, and construct CVaR risk constraints to establish a sub-Bruker model prediction and control model with a safety risk budget.

[0123] Step 7: Collect the initial state of charge of each energy storage device and electric vehicle in the virtual power plant in real time, and combine it with the real-time updated prediction sequence, price window and dynamic Wasserstein radius as input parameters for the sub-Bruker bar model predictive control model in Step 6 to obtain the optimal scheduling strategy set within the future time window; execute the charging, discharging and frequency regulation instructions of the first time step of the strategy set, update the state, and roll to the next scheduling time to solve, so as to realize the optimal scheduling of the virtual power plant.

[0124] The following example illustrates the superiority of the STGNN-based prediction-based virtual power plant split-rod optimization scheduling method for photovoltaic-storage-charging systems described in this invention.

[0125] A system of three transformer substations and eight photovoltaic-storage-charging stations is used as a case study. Two substations connect to three photovoltaic-storage-charging stations respectively, and the remaining substation connects to two photovoltaic-storage-charging stations. To compare the superiority of the proposed method, a stochastic optimization-based scheduling method, a fixed-radius bibliometric scheduling method, and the proposed STGNN-based bibliometric scheduling method for virtual power plants are compared for multi-market collaborative scheduling of virtual power plants. Simultaneously, backpropagation (BP) networks, long short-term memory (LSTM) networks, graph convolutional neural networks (GCN) networks, Transformer networks, and the proposed spatiotemporal graph neural network are used for prediction comparison. The algorithm of this invention is implemented using Python programming and employs the Gurobi solver to efficiently solve the transformed piecewise convex optimization problem.

[0126] Figure 2 and Figure 3 This paper presents the photovoltaic and load forecasts for a typical day using different deep learning models studied. Figure 2 and Figure 3 It can be seen that the spatiotemporal graph neural network model proposed in this invention exhibits better fitting results in both typical scenarios by alternately executing "temporal self-attention extraction - spatial graph interaction aggregation - deep temporal evolution fusion".

[0127] Figure 4 The decomposition results of the aggregated dispatch plan at the virtual power plant level are presented. Overall, the virtual power plant exhibits typical characteristics of "midday consumption, evening support, and rolling smoothing" throughout the day. During the peak daytime photovoltaic output period of 12:00-16:00, the system accurately captures the off-peak electricity price window, guiding the centralized charging of fixed energy storage and electric vehicles. This not only achieves full local consumption of new energy but also constructs a low-cost energy pool. During the morning and evening peak electricity consumption periods and high electricity price periods, the energy storage system and electric vehicles with V2G capabilities respond rapidly, switching to a discharge state to act as flexible peak power sources, effectively smoothing out the bi-peak fluctuations of the rigid base load.

[0128] Figure 5 The intraday variation curves of the dynamic Wasserstein radius for a typical transformer area are presented. Table 1 compares the scheduling results under two settings: a fixed radius (using the dynamic mean) and a dynamic radius. It can be seen that the Wasserstein radius exhibits a clear time-varying characteristic throughout the day: it is generally lower at night and in the early morning, indicating weaker system uncertainty; it rises rapidly after noon, reaching a high level between noon and afternoon, indicating that enhanced photovoltaic ramp-up and load fluctuations significantly expand the prediction error support set. Although it declines somewhat in the evening, it still maintains a certain level to cope with the operational risks during the evening peak period. This shows that the dynamic Wasserstein radius can accurately reflect the intraday risk evolution pattern.

[0129] Table 1. Scheduling Benefits under Different Radius Settings

[0130]

[0131] Table 2 compares the returns of stochastic optimization, robust optimization, and the method proposed in this invention. It can be seen that the proposed method achieves a better trade-off between return and risk, with a profit of 8863.92 yuan, significantly higher than robust optimization and only slightly lower than stochastic optimization. Furthermore, the total return of the method proposed in this invention is basically close to that of stochastic optimization and robust optimization, but the total cost is significantly lower than robust optimization. This indicates that it does not sacrifice return for security, but rather avoids overly conservative scheduling by more finely characterizing the uncertainty range. Therefore, the results show that the method proposed in this invention maintains strong risk defense capabilities while still possessing good economic efficiency.

[0132] Table 2. Scheduling benefits of different optimization methods

[0133]

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A STGNN prediction-based light storage charging virtual power plant distribution robust optimization scheduling method, characterized in that, The method includes the following steps: Step 1: Obtain the network topology parameters and historical operation scenario data of the virtual power plant; the network topology parameters include transformer capacity, photovoltaic-storage-charging station node connection relationship, distributed photovoltaic capacity, energy storage equipment parameters, and electric vehicle charging pile configuration; the historical operation scenario data includes the historical load demand, historical photovoltaic output, grid time-of-use electricity price, and frequency regulation market price data of each node. Step 2: Construct a virtual power plant physical topology adjacency matrix based on the node connection relationship of the photovoltaic-storage-charging station. Using the historical load demand and historical photovoltaic output data of each node as input, extract the spatiotemporal coupling features using a spatiotemporal graph neural network STGNN containing three spatiotemporal blocks of time-space-time, and output the photovoltaic output prediction sequence and load demand prediction sequence of each photovoltaic-storage-charging station node in the future multiple time steps. Step 3: Compare the predicted sequence with historical real operating data, extract the real prediction error of photovoltaic and load and convert it to the corresponding transformer substation node, and build a multi-scenario prediction error experience sample library for transformer substation node; Step 4: Based on the sample matrix and preset safety confidence level parameters obtained from the sample library in Step 3 at each rolling scheduling time, construct the empirical distribution and net disturbance error support set of the transformer substation nodes in the current prediction time domain, and dynamically calculate the Wasserstein radius using the measure set theorem, thereby generating the time series Wasserstein fuzzy set corresponding to each transformer substation node. Step 5: Based on the parameters and data obtained in Step 1, with energy storage charging and discharging power, electric vehicle V2G charging and discharging power and frequency regulation market application capacity as control variables, and taking into account the operation constraints of photovoltaic-storage charging stations and transformer capacity constraints, and with the objective function of maximizing the comprehensive profit of the virtual power plant in the energy market and frequency regulation market, establish a deterministic predictive control model for the virtual power plant. Step 6: Introduce the prediction error empirical sample library from Step 3 and the dynamic Wasserstein fuzzy set from Step 4 into the deterministic model prediction and control model from Step 5, and construct CVaR risk constraints to establish a sub-Bruker model prediction and control model with a safety risk budget. Step 7: Collect the initial state of charge of each energy storage device and electric vehicle in the virtual power plant in real time, and combine it with the real-time updated prediction sequence, price window and dynamic Wasserstein radius as input parameters for the sub-Bruker bar model predictive control model in Step 6 to obtain the optimal scheduling strategy set within the future time window; execute the charging, discharging and frequency regulation instructions of the first time step of the strategy set, update the state, and roll to the next scheduling time to solve, so as to realize the optimal scheduling of the virtual power plant.

2. The STGNN prediction-based distributed robust optimization scheduling method for optical storage and charging virtual power plants according to claim 1, characterized in that, In step 2, the spatiotemporal graph neural network model using three spatiotemporal blocks comprising time, space, and time is as follows: (1) Initial feature mapping and sine and cosine position coding (A-1) (A-2) (A-3) (A-4) In the formula, This represents the original input tensor composed of the historical load demand and photovoltaic output of each node; This represents the feature tensor mapped to the hidden layer dimension; This represents the linear projection weight matrix of the input layer; This represents the bias vector of the input layer; Represents the absolute position encoding matrix; Indicates the absolute position encoding matrix at the th The first time step Values ​​in each feature dimension; This indicates that the absolute position encoding matrix is ​​in the th... The first time step Values ​​in each feature dimension; This represents the dimension of the hidden layer features within the model network; Indicates the index of the time step in the sequence; Indicates the feature dimension index; This represents the initial spatiotemporal feature tensor of each node, which incorporates location encoding information. (2) First-layer temporal feature extraction (A-5) (A-6) (A-7) (A-8) (A-9) (A-10) (A-11) (A-12) In the formula, , and They represent the first The query vector, key vector, and value vector of each attention head; , , They represent the first The learnable weight matrix used to generate queries, keys, and values ​​in each attention head; Indicates the total number of nodes; Indicates that for a given first... The corresponding time series extracted from each node. ; Indicates the attention head index; Key vector transpose, Indicates the first Local temporal feature representation of the attention head output; This represents the exponential normalization function; Represents the feature dimensions of each attention head; This indicates a high level of attention from multiple parties. This indicates matrix concatenation along the feature dimension; This represents the total number of heads receiving multi-head attention. This represents the output weights of the multi-head attention concatenation; This represents the intermediate feature tensor after multi-head self-attention and the first residual connection; Indicates the layer normalization function; This represents a feedforward neural network; Represents a non-linear activation function; and This is the weight vector; and For bias terms; This indicates the node after the first layer of time features has been extracted. The characteristic output; (3) Spatial feature aggregation (A-13) (A-14) (A-15) In the formula, The original adjacency matrix representing the physical connection relationships of virtual power plant nodes; It is the identity matrix; This represents the adjacency matrix after introducing self-loops; Represents the symmetric normalized adjacency matrix; It is a diagonal degree matrix; This represents the shared weight matrix for feature propagation between nodes in a graph convolutional layer; This represents the bias vector of the graph convolutional layer; Indicates the regularization method; Indicates time step The feature input of all nodes after the first layer of temporal feature extraction at any given time; Indicates time step The node characteristics after spatial topological aggregation at any given moment; (4) Second-layer temporal feature extraction and global prediction (A-16) (A-17) In the formula, This indicates the output obtained by extracting temporal features again using the same multi-head self-attention and feedforward network structure as the first layer; This represents the fusion feature of the final output of the entire spatiotemporal block; Indicates the length of the input history sequence; This represents the comprehensive spatiotemporal characteristic state of all nodes at the last moment of the current historical sequence; and This represents the weight matrix of the output mapping layer; and This represents the bias vector of the output mapping layer; This represents the predicted values ​​of the output photovoltaic power and the load.

3. The method for optimal scheduling of a virtual power plant based on STGNN prediction for photovoltaic-storage-charging systems according to claim 2, characterized in that, In step 3, the method for constructing a multi-scenario prediction error experience sample library for transformer substation nodes is as follows: (A-18) (A-19) (A-20) (A-21) In the formula, Indicates an index of historical experience error scenarios; Indicates the number of valid historical error samples; Indicates the index of photovoltaic and energy storage charging stations; Indicates the first In the scenario, the first The actual load of each site; Indicates the first In the scenario, the first Forecast load for each site; Indicates the first In the scenario, the first The actual photovoltaic output of each site; Indicates the first In the scenario, the first Predicted photovoltaic output for each site; Indicates the first In the scenario, the first Load forecast residuals for each site; Indicates the first In the scenario, the first Photovoltaic prediction residuals at each site; Indicates the transformer substation index; This indicates that it belongs to the transformer substation area. A collection of power stations; Indicates transformer substation area In the Net prediction error samples in each scenario; Indicates transformer substation area A multi-scenario prediction error empirical sample library.

4. The method for optimal scheduling of a virtual power plant based on STGNN prediction for photovoltaic-storage-charging systems according to claim 3, characterized in that, The method for step 4 is as follows: (A-22) (A-23) (A-24) (A-25) (A-26) (A-27) (A-28) (A-29) (A-30) (A-31) (A-32) (A-33) (A-34) (A-35) (A-36) In the formula, Indicates the current rolling scheduling time; This represents the total number of time steps in the look-ahead prediction time domain; This represents the time step index within the look-ahead prediction time domain; Indicates the current rolling scheduling time Starting from the first The scheduling period corresponding to each prediction step; To extract from the historical error sample database Extracted corresponding transformer substations At the current rolling scheduling moment Corresponding future Subsample matrix of each prediction step ; Indicates transformer substation area In the scene Next The net perturbation error sample corresponding to each prediction step; Indicates transformer substation area In the Empirical distribution under each prediction step; Indicates the location of the net disturbance error sample Dirac measure at the location; Indicates the first The first prediction step Load forecast values ​​for each site; Indicates the first The first prediction step The historical load limit for each site; Indicates the first The first prediction step Lower bound of load residual for each site; Indicates the first The first prediction step Upper bound of load residuals for each site; Indicates transformer substation node Lower bound of the aggregated load residual; Indicates transformer substation node Upper bound of the aggregated load residual; Indicates the first The first prediction step Photovoltaic forecast values ​​for each site; Indicates the first The first prediction step The upper limit of historical photovoltaic experience for each site; Indicates the first The first prediction step Lower bound of photovoltaic residuals at each site; Indicates the first The first prediction step Upper bound of photovoltaic residuals at each site; Indicates transformer substation node Lower bound of the photovoltaic residual after aggregation; Indicates transformer substation node Upper bound of the photovoltaic residual after aggregation; This represents the lower bound of the transformer net disturbance support set after considering the combined load and photovoltaic uncertainties. This represents the upper bound of the transformer net disturbance support set after considering the combined load and photovoltaic uncertainties. Indicates the diameter of the support set for net disturbance error; This represents the preset significance level parameter for the sub-bars; Indicates the dynamic Wasserstein radius; Represents the true probability distribution; Indicates the distance from Wasserstein. A fuzzy set constructed for the radius; Represents the Wasserstein distance operator; The support set represents the probability distribution.

5. The method for optimal scheduling of a virtual power plant based on STGNN prediction for photovoltaic-storage-charging systems according to claim 4, characterized in that, The deterministic predictive control model for the virtual power plant in step 5 is as follows: (A-37) (A-38) (A-39) (A-40) (A-41) (A-42) (A-43) (A-44) (A-45) (A-46) (A-47) (A-48) (A-49) (A-50) (A-51) (A-52) (A-53) (A-54) (A-55) (A-56) (A-57) (A-58) (A-59) (A-60) (A-61) (A-62) in, Indicates an index for electric vehicles; Indicates the first A collection of electric vehicles within a single station; Indicates the charging station index; Indicates the first A collection of charging stations within a single site; Indicates the first The site is The power that is constantly supplied to the external power grid; Indicates the first The site is The power that is constantly purchased from the external power grid; Indicates the first The site is Solar power output at all times; Indicates the first The site is Base load at any given time; Indicates the first The site is Net energy storage power at any given time; Indicates the first The site is The net power of all electric vehicles at any given moment; Indicates the first Energy storage in individual power stations The state of charge at any given moment; Indicates the first The rated capacity of energy storage in each power station; Indicates the charging efficiency of energy storage; Indicates the discharge efficiency of energy storage; Indicates a time period; Indicates the first The minimum state-of-charge capacity of energy storage in a power station; Indicates the first The upper limit of the energy storage state of charge of a power station; Indicates the first Energy storage in individual power stations The charging power at any given time; Indicates the first The maximum charging power of energy storage in each power station; Indicates the first Energy storage in individual power stations Discharge power at any given moment; Indicates the first The maximum discharge power of the energy storage of each power station; Represents the energy storage charging state variables; Represents the state variables of energy storage discharge; Indicates the first Each power station The first moment Is the electric vehicle connected to the first...? One charging station; Indicates the first The electric car reached the first The moment of each power station; Indicates the first The electric car left the first The moment of each power station; Indicates the first The first power station Are the electric vehicles at the station? Indicates the first The first power station The charging power allowed by the electric vehicle itself; Indicates the first The first power station The maximum charging power of each charging station; This represents the first physical extreme value combining the vehicle and the pile. The maximum charging power of an electric vehicle; Indicates the first The first power station The permissible discharge power of an electric vehicle itself; Indicates the first The first power station The maximum discharge power of each charging pile; This represents the first physical extreme value combining the vehicle and the pile. The maximum discharge power of an electric vehicle; Indicates the first The first power station A number of electric vehicles The charging power at any given time; Indicates the first The first power station A number of electric vehicles Discharge power at any given moment; Indicates the first The charging state variables of an electric vehicle; Indicates the first One electric vehicle discharge state variable; Indicates the first The first power station A number of electric vehicles The state of charge at any given moment; Indicates the first The first power station The charging efficiency of an electric vehicle; Indicates the first The first power station The discharge efficiency of an electric vehicle; Indicates the first The first power station The rated capacity of the battery of an electric vehicle; Indicates the first The first power station The target state of charge of the electric vehicle when it leaves the station; Indicates the first The electric vehicles at the departure time The state of charge; Indicates the first The first power station The state of charge limit of an electric vehicle; Indicates the first The first power station The upper limit of the state of charge of an electric vehicle; Indicates transformer substation area exist Net power at any given time; Indicates the first Each power station The frequency regulation reserve capacity to be declared at any time; Indicates the first Each power station The frequency regulation reserve capacity to be declared at all times; Indicates transformer substation area The transformer capacity; express One step forward in predicting energy market electricity prices; express Price of frequency regulation reserve capacity for each predicted step; Indicates the first The degradation cost of energy storage batteries in a single power station; This indicates the charging service fee; This indicates the V2G compensation unit price.

6. The method for optimal scheduling of a virtual power plant based on STGNN prediction for photovoltaic-storage-charging systems according to claim 5, characterized in that, In step 6, the construction of CVaR risk constraints specifically involves transforming the transformer capacity over-limit risk into the following equivalent solvable convex optimization constraint set using duality theory: (A-63) (A-64) (A-65) (A-66) (A-67) (A-68) (A-69) (A-70) (A-71) (A-72) (A-73) (A-74) (A-75) (A-76) (A-77) In the formula, This represents the risk tolerance ratio coefficient; Indicates a security risk budget; This represents the transformer capacity over-limit loss function; Indicates transformer substation area In the prediction step Net power below; This indicates the set conditional risk value-safety confidence level; Describes the supremum operator; Indicate the infimum operator; Indicates transformer substation area exist Value at Risk (VaR) auxiliary optimization variables for each prediction step; Indicates transformer substation area exist The net perturbation error variable for each prediction step; Indicating targeting The first prediction step The local dual supremum function defined for each historical error scenario; The first norm of a vector; Indicates distribution The expected value of the following; Indicates correspondence The first prediction step Relaxed auxiliary variables for each error scenario; Indicates the dual multiplier variable; The above equivalent solvable convex optimization constraint set is introduced as a new safety risk constraint condition into the deterministic operation model described in step 5. Together with the original operation constraint equations (A-37) to (A-61) and objective function equation (A-62), it constitutes the sub-Blu-shaped bar model predictive control model with safety risk budget.

Images