A virtual power plant interaction model construction method and related device

Abstract

A virtual power plant interaction model construction method and related device, comprising: establishing an interactive physical model of a virtual power plant and a power distribution network, and the virtual power plant determining whether to respond to a new daily dispatch instruction issued by the power distribution network according to a remaining adjustable capacity; obtaining a training set by uniformly generating a dispatch instruction and obtaining a test set by randomly generating a dispatch instruction; normalizing the data set, establishing a dispatch instruction feasibility model for the normalized data set; and constructing a virtual power plant interaction cost model based on a kernel extreme learning machine-Gaussian process regression (KELM-GPR) based on the dispatch instruction feasibility model. The classification boundary can be used to learn the adjustable range of the virtual power plant, which is conducive to making reasonable scheduling plans for the virtual power plant itself and the superior power distribution network dispatch center. Finally, the virtual power plant interaction model is a data-driven model, and it does not need to be optimized and solved internally for the new daily dispatch instruction issued by the power distribution network, and it can respond within milliseconds, so it can meet the timeliness of the daily dispatch of the power distribution network.

Description

Technical Field

[0001] This invention belongs to the field of virtual power plant optimization scheduling, and specifically relates to a method for constructing a virtual power plant interaction model and related apparatus. Background Technology

[0002] While the installed capacity of new energy sources continues to increase, the intermittent and fluctuating nature of their output poses significant challenges to the operation and management of distribution networks. Virtual power plants, as one approach to address this issue, effectively aggregate distributed power sources, controllable loads, and energy storage into a virtual whole, providing an effective way for them to actively participate in distribution network dispatch and the electricity market. Currently, most virtual power plant participation in distribution network dispatch is based on the physical models of the aggregated units within the virtual power plant, constructing an optimized dispatch model with the goal of minimizing operating costs or maximizing economic benefits. However, this physical model-based method forces virtual power plants to provide detailed physical model parameters to the distribution network dispatch center, which can easily lead to the leakage of important information about the virtual power plant itself. Furthermore, for the distribution network dispatch center, considering that there may be multiple virtual power plants within its jurisdiction, establishing an optimized dispatch model based on physical models presents challenges such as model complexity, numerous variables, and difficulty in solving the problem. To address these issues, some scholars have proposed a two-layer optimization method. While the two-layer optimization scheduling method addresses the privacy leaks and numerous variables inherent in virtual power plants (VPPs) to some extent, the diversity of VPP components, temporal coupling, and numerous constraints make it difficult to analytically characterize the dispatchable boundaries of VPPs, hindering the provision of accurate dispatchable ranges to the distribution network dispatch center. Furthermore, the iterative solution relies on information exchange between the upper and lower layers, inevitably leading to lengthy solution times, particularly for intraday real-time optimization scheduling models, where timeliness is often unacceptable. Summary of the Invention

[0003] The purpose of this invention is to provide a method and related apparatus for constructing a virtual power plant interaction model, so as to solve the problems of leakage of important information, difficulty in parsing and characterizing dispatchable boundaries, and difficulty in meeting the timeliness of intraday dispatching of the distribution network when virtual power plants participate in distribution network dispatching.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for constructing a virtual power plant interaction model, comprising:

[0006] Establish an interactive physical model between the virtual power plant and the distribution network. The virtual power plant determines whether to respond to new intraday dispatch instructions issued by the distribution network based on its remaining adjustable capacity.

[0007] The training set is obtained by uniformly generating scheduling instructions, and the test set is obtained by randomly generating scheduling instructions.

[0008] The dataset is normalized, and a feasibility model for scheduling instructions is established for the normalized dataset.

[0009] Based on the scheduling instruction feasibility model, a virtual power plant interaction cost model based on KELM-GPR (Kinstantaneous Extreme Learning Machine-Gaussian Process Regression) is constructed.

[0010] Optionally, an interactive physical model of the virtual power plant and the distribution network can be established:

[0011] The virtual power plant, which utilizes surplus electricity resources, adopts a strategy of feeding surplus electricity into the grid. Its physical interaction model with the distribution network is as follows:

[0012] Objective function:

[0013]

[0014]

[0015] In the formula, F represents the operating cost of the virtual power plant for the remaining time slots within the day; k represents the current time slot (k = 0 if the interactive model was run the previous day); T represents the number of time slots within a day; F WT,t F PV,t F ES,t F MT,t F CL,t F Q,t F S,t These represent the following costs for time period t: wind power operation and management cost, photovoltaic operation and management cost, energy storage battery operation and management cost, gas turbine operation cost, demand response operation cost, off-load penalty cost, and surplus electricity grid connection cost; C WT C PV C ES a MT b MT C CL C q C sell The following are the cost coefficients for wind power operation and management, photovoltaic operation and management, energy storage battery operation and management, gas turbine operation, demand response compensation, and load shedding penalty, all calculated at 5 times the time-of-use tariff, plus the surplus electricity price for period t. P WT,t P PV,t P ES,t P MT,t P CL,t P q,t P sell,t These represent the wind power output, photovoltaic power output, energy storage battery output, gas turbine output, load-reducing output, off-load power, and surplus power fed into the grid during time period t, respectively; Δt is the duration of a time period.

[0016] Constraints:

[0017] Virtual power plant power balance constraints

[0018] P ES,t +P MT,t +P CL,t +P q,t -P sell,t =Δp G,t +P G,t -(P WT,t +P PV,t (3)

[0019] In the formula, Δp G,t P refers to new dispatch instructions issued by the distribution network within a certain period of time in the future. G,t The load is the one that has been determined by the distribution network dispatch command.

[0020] Energy storage battery operating constraints

[0021]

[0022] In the formula, P t dis P t char These represent the discharge power and charging power of the energy storage battery during time period t, respectively. This represents the maximum charge and discharge power of the energy storage battery. η represents the charge / discharge state of the energy storage battery, a 0-1 variable; char η dis These represent the charge and discharge efficiencies of the energy storage battery; E bat S represents the rated capacity of the energy storage battery. t S represents the state of charge of the energy storage battery during time period t; min S max These represent the minimum and maximum permissible states of charge (S) of the energy storage battery, respectively; start S end These represent the initial and final states of charge of the energy storage battery during a day;

[0023] Gas turbine operating constraints

[0024]

[0025] In the formula, u GT,t It is a 0-1 variable, representing the start-up and shutdown status of the gas turbine, which is determined by the day-ahead dispatch of the virtual power plant; These are the minimum and maximum power outputs of the gas turbine, respectively. These represent the downward and upward ramp rates of the gas turbine, respectively.

[0026] Reduced load operating constraints

[0027]

[0028] In the formula, The maximum output power that can be reduced during time period t.

[0029] Electricity sales constraints

[0030]

[0031] In the formula, The upper limit of the power output of a virtual power plant to the grid.

[0032] Based on intraday dispatch needs, the distribution network issues new dispatch instructions to the virtual power plant, including the dispatch start time ts, the dispatch duration tc, and the dispatch power Δp. G The virtual power plant assesses and responds to the feasibility of the instruction based on its own adjustability. m and the corresponding cost f c Firstly, the virtual power plant will not receive any new dispatch instructions, i.e., Δp. G When = 0, the interactive physical model is run to obtain the pre-scheduling plan and operating cost of each unit of the virtual power plant, denoted as . F 0 The regulations stipulate that when a virtual power plant receives a new dispatch instruction from the distribution network, it cannot increase the load shedding and surplus power grid connection under the existing load shedding power and pre-sale power plan. That is, constraint (7) is modified to constraint (8).

[0033]

[0034] Under this constraint, when the distribution network issues a new dispatch command, the virtual power plant operates an interactive physical model. If the model has a solution, it can respond to the new command while satisfying existing dispatch commands, i.e., f m =1, the cost of this scheduling instruction is f c =FF 0 If the model has no solution, then f m =0,f c =0.

[0035] Optionally, the virtual power plant interaction model adopts a strategy of training offline before the day and correcting online during the day:

[0036] The virtual power plant simulates new dispatch instructions that the distribution network may issue within a certain period of time based on its own operating status. A limited data sample set is obtained by solving an interactive physical model. Finally, data-driven techniques are used to train the data sample set to characterize the interaction characteristics between the virtual power plant and the distribution network within a certain period of time. A strategy of day-ahead offline training and intraday online correction is adopted. Specifically, during the day-ahead offline training phase, the virtual power plant solves the feasibility and interaction cost of dispatch instructions through the interactive physical model, generating intraday simulation data as the dataset. The dataset is then trained using WKNN and KELM-GPR from machine learning to determine the model structure and hyperparameters, including the number of neighbors in WKNN, the regularization coefficient in KELM, and the kernel function. The dataset size is as follows: During the intraday online correction phase, the virtual power plant decides whether to update the data-driven model parameters based on the dispatching needs of the distribution network dispatch center. The distribution network dispatch center informs the virtual power plant one time period in advance whether there is a dispatching need. For example, if the data-driven model needs to be called during time period k1, a dispatching invitation is sent to the virtual power plant during time period k1-1. During time period k1, the virtual power plant, based on the data-driven model structure and hyperparameters determined by the previous day's training and combined with the ultra-short-term wind and solar power output forecast data, generates incremental samples as a new dataset through the interaction physical model between the virtual power plant and the distribution network. This dataset is then used to train and update the model parameters, improving the model's prediction accuracy for use by the distribution network dispatch center. If there is no dispatching need, the virtual power plant does not update the model parameters.

[0037] Optionally, a training set can be obtained by uniformly generating scheduling instructions, and a test set can be obtained by randomly generating scheduling instructions.

[0038] Based on the adjustability of the virtual power plant, the feasible region for issuing new dispatch instructions to the distribution network is analyzed, that is, the reference boundary for generating new dispatch instructions needs to be known, in order to obtain the dataset, specifically:

[0039] 1) Reference boundary for generating new scheduling instructions

[0040] The new intraday dispatch instructions issued by the distribution network are [ts, tc, Δp] G Reference range for ts and tc values:

[0041]

[0042] In the formula, ts min The minimum allowed value for the start time period of scheduling; ts min , ts, and tc have the same unit and are taken as integers; when tc = 1, There are 16 possible cases when tc = 2. There are 15 possible combinations of ts and tc, resulting in 136 possible values. Ignoring the SOC of the energy storage battery and the ramp rate limitations of the gas turbine, the maximum possible dispatch power range of the virtual power plant, while satisfying existing dispatch instructions, is:

[0043]

[0044] In the formula, Pre-dispatch plans for each unit of the virtual power plant in the absence of new dispatch instructions; and Let these be the upper and lower boundaries of the virtual power plant's maximum possible output. Then, the actual output boundary of the virtual power plant must be included within these boundaries. Since the new scheduling instructions within the day are based on Δp during the scheduling start and end time period. G Since the power output is continuous, the dispatch power reference boundary can be taken as the minimum value of the maximum possible dispatch power boundary of the virtual power plant during the dispatch start and end time period, in order to further reduce the dispatch power reference range; then the upper and lower boundaries of the dispatch power corresponding to the i-th case are:

[0045]

[0046] In the formula, ts i , tc i , Let $i$ be the scheduling start time period, the scheduling duration period, and the upper and lower bounds of the scheduling power reference for the $i$ case, respectively. Therefore, the size of the scheduling power range for each case is:

[0047]

[0048] 2) Simulate new dispatch instructions issued by the distribution network

[0049] After obtaining the reference boundary of the dispatching instructions, new dispatching instructions that may be issued by the distribution network are generated, as follows:

[0050] (1) Uniform generation of scheduling instructions

[0051] Based on the set number of training sets N train The sample size and scheduling power for each case are as follows:

[0052]

[0053] In the formula, b i Let PW be the size of the power range corresponding to the i-th case. i The proportion of the sum of the power value ranges in all cases; floor(*) is the floor function; n i n is the number of samples taken in the i-th case; errorThe difference in sample size; mod(*) is the modulo function; sort(*) is the descending sort function; value and index are the descending sort value and the corresponding index number, respectively; index * The first n in the index err or index number; Δp G,i,j Let ΔP be the scheduling power value corresponding to case i, that is, the scheduling power value corresponding to case i is uniformly selected at equal intervals within the reference range; G,i,j ts corresponding to i cases i , tc i By combining them, a uniformly generated set of scheduling instructions can be obtained, as shown in equation (14).

[0054]

[0055] In the formula, These represent the sets of the start time period, duration period, and power of the scheduling instructions in the training set, respectively. These represent the start time period, scheduling duration, and scheduling power of the l-th scheduling instruction in the training set, respectively.

[0056] (2) Random generation of scheduling instructions

[0057] The scheduling instruction reference boundary and the number N of the test set, as known from the preceding text, are... test Through random functions in ts, tc, Δp G The scheduling instructions are generated by randomly sampling within the reference value range and must satisfy the constraint of equation (9); then a single randomly generated scheduling instruction is

[0058]

[0059] In the formula, U[ts min ,ts min +15] indicates that in [ts min ,ts min Random sampling within the interval [+15] Let the start time, duration, and power of the l-th scheduling instruction in the test set be the starting time, duration, and power, respectively. Then, the randomly generated scheduling instruction set is shown in equation (16), where, These represent the sets of the start time period, duration period, and power of the scheduling instructions in the test set, respectively.

[0060]

[0061] 3) The virtual power plant responds to dispatch instructions and generates a dataset.

[0062] After obtaining the dispatch instructions from equations (14) and (16), the feasibility and interaction cost of the dispatch instructions can be obtained through the interactive physical model based on the method for determining whether the virtual power plant responds to the distribution network dispatch instructions, as shown in equations (17)-(20); where, These are the sets of scheduling instruction feasibility and interaction costs, respectively, for the training and testing sets. These represent the scheduling instruction feasibility and interaction cost corresponding to the l-th scheduling instruction in the training set, respectively. These represent the feasibility of the scheduling instruction and the interaction cost corresponding to the l-th scheduling instruction in the test set, respectively.

[0063]

[0064]

[0065]

[0066]

[0067] In summary, the dataset required for model training is:

[0068]

[0069]

[0070] In the formula, x m x c These are the datasets required for the scheduling instruction feasibility model and the interaction cost model, respectively.

[0071] Optionally, normalize the dataset:

[0072] For the feasibility model of scheduling instructions, the dataset is normalized to a range of [-1, 1]. The calculation formula is as follows:

[0073]

[0074] In the formula, x represents the unnormalized original data. * The dataset is normalized; min(x) is the minimum value in the original dataset; max(x) is the maximum value in the original dataset. For the interaction cost model: first, abnormal data where scheduling instructions are not executable are removed from the interaction cost dataset; second, the scheduling instructions in the dataset are transformed from 3-dimensional features to 16-dimensional features to allow the interaction cost model to fully explore the complex nonlinear relationship between scheduling instructions and interaction costs. The transformation formula is as follows:

[0075]

[0076] In the formula, zeros(ts) l -ts min ,1) indicates generation (ts l -ts min A zero matrix of 1×1; ones(tc l ,1) indicates the generation of tc l A matrix of 1×1, all equal to 1; Al represents the l-th transformed scheduling instruction. Finally, the transformed dataset is normalized according to equation (23).

[0077] Optionally, a feasibility model for scheduling instructions can be established for the normalized dataset:

[0078] A scheduling command feasibility model is established using the weighted k-nearest neighbor (WKNN) algorithm. The WKNN model is then applied to construct the scheduling command feasibility model. For the key elements of WKNN, Euclidean distance is used as the distance metric, and the inverse squared distance is used as the distance weight. The model prediction accuracy is used as the objective function, and the Particle Swarm Optimization (PSO) algorithm is employed for optimization. The AUC index in the confusion matrix, which is insensitive to sample imbalance, is selected as the evaluation criterion for model prediction accuracy. The training process of the scheduling command feasibility model is as follows:

[0079] During the initial training, after data preprocessing, the number of neighboring points in the model was first optimized using the particle swarm optimization algorithm. Then, the classification model was trained and tested. If the prediction accuracy of the classification model did not meet the requirements, the above steps were repeated by increasing the dataset size until the required prediction accuracy was achieved. This process determined the number of neighboring points and the required dataset size. In real-time operation, based on the model structure and hyperparameters obtained from the initial training, combined with the real-time operating status of the virtual power plant and ultra-short-term wind and solar forecasts, incremental samples were generated as a new dataset through an interactive physical model to train and update the classification model parameters. The trained scheduling instruction feasibility model is denoted as...

[0080]

[0081] The obtained scheduling instruction reference boundary is also used as a model to determine the feasibility of scheduling instructions; that is, instructions outside the reference boundary are infeasible, and instructions inside the reference boundary are feasible, denoted as .

[0082]

[0083] Will and By complementing each other's strengths, a feasibility model for the final scheduling instructions was obtained:

[0084]

[0085] According to this model, the virtual power plant can know model f MThe classification boundaries are used to determine the dispatchable range of the virtual power plant and provide this information to the distribution network.

[0086] Optionally, a virtual power plant interaction cost model based on KELM-GPR (Kernel Extreme Learning Machine-Gaussian Process Regression) is constructed, specifically as follows:

[0087] A KELM-based interaction cost model was established, and the GPR (Growth Percentage Reduction) model was introduced as an error compensation model to improve the prediction accuracy of the interaction cost model. Typical radial basis functions and exponential covariance functions were used as the kernel functions of KELM and GPR, respectively. Mean absolute percentage error (MAPE) and coefficient of determination (R²) were used as evaluation metrics for the model's prediction performance. The training process of the interaction cost model is as follows:

[0088] During the initial training, the original dataset was first preprocessed. Then, the regularization coefficient C of KELM and the kernel parameter σ of the kernel function were optimized using the particle swarm optimization algorithm. Finally, the interaction cost model was trained and tested. If the model's prediction performance did not meet the requirements, the dataset size was increased and the above steps were repeated until the model's prediction performance met the requirements. During intraday real-time operation, based on the model structure and hyperparameters obtained from the initial training, combined with the real-time operating status of the virtual power plant and ultra-short-term wind and solar forecasts, incremental samples were generated as a new dataset through the interaction physics model to train and update the interaction cost model parameters. The trained interaction cost model is denoted as […].

[0089]

[0090] Provided that the training indicators are met, the dataset size required for daily updates of the virtual power plant interaction model is the maximum of the dataset sizes required for updates of the scheduling instruction feasibility model and the interaction cost model.

[0091] Secondly, the present invention provides a virtual power plant interactive model construction system, comprising:

[0092] The interactive physical model construction module is used to establish an interactive physical model between the virtual power plant and the distribution network. The virtual power plant determines whether to respond to the new intraday dispatch instructions issued by the distribution network based on its remaining adjustable capacity.

[0093] The data acquisition module is used to obtain the training set by uniformly generating scheduling instructions and to obtain the test set by randomly generating scheduling instructions;

[0094] The data processing module is used to normalize the dataset and establish a feasibility model for scheduling instructions on the normalized dataset.

[0095] The interaction cost model construction module is used to construct a virtual power plant interaction cost model based on Kernel Extreme Learning Machine-Gaussian Process Regression (KELM-GPR) based on the scheduling instruction feasibility model.

[0096] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a method for constructing a virtual power plant interaction model.

[0097] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for constructing a virtual power plant interaction model.

[0098] Compared with the prior art, the present invention has the following technical effects:

[0099] First, this invention employs data-driven technology in its virtual power plant interaction model, making it a function with defined parameters. External personnel cannot know the specific information of each internal operating unit, thus protecting the internal information security of the virtual power plant. Second, the scheduling command feasibility model in the virtual power plant interaction model is essentially a binary classification model. Based on its classification boundary, the dispatchable range of the virtual power plant can be determined, which is beneficial for both the virtual power plant itself and the upper-level distribution network dispatch center to make reasonable scheduling plans. Finally, because the virtual power plant interaction model is a data-driven model, it does not require internal optimization and solution for new intraday dispatch commands issued by the distribution network and can respond in milliseconds, thus meeting the timeliness requirements of intraday dispatching of the distribution network. Attached Figure Description

[0100] Figure 1 This is a diagram of the virtual power plant resource composition of the present invention.

[0101] Figure 2 This is a diagram illustrating the interaction process between the virtual power plant and the distribution network in this invention.

[0102] Figure 3 This is a training strategy diagram for the virtual power plant interaction model of the present invention.

[0103] Figure 4 This is a training diagram of the feasibility model for the scheduling instructions of this invention.

[0104] Figure 5 This is the training graph of the interaction cost model of the present invention.

[0105] Figure 6 This is a wind and solar power output forecast and dispatch command load curve for a typical day before the VPP of the present invention.

[0106] Figure 7 This is the VPP day-ahead scheduling plan diagram of the present invention. Detailed Implementation

[0107] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0108] Step 1: Establish an interactive physical model between the virtual power plant and the distribution network. The virtual power plant determines whether to respond to new intraday dispatch instructions issued by the distribution network based on its remaining adjustable capacity.

[0109] Step 2: Considering the time-varying nature of the virtual power plant's external interaction characteristics, the virtual power plant interaction model adopts a strategy of offline training before the day and online correction within the day.

[0110] Step 3: Considering the boundaries of scheduling instructions, obtain the training set by uniformly generating scheduling instructions and obtain the test set by randomly generating scheduling instructions.

[0111] Step 4: Normalize the dataset.

[0112] Step 5: For the normalized dataset, establish a feasibility model for scheduling instructions using the weighted K-nearest neighbor (WKNN) algorithm.

[0113] Step 6: Construct a virtual power plant interaction cost model based on Kernel Extreme Learning Machine-Gaussian Process Regression (KELM-GPR) based on the scheduling instruction feasibility model.

[0114] Specifically:

[0115] In step 1, taking a virtual power plant (VPP) consisting of wind turbine generators, photovoltaic power plants, gas turbines, energy storage batteries, and user-side load shedding components as an example, such as... Figure 1 As shown. The VPP adopts a centralized control mode, aiming at optimal operating costs, and completes the distribution network dispatch instructions by coordinating internal aggregated resources. Among them, wind power and photovoltaic are non-adjustable units, and their full output is consumed. When the VPP's own output cannot meet the distribution network dispatch instructions, it will face a load shedding penalty. The VPP's surplus power resources adopt a surplus power grid connection strategy. Its interaction physical model with the distribution network is as follows.

[0116] Objective function:

[0117]

[0118]

[0119] In the formula, F is the operating cost of the VPP for the remaining time slots within the day, k is the current time slot (k = 0 if the interaction model was run the previous day), T is the number of time slots within a day; F WT,t F PV,t F ES,t F MT,t F CL,t F Q,t F S,tThese represent the following costs for time period t: wind power operation and management cost, photovoltaic operation and management cost, energy storage battery operation and management cost, gas turbine operation cost, demand response operation cost, off-load penalty cost, and surplus electricity grid connection cost; C WT C PV C ES a MT b MT C CL C q C sell The following are the cost coefficients for wind power operation and management, photovoltaic operation and management, energy storage battery operation and management, gas turbine operation, demand response compensation, and load shedding penalty, all calculated at 5 times the time-of-use tariff, plus the surplus electricity price for period t. P WT,t P PV,t P ES,t P MT,t P CL,t P q,t P sell,t These represent the wind power output, photovoltaic power output, energy storage battery output, gas turbine output, load reduction output, off-load power, and surplus power fed into the grid during time period t, respectively; Δt is the duration of a time period.

[0120] Constraints:

[0121] 1) VPP power balance constraint

[0122] P ES,t +P MT,t +P CL,t +P q,t -P sell,t =Δp G,t +P G,t -(P WT,t +P PV,t (3)

[0123] In the formula, Δp G,t P refers to new dispatch instructions issued by the distribution network within a certain period of time in the future. G,t The load is the one that has been determined by the distribution network dispatch command.

[0124] Energy storage battery operating constraints

[0125]

[0126] In the formula, P t dis P t char These represent the discharge power and charging power of the energy storage battery during time period t, respectively. This represents the maximum charge and discharge power of the energy storage battery. η represents the charge / discharge state of the energy storage battery, a 0-1 variable; char η dis These represent the charge and discharge efficiencies of the energy storage battery; E bat S represents the rated capacity of the energy storage battery. t S represents the state of charge of the energy storage battery during time period t; min S max These represent the minimum and maximum permissible states of charge (S) of the energy storage battery, respectively; start S end These represent the initial and final states of charge of the energy storage battery during a day, respectively.

[0127] Gas turbine operating constraints

[0128]

[0129] In the formula, u GT,t It is a 0-1 variable, representing the start-up and shutdown status of the gas turbine, which is determined by the VPP day-ahead dispatch; These are the minimum and maximum power outputs of the gas turbine, respectively. These represent the downward and upward ramp rates of the gas turbine, respectively.

[0130] Reduced load operating constraints

[0131]

[0132] In the formula, The maximum output power that can be reduced during time period t.

[0133] Electricity sales constraints

[0134]

[0135] In the formula, The upper limit of the power output that a VPP can sell to the grid.

[0136] Based on intraday dispatch needs, the distribution network issues new dispatch instructions to the VPP, including the dispatch start time ts, the dispatch duration tc, and the dispatch power Δp. G The VPP, based on its own adjustability, determines and provides feedback on the feasibility of the instruction. m and the corresponding cost f c First, VPP will not issue any new scheduling instructions, i.e., Δp. G When = 0, run the interactive physical model to obtain the pre-scheduling plan and operating cost of each VPP unit, denoted as . F 0The regulations stipulate that when the VPP receives a new dispatch instruction from the distribution network, it cannot increase the load shedding and surplus power supply under the existing load shedding power and pre-sale power plan. That is, constraint (7) is modified to constraint (8).

[0137]

[0138] Under this constraint, when the distribution network issues a new dispatch command, the VPP runs the interactive physical model. If the model has a solution, it can respond to the command while satisfying existing dispatch commands, i.e., f m =1, the cost of this scheduling instruction is f c =FF 0 If the model has no solution, then f m =0,f c =0. For example... Figure 2 As shown in (a).

[0139] Furthermore, in step 2, the VPP simulates new dispatch instructions that the distribution network may issue within a certain period of time based on its own operating status, and obtains a finite data sample set through an interactive physical model. Finally, the samples are trained using data-driven technology to characterize the interaction characteristics between the VPP and the distribution network within a certain period of time. Figure 2 As shown in (b), the VPP can obtain the dispatchable range information of the VPP through the dispatch command feasibility model and provide it to the distribution network, such as Figure 2 (b) is shown by the dashed line. Considering the time-varying nature of wind and solar power output forecasts, the interaction characteristics between the VPP and the distribution network also become time-varying. Therefore, the data-driven model should be updated in real time according to the real-time operating status of the VPP, the latest wind and solar power output forecasts, and the dispatching needs of the distribution network. To meet the timeliness requirements of model updates, a strategy of day-ahead offline training and intraday online correction is adopted. Specifically, as shown... Figure 3As shown: During the offline training phase, the Virtual Power Plant (VPP) uses an interactive physical model to solve for the feasibility and interaction cost of dispatch instructions, generating intraday simulated data as the dataset. The dataset is trained using WKNN and KELM-GPR machine learning methods to determine the model structure and hyperparameters, including the number of neighbors in WKNN, the regularization coefficient in KELM, the kernel function, and the dataset size. During the intraday online correction phase, the VPP decides whether to update the data-driven model parameters based on the dispatch needs of the distribution network dispatch center. The distribution network dispatch center can notify the VPP of any dispatch needs one time period in advance. For example, if the data-driven model needs to be invoked during time period k1, a dispatch invitation is sent to the VPP during time period k1-1. During time period k1, the VPP, based on the previously determined data-driven model structure and hyperparameters and combined with ultra-short-term wind and solar power output forecast data, generates incremental samples as a new dataset through the virtual power plant and distribution network interactive physical model. This dataset is then used to train and update the model parameters, improving the model's prediction accuracy for use by the distribution network dispatch center. If there is no dispatch need, the VPP does not update the model parameters.

[0140] Furthermore, in step 3, based on the adjustability of the VPP, the feasible region for issuing new dispatch instructions to the distribution network is analyzed, that is, the reference boundary for generating new dispatch instructions needs to be known, thereby obtaining the dataset, specifically:

[0141] 1) Reference boundary for generating new scheduling instructions

[0142] The new intraday dispatch instructions issued by the distribution network are [ts, tc, Δp] G Considering that daily dispatching in distribution networks often employs rolling optimization dispatching, executed every 15 minutes with an optimization time window of 4 hours, the data-driven model constructed in this invention characterizes the interaction characteristics between the VPP and the distribution network within the next 4 hours (15-minute time scale, 16 time periods). This model can be adjusted as needed in practical applications. Therefore, the reference ranges for the values ​​of ts and tc are readily apparent, namely:

[0143]

[0144] In the formula, ts min The minimum allowed value for the start time period of scheduling; ts min If ts and tc have the same unit, take the integer part. Let [ts, tc, Δp] be the integer part. G For example, [9,8,1] means that starting from 2:00 (there are 96 time periods in a day, and 2:00-2:15 is the 9th time period), for 8 consecutive time periods (2 hours), the VPP output will increase by 1MW. When tc=1, There are 16 possible cases when tc = 2. There are 15 possible combinations of ts and tc, resulting in 136 possible values. The range of dispatch power for each combination is difficult to characterize analytically due to the temporal coupling of the energy storage battery output within the VPP and the ramp rate constraint of the gas turbine. To find a suitable reference power boundary, this invention ignores the SOC of the energy storage battery and the ramp rate limitation of the gas turbine. Therefore, the maximum possible dispatch power range of the VPP, while satisfying existing dispatch commands, is...

[0145]

[0146] In the formula, This is a pre-scheduling plan for each unit of the VPP in the absence of new scheduling instructions. and Let these be the upper and lower boundaries of the maximum possible output of the VPP. Then, the actual output boundary of the VPP must be contained within these boundaries. Since the new scheduling instructions within the day are based on Δp during the scheduling start and end time period. G Since continuous power output is required, the scheduling power reference boundary can be taken as the minimum value of the maximum possible scheduling power boundary of the VPP during the scheduling start and end periods, in order to further reduce the scheduling power reference range. Therefore, the upper and lower boundaries of the scheduling power corresponding to the i-th case are:

[0147]

[0148] In the formula, ts i , tc i , Let $\mathbf{i}$ be the scheduling start time, scheduling duration, and upper and lower bounds of the scheduling power reference for the $i$ case, respectively. Therefore, the size of the scheduling power range for each case can be calculated as follows:

[0149]

[0150] 2) Simulate new dispatch instructions issued by the distribution network

[0151] After obtaining the reference boundary of the dispatching instruction, new dispatching instructions that may be issued by the distribution network are generated, as follows.

[0152] (1) Uniform generation of scheduling instructions

[0153] Based on the set number of training sets N train The sample size and scheduling power for each case are as follows:

[0154]

[0155] In the formula, b i Let PW be the size of the power range corresponding to the i-th case. iThe proportion of the sum of the power value ranges in all cases; floor(*) is the floor function; n i n is the number of samples taken in the i-th case; error The difference in sample size; mod(*) is the modulo function; sort(*) is the descending sort function; value and index are the descending sort value and the corresponding index number, respectively; index * The first n in the index error Index number. Δp G,i,j Let ΔP be the scheduling power value corresponding to case i, that is, the scheduling power value for case i is uniformly selected at equal intervals within the reference range. G,i,j ts corresponding to i cases i , tc i By combining them, a uniformly generated set of scheduling instructions can be obtained, as shown in equation (14).

[0156]

[0157] In the formula, These represent the sets of the start time period, duration period, and power of the scheduling instructions in the training set, respectively. These represent the start time period, scheduling duration, and scheduling power of the l-th scheduling instruction in the training set, respectively.

[0158] (2) Random generation of scheduling instructions

[0159] The scheduling instruction reference boundary and the number N of the test set, as known from the preceding text, are... test Through random functions in ts, tc, Δp G The scheduling instructions are generated by randomly sampling within the reference value range and must satisfy the constraint of equation (9). Then, a single randomly generated scheduling instruction is...

[0160]

[0161] In the formula, U[ts min ,ts min +15] indicates that in [ts min ,ts min Random sampling within the interval [+15] Let be the start time period, scheduling duration period, and scheduling power of the l-th scheduling instruction in the test set, respectively. Then, the randomly generated scheduling instruction set can be obtained as shown in equation (16), where These represent the sets of the start time period, duration period, and power of the scheduling instructions in the test set, respectively.

[0162]

[0163] 3) The VPP responds to scheduling instructions and generates a dataset.

[0164] After obtaining the dispatch instructions from equations (14) and (16), according to the method for determining whether the VPP responds to the distribution network dispatch instructions as described in claim 2, the feasibility and interaction cost of the dispatch instructions can be obtained through the interactive physical model, as shown in equations (17)-(20). Where, These are the sets of scheduling instruction feasibility and interaction costs, respectively, for the training and testing sets. These represent the scheduling instruction feasibility and interaction cost corresponding to the l-th scheduling instruction in the training set, respectively. These represent the feasibility of the scheduling instruction and the interaction cost corresponding to the l-th scheduling instruction in the test set, respectively.

[0165]

[0166]

[0167]

[0168]

[0169] In summary, the dataset required for model training is:

[0170]

[0171]

[0172] In the formula, x m x c These are the datasets required for the scheduling instruction feasibility model and the interaction cost model, respectively.

[0173] Furthermore, in step 4, the dataset is preprocessed, specifically as follows:

[0174] For the feasibility model of scheduling instructions: to eliminate the impact of differences in the units of measurement in the dataset on model training, the dataset needs to be normalized to a value between [-1, 1]. The calculation formula is as follows:

[0175]

[0176] In the formula, x represents the unnormalized original data. * The dataset is normalized; min(x) is the minimum value in the original dataset; max(x) is the maximum value in the original dataset. For the interaction cost model: first, abnormal data where scheduling instructions are not executable are removed from the interaction cost dataset; second, the scheduling instructions in the dataset are transformed from 3-dimensional features to 16-dimensional features to allow the interaction cost model to fully explore the complex nonlinear relationship between scheduling instructions and interaction costs. The transformation formula is as follows:

[0177]

[0178] In the formula, zeros(ts) l -ts min ,1) indicates generation (ts l -ts min A zero matrix of 1×1; ones(tc l ,1) indicates the generation of tc l A × 1 matrix consisting entirely of 1s; l This represents the l-th transformed scheduling instruction. Finally, the transformed dataset is normalized according to equation (23).

[0179] Furthermore, in step 5, a feasibility model for scheduling instructions is established using the Weighted K-Nearest Neighbors (WKNN) algorithm, specifically as follows:

[0180] Determining the feasibility of scheduling instructions is a binary classification problem. The powerful classification capabilities of existing machine learning models can approximate the actual output boundary of the Virtual Power Plant (VPP). WKNN is an easy-to-implement and highly effective learning classifier, and this invention applies WKNN to construct a scheduling instruction feasibility model. For the key elements of WKNN, Euclidean distance is used as the distance metric, and the inverse squared distance is used as the distance weight. The number of neighbors, *r*, has a significant impact on the model's prediction accuracy. To avoid the inefficiency of manually selecting parameters, the model's prediction accuracy is used as the objective function, and the Particle Swarm Optimization (PSO) algorithm is employed for optimization. Considering the large number of feasible scheduling instructions in the training set, leading to imbalanced training samples, the AUC index in the confusion matrix, which is insensitive to sample imbalance, is chosen as the evaluation criterion for model prediction accuracy. The training process of the scheduling instruction feasibility model is as follows: Figure 4 As shown.

[0181] During the initial training, after data preprocessing, the number of neighboring points in the model was first optimized using the particle swarm optimization algorithm. Then, the classification model was trained and tested. If the prediction accuracy of the classification model did not meet the requirements, the above steps were repeated by increasing the dataset size until the required prediction accuracy was achieved. This process determined the number of neighboring points and the required dataset size. In real-time operation, based on the model structure and hyperparameters obtained from the initial training, combined with the real-time running status of the VPP and ultra-short-term wind and light predictions, incremental samples were generated as a new dataset through the interactive physics model to train and update the classification model parameters. The trained scheduling instruction feasibility model is denoted as...

[0182]

[0183] Considering that the scheduling instructions in the dataset are generated based on the scheduling instruction reference boundary, in some cases, instructions on the reference boundary can still be satisfied, making the datasets in these cases all feasible. Therefore, a classification model trained on these datasets can effectively determine whether instructions inside the reference boundary are feasible, but it is highly likely to classify instructions outside the reference boundary as feasible. To address this problem, the obtained scheduling instruction reference boundary is also used as a model to determine the feasibility of scheduling instructions; that is, instructions outside the reference boundary are infeasible, and instructions inside the reference boundary are feasible, denoted as .

[0184]

[0185] However Unable to effectively identify commands within the reference boundary, this paper will... and By complementing each other's strengths, a feasibility model for the final scheduling instructions was obtained:

[0186]

[0187] According to this model, VPP can know model f M By defining the classification boundaries, the dispatchable range of the VPP can be determined and provided to the distribution network.

[0188] Furthermore, in step 6, a virtual power plant interaction cost model based on Kernel Extreme Learning Machine-Gaussian Process Regression (KELM-GPR) is constructed, specifically as follows:

[0189] The interaction cost model aims to establish a complex nonlinear mapping relationship between scheduling instructions and interaction costs, belonging to the regression model category. Considering the strong nonlinear fitting capabilities of KELM and GPR, and the advantages of KELM such as fast learning, good generalization, and no need to specify the number of hidden layer nodes, while GPR has good adaptability to handling complex problems such as nonlinearity and high dimensionality, this invention uses KELM to establish the interaction cost model. To further reduce the model's prediction error, GPR is introduced as an error compensation model to improve the prediction accuracy of the interaction cost model. This invention uses typical radial basis functions and exponential covariance functions as the kernel functions of KELM and GPR, respectively, and uses the mean absolute percentage error (MAPE) and coefficient of determination (R²) as evaluation indicators of the model's prediction performance. The interaction cost model training process is as follows... Figure 5 As shown.

[0190] During the initial training, the original dataset was first preprocessed. Then, the regularization coefficient C of KELM and the kernel parameter σ of the kernel function were optimized using the particle swarm optimization algorithm. Finally, the interaction cost model was trained and tested. If the model's prediction performance did not meet the requirements, the dataset size was increased and the above steps were repeated until the model's prediction performance met the requirements. During real-time operation, based on the model structure and hyperparameters obtained from the initial training, combined with the real-time running status of VPP and ultra-short-term wind and light predictions, incremental samples were generated through the interaction physics model as a new dataset to train and update the interaction cost model parameters. The trained interaction cost model is denoted as […].

[0191]

[0192] Provided that the training indicators are met, the dataset size required for daily updates of the virtual power plant interaction model is the maximum of the dataset sizes required for updates of the scheduling instruction feasibility model and the interaction cost model.

[0193] Example:

[0194] Example: Create as shown in the attached document Figure 1 Taking the VPP in the resource structure diagram shown as an example, the wind and solar power output forecast and dispatch command load curves for a typical day of VPP are as follows: Figure 6 As shown in Table A1, the main parameters of the equipment within the VPP are as follows: the time-of-use electricity price is RMB 1390.8 / (MWh) during peak hours (11:00-15:00, 18:00-21:00), RMB 347.7 / (MWh) during off-peak hours (00:00-07:00, 23:00-24:00), and RMB 869.25 / (MWh) during normal hours; the surplus electricity price is RMB 200 / (MWh). The maximum load reduction is 10% of the day-ahead dispatch command load curve, with a dispatch unit time interval of 15 minutes. The simulation environment is Matlab 2021a, CPU Ryzen 7 5800H, RAM 16GB, and the interactive physics model is solved using the cplex optimizer. The VPP day-ahead dispatch plan is as follows. Figure 7 As shown, the scheduling cost is 53,007 yuan, which is the pre-scheduling plan and scheduling cost of the VPP under the condition that no new scheduling instructions are issued by the distribution network.

[0195] Table A1 VPP Operating Parameters

[0196]

[0197]

[0198] According to the technical solution described in the manual, a dataset was acquired during the period from 0:00 to 4:00, and the scheduling instruction feasibility model and interaction cost model were trained. The ratio of training to test set samples in the dataset was set at 4:1. The training accuracy target for the scheduling instruction feasibility model was AUC = 0.99, and the training accuracy target for the interaction cost model was MAPE = 0.25% and R² = 0.99. These can be adjusted as needed in practical applications.

[0199] 1) Training of the feasibility model for scheduling instructions

[0200] The model was trained using WKNN. The training and testing results are shown in Table 1. The training time does not include the time spent optimizing model parameters using PSO; similarly, Table 2 shows the results.

[0201] Table 1. Training and testing results of the WKNN-based scheduling instruction feasibility model.

[0202]

[0203] As shown in Table 1, the AUC value increases with the number of samples (dataset size). A higher AUC value indicates a better model classification performance. When the number of samples is 2500, the model's AUC metric reaches the training target value, and the training time is less than 3.6 seconds. Therefore, during intraday online correction, 2500 samples can be incrementally generated as a new dataset to train and update the model parameters.

[0204] 2) Training the interaction cost model

[0205] The interaction cost model was trained using KELM-GPR, and the training and testing results of the interaction cost model are shown in Table 2.

[0206] Table 2. Training and testing results of the interaction cost model based on KELM-GPR

[0207]

[0208]

[0209] As shown in Table 2, the MAPE of the model decreases with the increase of the number of samples, while the R² gradually increases, indicating that the model's prediction performance is getting better and better. When the number of samples is 3000, the model's prediction index has reached the training target value, and the training time is only 3.68 seconds. Therefore, for intraday online correction, KELM with C=2.7 and σ=7762, combined with the GPR error compensation model, can be used to incrementally generate 3000 samples as a new dataset to train and update the model parameters.

[0210] In summary, the number of samples required for online updates of the data-driven model should simultaneously meet the needs of both the scheduling instruction feasibility model and the interaction cost model, and should be set to 3000.

[0211] 3) Model Testing

[0212] Based on the pre-scheduling plan for VPP to issue new dispatching instructions without distribution network, several new intraday dispatching instructions issued by distribution network are simulated, and the trained virtual power plant interaction model and the traditional interaction physical model in step 1 of the instruction manual are used to respond to them respectively. The response results are shown in Table 3.

[0213] Table 3 Test Results of Virtual Power Plant Interaction Model

[0214]

[0215] In Table 3, the dispatch instruction [1,4,-1.65] indicates that from 0:00 (there are 96 time periods in a day, with 0:00-0:15 being the first time period), for four consecutive time periods (1 hour), the VPP output is reduced by 1.65MW. During this time period, the dispatchable range of the VPP corresponds to [-0.255,1.956] in Table 3. The other dispatch instructions follow the same pattern. As shown in Table 3, the virtual power plant interaction model has the same interaction characteristics as the traditional physical interaction model. Its response time to a single dispatch instruction is in the millisecond range, which is significantly better than the latter. Moreover, compared to the traditional physical interaction model, which is not easy to analytically characterize the dispatchable boundary of the VPP, the virtual power plant interaction model can provide information on the dispatchable range for the corresponding dispatch period, thereby better meeting the timeliness requirements of the upper-level distribution network dispatch. At the same time, since the virtual power plant interaction model is a data-driven model, it is equivalent to a "black box" externally, which can protect the information security inside the VPP and is suitable for the modeling needs in a competitive market environment. In practical applications, the incremental samples generated during the virtual power plant interactive model update process are obtained through parallel computing to reduce model update time.

[0216] In another embodiment of the present invention, a virtual power plant interaction model construction system is provided, which can be used to implement the above-described virtual power plant interaction model construction method. Specifically, the virtual power plant interaction model construction system includes:

[0217] The interactive physical model construction module is used to establish an interactive physical model between the virtual power plant and the distribution network. The virtual power plant determines whether to respond to the new intraday dispatch instructions issued by the distribution network based on its remaining adjustable capacity.

[0218] The data acquisition module is used to obtain the training set by uniformly generating scheduling instructions and to obtain the test set by randomly generating scheduling instructions;

[0219] The data processing module is used to normalize the dataset and establish a feasibility model for scheduling instructions on the normalized dataset.

[0220] The interaction cost model construction module is used to construct a virtual power plant interaction cost model based on Kernel Extreme Learning Machine-Gaussian Process Regression (KELM-GPR) based on the scheduling instruction feasibility model.

[0221] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0222] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a virtual power plant interactive model construction method.

[0223] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the operating system of the terminal. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the virtual power plant interaction model construction method in the above embodiments.

[0224] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0225] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0226] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0227] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0228] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing a virtual power plant interaction model, characterized in that, include: Establish an interactive physical model between the virtual power plant and the distribution network. The virtual power plant determines whether to respond to new intraday dispatch instructions issued by the distribution network based on its remaining adjustable capacity. The training set is obtained by uniformly generating scheduling instructions, and the test set is obtained by randomly generating scheduling instructions. The dataset is normalized, and a feasibility model for scheduling instructions is established for the normalized dataset. A virtual power plant interaction cost model based on KELM-GPR (Kernel Extreme Learning Machine-Gaussian Process Regression) is constructed based on the scheduling instruction feasibility model. Establish an interactive physical model of the virtual power plant and the distribution network: The virtual power plant, which utilizes surplus electricity resources, adopts a strategy of feeding surplus electricity into the grid. Its physical interaction model with the distribution network is as follows: Objective function: (1) (2) In the formula, The operating cost of the virtual power plant for the remaining time of the day. For the current time period, if this interaction model had been run a day earlier, then ; Number of time periods within a day; , , , , , , They represent The operating and management costs of wind power, photovoltaic power, energy storage batteries, gas turbines, demand response, load shedding penalties, and surplus power to the grid are included. , , , , , , , These are the wind power operation and management cost coefficient, photovoltaic operation and management cost coefficient, energy storage battery operation and management cost coefficient, gas turbine operation cost coefficient, gas turbine fixed cost coefficient, demand response compensation cost coefficient, and off-load penalty coefficient (5 times the time-of-use electricity price). The on-grid electricity price for surplus electricity during certain periods; , , , , , , They are respectively Wind power output, photovoltaic power output, energy storage battery output, gas turbine output, load reduction output, off-load power, and surplus power fed into the grid during specific time periods; The duration of a time period; Constraints: Virtual power plant power balance constraints (3) In the formula, These are new dispatch instructions to be issued by the distribution network within a certain period of time in the future. For the loads already determined by the distribution network dispatch instructions; Energy storage battery operating constraints (4) In the formula, , For energy storage batteries Discharge power and charging power during the time period; This represents the maximum charge and discharge power of the energy storage battery. , These represent the charge and discharge states of the energy storage battery, and are variables ranging from 0 to 1. , These are the charge and discharge efficiencies of the energy storage battery; This refers to the rated capacity of the energy storage battery. For energy storage batteries State of charge over a period of time; , These represent the minimum and maximum permissible states of charge of the energy storage battery, respectively. , These represent the initial and final states of charge of the energy storage battery during a day; Gas turbine operating constraints (5) In the formula, It is a 0-1 variable, representing the start-up and shutdown status of the gas turbine, which is determined by the day-ahead dispatch of the virtual power plant; , These are the minimum and maximum power outputs of the gas turbine, respectively. , These represent the downward and upward ramp rates of the gas turbine, respectively. Reduced load operating constraints (6) In the formula, In order to be in During certain periods, the maximum output power of the load can be reduced; Electricity sales constraints (7) In the formula, The upper limit of the power output of a virtual power plant to the grid. The distribution network issues new dispatch instructions to the virtual power plant based on intraday dispatch needs, including the dispatch start time period. Scheduling duration and scheduling power The virtual power plant assesses and responds to the feasibility of the instruction based on its own adjustability. and the corresponding costs Firstly, the virtual power plant will not issue new dispatch instructions. In this case, the interactive physical model is run to obtain the pre-scheduling plan and operating cost of each unit of the virtual power plant, denoted as , , , , , The regulations stipulate that when a virtual power plant receives a new dispatch instruction from the distribution network, it cannot increase the load shedding and surplus power grid connection under the existing load shedding power and pre-sale power plan; that is, constraint (7) is modified to constraint (8). (8) Under this constraint, when the distribution network issues a new dispatch command, the virtual power plant operates an interactive physical model. If the model has a solution, it can respond to the new command while satisfying existing dispatch commands. The cost corresponding to this scheduling instruction is ; If the model has no solution, then ; For the normalized dataset, establish a feasibility model for scheduling instructions: A scheduling command feasibility model is established using the weighted k-nearest neighbor (WKNN) algorithm. The WKNN model is then applied to construct the scheduling command feasibility model. For the key elements of WKNN, Euclidean distance is used as the distance metric, and the inverse squared distance is used as the distance weight. The model prediction accuracy is used as the objective function, and the Particle Swarm Optimization (PSO) algorithm is employed for optimization. The AUC index in the confusion matrix, which is insensitive to sample imbalance, is selected as the evaluation criterion for model prediction accuracy. The training process of the scheduling command feasibility model is as follows: During the initial training, after data preprocessing, the number of neighboring points in the model was first optimized using the particle swarm optimization algorithm. Then, the classification model was trained and tested. If the prediction accuracy of the classification model did not meet the requirements, the above steps were repeated by increasing the dataset size until the required prediction accuracy was achieved. This process determined the number of neighboring points and the required dataset size. In real-time operation, based on the model structure and hyperparameters obtained from the initial training, combined with the real-time operating status of the virtual power plant and ultra-short-term wind and solar forecasts, incremental samples were generated as a new dataset through an interactive physical model to train and update the classification model parameters. The trained scheduling instruction feasibility model is denoted as... (25) The obtained scheduling instruction reference boundary is also used as a model to determine the feasibility of scheduling instructions; that is, instructions outside the reference boundary are infeasible, and instructions inside the reference boundary are feasible, denoted as . (26) Will and By complementing each other's strengths, a feasibility model for the final scheduling instructions was obtained: (27) According to this model, the virtual power plant can know the model. The classification boundaries are used to determine the dispatchable range of the virtual power plant and provide this information to the distribution network. A virtual power plant interaction cost model based on KELM-GPR (Kind Extreme Learning Machine-Gaussian Process Regression) is constructed as follows: A KELM-based interaction cost model was established, and the GPR (Growth Percentage Reduction) model was introduced as an error compensation model to improve the prediction accuracy of the interaction cost model. Typical radial basis functions and exponential covariance functions were used as the kernel functions of KELM and GPR, respectively. Mean absolute percentage error (MAPE) and coefficient of determination (R²) were used as evaluation metrics for the model's prediction performance. The training process of the interaction cost model is as follows: In the training process, the original dataset was first preprocessed, and then the regularization coefficients of KELM were adjusted using the particle swarm optimization algorithm. Kernel parameters of kernel function Optimization is performed, and finally, the interaction cost model is trained and tested. If the model's prediction performance does not meet the requirements, the dataset size is increased and the above steps are repeated until the model's prediction performance meets the requirements. During intraday real-time operation, based on the model structure and hyperparameters trained the previous day, combined with the real-time operating status of the virtual power plant and ultra-short-term wind and solar forecasts, incremental samples are generated through the interaction physics model as a new dataset to train and update the interaction cost model parameters. The trained interaction cost model is denoted as […]. (28) Provided that the training indicators are met, the dataset size required for daily updates of the virtual power plant interaction model is the maximum of the dataset sizes required for updates of the scheduling instruction feasibility model and the interaction cost model.

2. The method for constructing a virtual power plant interaction model according to claim 1, characterized in that, The virtual power plant interaction model adopts a strategy of offline training before the day and online correction during the day: The virtual power plant simulates new dispatch instructions that the distribution network may issue within a certain period of time in the future based on its own operating status, and obtains a limited data sample set by solving through an interactive physical model. Finally, the data-driven technology is used to learn and train the samples to characterize the interaction characteristics between the virtual power plant and the distribution network within a certain period of time in the future; a strategy of day-ahead offline training and intraday online correction is adopted. Specifically: During the offline training phase, the virtual power plant solves the feasibility and interaction cost of dispatch instructions through an interactive physical model, generating intraday simulation data as the dataset; the dataset is then trained using WKNN and KELM-GPR in machine learning to determine the model structure and hyperparameters, including the number of neighbors in WKNN, the regularization coefficient of KELM, the kernel function, and the dataset size; during the intraday online correction phase, the virtual power plant decides whether to update the data-driven model parameters based on the dispatch requirements of the distribution network dispatch center. The distribution network dispatch center should notify the user of any dispatching needs one time in advance, such as... If a data-driven model needs to be invoked during a specific time period, then... The time period sends a dispatch invitation to the virtual power plant, and the virtual power plant then... During a given period, based on the data-driven model structure and hyperparameters determined in previous training and combined with ultra-short-term wind and solar power output forecast data, incremental samples are generated as new datasets through the interaction physical model between the virtual power plant and the distribution network. The model parameters are then trained and updated to improve the model's prediction accuracy for use by the distribution network dispatch center. If there is no dispatch requirement, the virtual power plant will not update the model parameters.

3. The method for constructing a virtual power plant interaction model according to claim 1, characterized in that, The training set is obtained by uniformly generating scheduling instructions, and the test set is obtained by randomly generating scheduling instructions. Based on the adjustability of the virtual power plant, the feasible region for issuing new dispatch instructions to the distribution network is analyzed, that is, the reference boundary for generating new dispatch instructions needs to be known, in order to obtain the dataset, specifically: 1) Generate the reference boundary for new scheduling instructions The new intraday dispatch instructions issued by the distribution network are ; , Reference range for values: (9) In the formula, This is the minimum value allowed for the start of the scheduling period; , , Take the integer part if the units are the same; when hour, There are 16 situations when hour, There are 15 possible scenarios. and There are 136 possible combinations. Ignoring the SOC of the energy storage battery and the ramp rate limit of the gas turbine, the maximum possible dispatch power range of the virtual power plant, based on existing dispatch instructions, is: (10) In the formula, , , , , Pre-dispatch plans for each unit of the virtual power plant in the absence of new dispatch instructions; and Let these be the upper and lower boundaries of the virtual power plant's maximum possible output. Then, the actual output boundary of the virtual power plant must be included within these boundaries. Because the new scheduling instructions within the day are within the scheduling start and end time period... Continuous power output means that the dispatch power reference boundary can be taken as the minimum value of the maximum possible dispatch power boundary of the virtual power plant within the dispatch start and end time period, in order to further reduce the dispatch power reference range; then the... The upper and lower bounds of the scheduling power corresponding to this case are: (11) In the formula, , , , The first The scheduling start time, scheduling duration, and upper and lower bounds for each of the following scenarios are used to determine the scheduling power range for each scenario. (12) 2) New dispatch instructions issued by the simulated distribution network After obtaining the reference boundary of the dispatching instructions, new dispatching instructions that may be issued by the distribution network are generated, as follows: (1) Uniform generation of scheduling instructions Based on the set number of training sets The sample size and scheduling power for each case are as follows: (13) In the formula, For the first The size of the power range corresponding to this case The proportion of the sum of the power value ranges in all cases; This is the floor function; For the first Number of samples taken in this case; This represents the difference in the number of samples taken. It is the remainder function; It is a function for sorting in descending order; , These are the values ​​sorted in descending order and their corresponding index numbers; for Center front Each index number; for The corresponding scheduling power value under this condition is, i.e. The scheduling power reference range corresponding to this case is taken at equal intervals and uniformly; and The corresponding situation , By combining them, a uniformly generated set of scheduling instructions can be obtained, as shown in equation (14). (14) In the formula, , , These represent the sets of the start time period, duration period, and power of the scheduling instructions in the training set, respectively. , , The training set The start time, duration, and power of each scheduling instruction; (2) Random generation of scheduling instructions The scheduling instruction reference boundary and the number of test sets set, as known from the preceding text, are as follows. Through a random function , , The scheduling instructions are generated by randomly sampling within the reference value range and must satisfy the constraint of equation (9); then a single randomly generated scheduling instruction is (15) In the formula, Indicates in Random sampling within an interval , , The test set number The scheduling instruction start time, scheduling duration, and scheduling power are used to obtain a randomly generated set of scheduling instructions as shown in equation (16), where, , , These represent the sets of the start time period, duration period, and power of the scheduling instructions in the test set, respectively. (16) 3) The virtual power plant responds to dispatch instructions and generates a dataset. After obtaining the dispatch instructions from equations (14) and (16), the feasibility and interaction cost of the dispatch instructions can be obtained through the interactive physical model based on the method for determining whether the virtual power plant responds to the distribution network dispatch instructions, as shown in equations (17)-(20); where, , , , These are the sets of scheduling instruction feasibility and interaction costs, respectively, for the training and testing sets. , The training set The feasibility and interaction cost of each scheduling instruction; , The test set number The feasibility and interaction cost of each scheduling instruction; (17) (18) (19) (20) In summary, the dataset required for model training is: (21) (22) In the formula, , These are the datasets required for the scheduling instruction feasibility model and the interaction cost model, respectively.

4. The method for constructing a virtual power plant interaction model according to claim 1, characterized in that, Normalize the dataset: For the feasibility model of scheduling instructions, the dataset is normalized to a range of [-1, 1]. The calculation formula is as follows: (23) In the formula, The original data is unnormalized. This is the normalized dataset; It is the minimum value in the original dataset; The maximum value in the original dataset; for the interaction cost model: first, remove outlier data where scheduling instructions are inexecutable from the interaction cost dataset; second, transform the scheduling instructions in the dataset from 3-dimensional features to 16-dimensional features, so that the interaction cost model can fully explore the complex nonlinear relationship between scheduling instructions and interaction costs, and the transformation formula is: (24) In the formula Indicates generation The zero matrix; Indicates generation A matrix consisting entirely of 1s; Indicates the first The transformed scheduling instructions are then processed according to equation (23) to normalize the transformed dataset.

5. A virtual power plant interactive model construction system, characterized in that, include: The interactive physical model construction module is used to establish an interactive physical model between the virtual power plant and the distribution network. The virtual power plant determines whether to respond to the new intraday dispatch instructions issued by the distribution network based on its remaining adjustable capacity. The data acquisition module is used to obtain the training set by uniformly generating scheduling instructions and to obtain the test set by randomly generating scheduling instructions; The data processing module is used to normalize the dataset and establish a feasibility model for scheduling instructions on the normalized dataset. The interaction cost model construction module is used to construct a virtual power plant interaction cost model based on KELM-GPR (Kernel Extreme Learning Machine-Gaussian Process Regression) based on the scheduling instruction feasibility model. Establish an interactive physical model of the virtual power plant and the distribution network: The virtual power plant, which utilizes surplus electricity resources, adopts a strategy of feeding surplus electricity into the grid. Its physical interaction model with the distribution network is as follows: Objective function: (1) (2) In the formula, The operating cost of the virtual power plant for the remaining time of the day. For the current time period, if this interaction model had been run a day earlier, then ; Number of time periods within a day; , , , , , , They represent The operating and management costs of wind power, photovoltaic power, energy storage batteries, gas turbines, demand response, load shedding penalties, and surplus power to the grid are included. , , , , , , , These are the wind power operation and management cost coefficient, photovoltaic operation and management cost coefficient, energy storage battery operation and management cost coefficient, gas turbine operation cost coefficient, gas turbine fixed cost coefficient, demand response compensation cost coefficient, and off-load penalty coefficient (5 times the time-of-use electricity price). The on-grid electricity price for surplus electricity during certain periods; , , , , , , They are respectively Wind power output, photovoltaic power output, energy storage battery output, gas turbine output, load reduction output, off-load power, and surplus power fed into the grid during specific time periods; The duration of a time period; Constraints: Virtual power plant power balance constraints (3) In the formula, These are new dispatch instructions to be issued by the distribution network within a certain period of time in the future. For the loads already determined by the distribution network dispatch instructions; Energy storage battery operating constraints (4) In the formula, , For energy storage batteries Discharge power and charging power during the time period; This represents the maximum charge and discharge power of the energy storage battery. , These represent the charge and discharge states of the energy storage battery, and are variables ranging from 0 to 1. , These are the charge and discharge efficiencies of the energy storage battery; This refers to the rated capacity of the energy storage battery. For energy storage batteries State of charge over a period of time; , These represent the minimum and maximum permissible states of charge of the energy storage battery, respectively. , These represent the initial and final states of charge of the energy storage battery during a day; Gas turbine operating constraints (5) In the formula, It is a 0-1 variable, representing the start-up and shutdown status of the gas turbine, which is determined by the day-ahead dispatch of the virtual power plant; , These are the minimum and maximum power outputs of the gas turbine, respectively. , These represent the downward and upward ramp rates of the gas turbine, respectively. Reduced load operating constraints (6) In the formula, In order to be in During certain periods, the maximum output power of the load can be reduced; Electricity sales constraints (7) In the formula, The upper limit of the power output of a virtual power plant to the grid. The distribution network issues new dispatch instructions to the virtual power plant based on intraday dispatch needs, including the dispatch start time period. Scheduling duration and scheduling power The virtual power plant assesses and responds to the feasibility of the instruction based on its own adjustability. and the corresponding costs Firstly, the virtual power plant will not issue new dispatch instructions. In this case, the interactive physical model is run to obtain the pre-scheduling plan and operating cost of each unit of the virtual power plant, denoted as , , , , , The regulations stipulate that when a virtual power plant receives a new dispatch instruction from the distribution network, it cannot increase the load shedding and surplus power grid connection under the existing load shedding power and pre-sale power plan; that is, constraint (7) is modified to constraint (8). (8) Under this constraint, when the distribution network issues a new dispatch command, the virtual power plant operates an interactive physical model. If the model has a solution, it can respond to the new command while satisfying existing dispatch commands. The cost corresponding to this scheduling instruction is ; If the model has no solution, then ; For the normalized dataset, establish a feasibility model for scheduling instructions: A scheduling command feasibility model is established using the weighted k-nearest neighbor (WKNN) algorithm. The WKNN model is then applied to construct the scheduling command feasibility model. For the key elements of WKNN, Euclidean distance is used as the distance metric, and the inverse squared distance is used as the distance weight. The model prediction accuracy is used as the objective function, and the Particle Swarm Optimization (PSO) algorithm is employed for optimization. The AUC index in the confusion matrix, which is insensitive to sample imbalance, is selected as the evaluation criterion for model prediction accuracy. The training process of the scheduling command feasibility model is as follows: During the initial training, after data preprocessing, the number of neighboring points in the model was first optimized using the particle swarm optimization algorithm. Then, the classification model was trained and tested. If the prediction accuracy of the classification model did not meet the requirements, the above steps were repeated by increasing the dataset size until the required prediction accuracy was achieved. This process determined the number of neighboring points and the required dataset size. In real-time operation, based on the model structure and hyperparameters obtained from the initial training, combined with the real-time operating status of the virtual power plant and ultra-short-term wind and solar forecasts, incremental samples were generated as a new dataset through an interactive physical model to train and update the classification model parameters. The trained scheduling instruction feasibility model is denoted as... (25) The obtained scheduling instruction reference boundary is also used as a model to determine the feasibility of scheduling instructions; that is, instructions outside the reference boundary are infeasible, and instructions inside the reference boundary are feasible, denoted as . (26) Will and By complementing each other's strengths, a feasibility model for the final scheduling instructions was obtained: (27) According to this model, the virtual power plant can know the model. The classification boundaries are used to determine the dispatchable range of the virtual power plant and provide this information to the distribution network. A virtual power plant interaction cost model based on KELM-GPR (Kind Extreme Learning Machine-Gaussian Process Regression) is constructed as follows: A KELM-based interaction cost model was established, and the GPR (Growth Percentage Reduction) model was introduced as an error compensation model to improve the prediction accuracy of the interaction cost model. Typical radial basis functions and exponential covariance functions were used as the kernel functions of KELM and GPR, respectively. Mean absolute percentage error (MAPE) and coefficient of determination (R²) were used as evaluation metrics for the model's prediction performance. The training process of the interaction cost model is as follows: In the training process, the original dataset was first preprocessed, and then the regularization coefficients of KELM were adjusted using the particle swarm optimization algorithm. Kernel parameters of kernel function Optimization is performed, and finally, the interaction cost model is trained and tested. If the model's prediction performance does not meet the requirements, the dataset size is increased and the above steps are repeated until the model's prediction performance meets the requirements. During intraday real-time operation, based on the model structure and hyperparameters trained the previous day, combined with the real-time operating status of the virtual power plant and ultra-short-term wind and solar forecasts, incremental samples are generated through the interaction physics model as a new dataset to train and update the interaction cost model parameters. The trained interaction cost model is denoted as […]. (28) Provided that the training indicators are met, the dataset size required for daily updates of the virtual power plant interaction model is the maximum of the dataset sizes required for updates of the scheduling instruction feasibility model and the interaction cost model.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the virtual power plant interactive model construction method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the virtual power plant interactive model construction method as described in any one of claims 1 to 4.

Images