A model data double-driven security constrained unit commitment decision method and system

Abstract

This invention discloses a model-data dual-driven method and system for safety-constrained unit combination decision-making, relating to the field of power system optimization and dispatching technology. The method includes the following steps: Based on active and reactive load scenarios generated from ellipsoidal uncertain sets, a multi-scenario AC power flow constrained unit combination model is constructed; a neural network surrogate model is constructed, taking unit start-up and shutdown states and topology enhancement features as inputs and node voltage exceedances as outputs; the master problem of multi-scenario DC safety-constrained unit combination is solved to obtain initial unit start-up and shutdown schemes and output plans; the trained neural network surrogate model is used to quickly predict the node voltage exceedance situations in multiple scenarios within the initial unit start-up and shutdown schemes, and the prediction results are verified to obtain the final unit start-up and shutdown scheme. This invention overcomes the problems of existing AC-SCUC solutions, such as difficulty in balancing accuracy and efficiency, and insufficient adaptability to multiple scenarios.

Description

Technical Field

[0001] This invention relates to the field of power system optimization and dispatching technology, and in particular to a model-data dual-driven method and system for safety-constrained unit combination decision-making. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Security-constrained unit commitment is the core of day-ahead dispatch, aiming to minimize costs while meeting security constraints. The integration of renewable energy sources exacerbates uncertainty. Traditional DC power flow security-constrained unit commitment (SCUC) ignores reactive voltage, making it difficult to guarantee AC feasibility; while precise AC power flow unit commitment (AC-SCUC) is a non-convex, nonlinear mixed-integer programming problem, difficult to solve. Existing physical model-driven methods (such as linearization and convex relaxation) offer high solution security but suffer from approximation errors and low efficiency at large scales; data-driven methods (such as neural networks) are fast but lack security and generalization capabilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the purpose of this invention is to provide a model-data dual-driven safety constraint unit combination decision-making method and system, in order to overcome the problems of existing technologies where AC-SCUC solutions are difficult to balance accuracy and efficiency, and lack adaptability to multiple scenarios.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a model-data dual-driven safety-constrained unit combination decision-making method, comprising the following steps: Based on the active and reactive load scenarios generated by the ellipsoidal uncertainty set, a multi-scenario AC power flow constraint unit combination model is constructed according to the active and reactive load scenarios as the main problem of multi-scenario DC safety constraint unit combination. A neural network surrogate model is constructed with unit start-up and shutdown status and topology enhancement features as inputs and node voltage over-limit as output. Through multi-objective training, the neural network surrogate model is made to approximate the optimal AC power flow mapping. The solution is obtained by a two-stage iterative method with analog and digital dual drive. First, the main problem of unit combination under DC safety constraints in multiple scenarios is solved to obtain the initial unit start-up and shutdown scheme and output plan. Then, a trained neural network surrogate model is used to quickly predict the voltage over-limit situation of each node in multiple scenarios in the initial unit start-up and shutdown scheme, and the prediction results are verified to obtain the final unit start-up and shutdown scheme.

[0006] Furthermore, based on the ellipsoidal uncertainty set, active and reactive load scenario sets are generated. The covariance matrix is ​​used to characterize the intensity and correlation of multi-node load fluctuations, and the overall amplitude of the disturbance is controlled by the ellipsoidal radius to generate a finite number of active and reactive load scenarios.

[0007] Furthermore, the multi-scenario DC safety constraint unit combination master problem aims to minimize the expected operating cost in multiple scenarios. The constraints include power balance constraints, DC power flow constraints, unit output upper and lower limit constraints, unit minimum start-up and shutdown time constraints, unit spinning reserve constraints, and unit ramping constraints.

[0008] Furthermore, the topology enhancement features include the start-up and shutdown status of units in each time period, the system load level, and the topology enhancement features composed of the coordination features of online units, load coverage features, spatial dispersion features, and the number of online units.

[0009] Furthermore, the specific steps for making the neural network surrogate model approximate the optimal power flow mapping through multi-objective training are as follows: The neural network proxy model adopts a multi-layer feedforward neural network structure, and through offline training, the neural network approximates the nonlinear mapping relationship of AC optimal power flow calculation. The neural network agent model training process employs multi-objective loss functions, including Huber loss, quantile loss, and binary cross-entropy loss.

[0010] Furthermore, the neural network agent model adopts a time-sharing training strategy, that is, it trains independent neural network models for each of the 24 time periods of the day, and the network for each time period is trained only for the load level and operating characteristics of that time period.

[0011] Furthermore, the analog-digital dual-drive two-stage iterative solution method includes a physical model-driven main problem module based on multi-scenario DC safety constraint unit combination, a data-driven sub-problem module based on neural network multi-scenario AC power flow recovery, a sensitivity calculation module, a feedback constraint generation module, and a physical verification module based on accurate AC power flow. The main problem module and sub-problem modules iterate in a closed loop using unit start-up / shutdown schemes and voltage over-limit feedback information. The main problem module outputs the unit start-up / shutdown scheme to the sub-problem module. The sub-problem module quickly predicts voltage over-limits using a neural network surrogate model; if no over-limit occurs, it proceeds to the physical verification module; if an over-limit occurs, it generates over-limit information and transmits it to the sensitivity calculation module. The sensitivity calculation module calculates the sensitivity matrix of node voltage to unit reactive power output, and the feedback constraint generation module constructs linear feedback constraints. These feedback constraints are added to the main problem module to guide the next round of main problem solving.

[0012] Furthermore, the specific steps for solving the master problem of DC safety-constrained unit combination in multiple scenarios are as follows: First, initialize the parameters; Secondly, based on active loads in multiple scenarios, CPLEX is used to solve the master problem of unit combination under DC safety constraints in multiple scenarios.

[0013] Furthermore, the specific steps for quickly predicting voltage over-limit situations at each node under multiple scenarios in the initial unit start-up and shutdown scheme using the trained neural network surrogate model are as follows: Read the offline trained neural network agent model; By constructing online inputs for each time period, the decision-making results of unit combinations under DC safety constraints in multiple scenarios are reconstructed. The power flow exceedance status at each time period is calculated based on the trained neural network proxy model.

[0014] Furthermore, the prediction results are verified by determining whether the current unit combination has basically eliminated the undervoltage at all times under the neural network surrogate model based on the undervoltage over-limit results output by the neural network surrogate model.

[0015] Furthermore, if the undervoltage limit violation result shows a limit violation, the maximum limit violation value of each node in all scenarios is updated. If the undervoltage limit violation result shows no limit violation, the physical verification model based on accurate AC power flow is called for final verification. If the physical verification model passes the verification, the final SCUC scheme that satisfies the AC power flow constraints is output. If the physical verification model fails the verification, the limit violation information obtained from the actual physical verification is used to replace the proxy result, the feedback information is updated again, and the problem is returned to continue iterating.

[0016] A second aspect of the present invention provides a model-data dual-driven safety-constrained unit combination decision system, comprising: The physical model module is configured to generate active and reactive load scenarios based on ellipsoidal uncertainty sets. Based on the active and reactive load scenarios, a multi-scenario AC power flow constraint unit combination model is constructed as the main problem of multi-scenario DC safety constraint unit combination. The historical data-driven module is configured to build a neural network proxy model with unit start-up and shutdown status and topology enhancement features as inputs and node voltage overruns as outputs. Through multi-objective training, the neural network proxy model is made to approximate the optimal AC power flow mapping. The two-stage solution module is configured to use a two-stage iterative solution method driven by both analog and digital methods. First, it solves the master problem of unit combination under DC safety constraints in multiple scenarios to obtain the initial unit start-up and shutdown scheme and output plan. Then, it uses a trained neural network surrogate model to quickly predict the voltage over-limit situation of each node in multiple scenarios in the initial unit start-up and shutdown scheme, and verifies the prediction results to obtain the final unit start-up and shutdown scheme.

[0017] A third aspect of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute steps in the model data dual-driven safety constraint unit combination decision method as described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the model data dual-driven safety constraint unit combination decision-making method as described in the first aspect of the present invention.

[0019] A fifth aspect of the present invention provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the model data dual-driven safety constraint unit combination decision-making method as described in the first aspect of the present invention.

[0020] The above one or more technical solutions have the following beneficial effects: This invention discloses a model-data dual-driven method and system for safety-constrained unit combination decision-making. Addressing the shortcomings of existing AC power flow constraint unit combination solutions, such as difficulty in balancing accuracy and efficiency and insufficient adaptability to multiple scenarios, this invention first constructs a multi-scenario active and reactive load model based on ellipsoidal uncertainty sets and a master problem for DC safety-constrained unit combination. Second, it constructs a neural network surrogate model with unit start-up / shutdown states and topology enhancement features as inputs and node voltage exceedances as outputs, training it through multi-objectives to approximate the optimal AC power flow mapping. Finally, it employs an iterative framework combining master problem optimization data with rapid sensitivity verification and feedback correction, replacing high-frequency power flow calculations with a neural network, and finally verifying the results with an accurate physical model after convergence. This invention significantly improves the solution efficiency for multiple scenarios while ensuring that the unit combination scheme strictly meets AC power flow safety and economic requirements, making it suitable for day-ahead dispatch decisions in power systems with a high proportion of renewable energy integration.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the model data dual-driven safety constraint unit combination decision-making method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the neural network model structure in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of a feasible solution recovery SCUC framework in Embodiment 1 of the present invention; Figure 4 This is a comparison diagram of node voltages considering reactive power uncertainty in Embodiment 1 of the present invention; Figure 5 This is a scatter plot of the limit fitting in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of the neural network performance in Embodiment 1 of the present invention; Figure 7 This is a schematic diagram of the unit start-up and shutdown results in Method 1 of Embodiment 1 of the present invention; Figure 8 This is a schematic diagram of the unit start-up and shutdown results in method 2 of embodiment 1 of the present invention; Figure 9 This is a comparison chart of the changes in the objective function values ​​of methods 1 and 2 in Embodiment 1 of the present invention; Figure 10 This is a comparison chart of the cumulative operating hours of methods 1 and 2 in Embodiment 1 of the present invention; Figure 11 This is a comparison chart of the maximum voltage over-limit changes in methods 1 and 2 in Embodiment 1 of the present invention. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] Example 1: Embodiment 1 of the present invention provides a model-data dual-driven safety-constrained unit combination decision-making method, such as... Figure 1 As shown, it includes the following steps: Step 1: Based on the active and reactive load scenarios generated by the ellipsoidal uncertainty set, construct a multi-scenario AC power flow constraint unit combination model as the main problem of multi-scenario DC safety constraint unit combination.

[0027] Step 1.1: Generate active and reactive load scenario sets based on the ellipsoidal uncertainty set, use the covariance matrix to characterize the intensity and correlation of multi-node load fluctuations, and use the ellipsoidal radius to control the overall amplitude of the disturbance to generate a finite number of active and reactive load scenarios.

[0028] In one specific implementation, the ellipsoidal uncertainty set is used to characterize the range of uncertain fluctuations in load correlation. This uncertainty set is constructed based on historical load data, describes the correlation and intensity of load fluctuations at multiple nodes through the covariance matrix, and controls the overall disturbance range through the ellipsoid radius parameter. Its mathematical expression is shown in formula (5). In this embodiment, based on this uncertainty set, a finite number of active and reactive load scenarios are sampled and generated within the set for subsequent multi-scenario safety-constrained unit combination optimization models.

[0029] It should be noted that, in this application, "node" refers to the set of bus nodes in a power system, represented by the symbol... m This indicates that the system includes load nodes, generation nodes, and balancing nodes. In the load uncertainty modeling section, active / reactive load modeling is performed only for nodes with loads. However, in the power flow constraint and unit combination optimization model, the node set covers all bus nodes in the system to ensure the integrity of the model and the correct expression of power balance constraints.

[0030] The scene generation function is: (1).

[0031] In the formula: and Let represent the set of active and reactive scenarios at time t, where and This represents the active and reactive load scenario vector of the node in the k-th scenario. , They represent t time k Scenario, non-zero load node m The active and reactive load power, m =1,2…N D N D This represents the total number of nodes with non-zero load. This represents the scene generation function; and express t The number of time nodes is The example includes the active and reactive baseline load vectors; and They are respectively t Time of the first m One active load and one reactive load baseline value; This represents the covariance matrix, used to describe the fluctuation intensity and correlation of nodal disturbances; This represents the radius of the ellipsoid, used to control the overall magnitude of the disturbance; Indicates the number of scenes.

[0032] The load of a non-zero load node in the k-th scenario is defined as follows: (2), (3).

[0033] In the formula: For a set of nodes with non-zero load, Indicates the first m The reactive load baseline value for each node, corresponding to the number of nodes is: , representing the number of non-zero load nodes participating in uncertainty modeling; Let be the relative disturbance of the non-zero load node in scenario k. This represents the reactive load value of the m-th node at time t.

[0034] covariance matrix Represented as: (4).

[0035] In the formula: Let be the diagonal matrix of nodal fluctuation intensity, representing the relative fluctuation intensity of non-zero load nodes. Indicates the first The relative fluctuation intensity of each non-zero load node. , R represents the total number of nodes with non-zero load; R is the correlation coefficient matrix, whose elements are... Indicates the first The and the first The correlation coefficient between non-zero load node disturbances is used to characterize the spatial correlation of load disturbances at each node.

[0036] The ellipsoidal uncertainty set constraint is: (5).

[0037] In the formula: W is used to describe the intensity and correlation of fluctuations between nodes. Used to control the overall deviation of the scenario from the baseline operating condition.

[0038] Step 1.2: Based on the active and reactive load scenarios, establish a multi-scenario DC safety constraint unit combination master problem model with the objective of minimizing the expected operating cost of multiple scenarios and taking into account unit operation constraints and DC power flow constraints.

[0039] The objective function of the multi-scenario DC-SCUC model is to minimize the total cost. Its main components include the start-up and shutdown costs of the generating units. Weighted operating costs The details are as follows: (6), in (7), (8).

[0040] In the formula: This represents a unit combination model constrained by AC power flow. Indicates the unit i At any moment t The start / stop status; Indicates the unit i At any moment t Scene k Those who have made meritorious contributions; This represents the total number of system units. This represents the total number of system lines. a i , b i , c i These represent the secondary cost coefficient, linear cost coefficient, and fixed operating cost item of the unit, respectively, used to characterize the nonlinear characteristics of the unit's power generation cost as output changes. The units Start-up and shutdown costs; The units At any moment The start and stop signals, among which , . Indicates the unit At any moment The start / stop status.

[0041] The constraints include power balance constraints, DC power flow constraints, upper and lower limits of unit output constraints, minimum start-up and shutdown time constraints, unit spinning reserve constraints, and unit ramping constraints.

[0042] (1) Power balance constraint.

[0043] (9).

[0044] (2) DC power flow constraint (10) (11), (12).

[0045] In the formula: , Let represent the active power and reactive power injected by node m in scenario k at time t, respectively. and These represent the power on the line from node m to node n in scenario k at time t; The mutual susceptance between node m and node n; and Let represent the voltage phase angles at nodes m and n in scenario k at time t, respectively; Let be a 0-1 integer variable, representing whether, at time t in scenario k, unit i is at node m and in the active state. For nodes without connected generators, .

[0046] (3) Upper and lower limits of unit output constraints.

[0047] (13).

[0048] In the formula: and These represent the lower and upper limits of the active power output of unit i, respectively.

[0049] (4) Minimum start-up and shutdown time constraints for the unit.

[0050] (14) (15).

[0051] In the formula: and They represent the generating units. At any moment The continuous start and stop time; and They represent the generating units. The minimum continuous start and stop time.

[0052] (5) Unit rotation standby constraint.

[0053] (16).

[0054] In the formula: and These are the positive and negative spinning reserve requirements of the system caused by load forecasting errors, respectively. and They represent the generating units. Uphill and downhill constraints.

[0055] (6) Unit ramping constraints.

[0056] (17).

[0057] The multi-scenario PFR model will calculate the unit start-up and shutdown based on DC-SCUC. Contributing to various scenarios As input, the reactive power output of the units in each scenario As the decision variable, the voltage over-limit value of each scenario is minimized by AC-OPF with voltage soft constraint penalty, and finally the voltage and phase angle of each node are output.

[0058] This invention achieves multi-scenario SCUC decision-making that satisfies precise power flow constraints by using a framework that decouples active and reactive power. Based on the multi-scenario DC-SCUC master problem and the multi-scenario reactive load synchronous verification, it overcomes the shortcomings of traditional methods that ignore reactive power uncertainty.

[0059] It should be noted that, to evaluate the predictive performance of the neural network surrogate model, this embodiment uses the coefficient of determination (R²) and root mean square error (RMSE) as evaluation metrics. R² measures the goodness of fit of the model to the real data, representing the model's ability to explain changes in the output variable; a value closer to 1 indicates a better fit. RMSE measures the degree of deviation between the predicted and actual values.

[0060] Step 2: Construct a neural network proxy model with unit start-up and shutdown status and topology enhancement features as inputs and node voltage overruns as outputs. Through multi-objective training, make the neural network proxy model approximate the optimal AC power flow mapping.

[0061] Step 2.1: Construct a neural network surrogate model with unit start-up and shutdown status and topology enhancement features as inputs and node voltage over-limit as output.

[0062] In one specific implementation, from an input and output perspective, the data-driven PFR model can be viewed as a nonlinear function mapping unit combination states to node undervoltage exceedance vectors. This is considering the relationship between the power flow exceedance degree and the load factor. Since the node where the unit is located is relevant, the neural network is set as follows: (18).

[0063] In the formula: Let represent the voltage over-limit value of each node in the s-th sample group at time t; This represents the start-up / shutdown status of the unit in the s-th sample group at time t; This represents the topology enhancement features of sample s oriented towards the grid structure at time t. These features include information such as the coordination characteristics of online unit pairs, the coverage capability of units to the system load center, the spatial dispersion of online units, and the total number of online units. This represents the neural network function at time t.

[0064] like Figure 2 As shown, the nonlinear mapping defined in equation (18) is implemented by the input layer, several hidden layers, and the output layer. The input layer vector is defined as: (19).

[0065] in, Represents the input layer vector. Let represent the neural network input of the s-th sample at time t.

[0066] Its dimension is determined by both the basic input features and the topology enhancement features; the output layer dimension is equal to the number of system nodes, corresponding to the voltage over-limit prediction value of each node. The topology enhancement features include the start-up and shutdown status of units in each time period, the system load level, and the topology enhancement features composed of the coordination features of online units, load coverage features, spatial dispersion features, and the number of online units.

[0067] The start-up and shutdown status of generator units at different time periods indicates the operating status (start-up / shutdown) of each generator unit at different times, usually represented by 0-1 variables. This feature reflects the dynamic changes in the system's dispatchable resources and is an important basic information for characterizing the system's power supply capacity.

[0068] System load level represents the total load demand of the system at different times, reflecting the overall operating pressure of the system. This characteristic reflects the changes in load intensity of the power grid and is an important macroscopic indicator of the system's operating status.

[0069] Online unit coordination characteristics represent the coordination relationships between units currently in operation, such as output distribution, geographical or electrical coupling relationships, and are used to characterize the coordinated power supply structure characteristics during unit operation. Load coverage characteristics represent the power supply coverage capability of online units to the system load, that is, the degree to which the currently operating units meet the load of each node, and are used to measure power supply adequacy and local power supply capacity.

[0070] Spatial dispersion characteristics represent the degree of distribution of online generating units in the power grid topology. They are used to characterize the spatial concentration or dispersion of generating units, thereby reflecting the balance of the power supply structure.

[0071] The number of online generating units represents the total number of generating units in operation during the current time period. It is used to reflect the scale and level of system scheduling resources and is a direct indicator of the system's operating status.

[0072] Step 2.2: The neural network proxy model adopts a multi-layer feedforward neural network structure. Through offline training, the neural network approximates the nonlinear mapping relationship of AC optimal power flow calculation.

[0073] In one specific implementation, the training dataset of this embodiment is derived from historical power system operation data and publicly available datasets. The neural network surrogate model adopts a time-sharing training strategy, that is, independent neural network models are trained for each of the 24 time periods of the day, with each time period's network trained only for the load level and operating characteristics of that time period. The neural network surrogate model adopts a multi-layer feedforward structure, including an input layer, several hidden layers, and an output layer. If the first... The hidden layer output is denoted as Then the forward propagation process can be represented as: (20).

[0074] In the formula: and These are the weight matrix and bias vector of the lq-th layer, respectively. It is a non-linear activation function. Indicates the first Hidden layer output.

[0075] After After the hidden layer, the network output can be written as: (twenty one).

[0076] In the formula: This represents the network's prediction result for the voltage exceeding the limit of the s-th sample node at time t; The number of hidden layers; This is the output mapping function. Since the node voltage limit itself is non-negative, this invention sets the output target to a non-negative undervoltage amplitude.

[0077] Step 2.3: During the training of the neural network agent model, a multi-objective loss function is used, including Huber loss, quantile loss and binary cross-entropy loss.

[0078] In one specific implementation, the neural network is trained using a multi-objective loss function, including: Huber loss for regression fitting, quantile loss for imposing a stronger penalty on undervoltage underestimation, and binary cross-entropy loss for enhancing the ability to identify node over-limit states. The multi-loss objective function is... To improve the model's ability to fit continuous undervoltage amplitudes while meeting engineering requirements for safety assessment, this invention employs a multi-objective training strategy. The comprehensive objective function... for: (twenty two).

[0079] In the formula: Huber loss is used to improve the accuracy of regression fitting. Quantile loss is used to impose a stronger penalty on underestimation of undervoltage, making the model as a whole tend to make conservative predictions. The binary cross-entropy loss is constructed based on whether the limit is exceeded, and is used to enhance the model's ability to identify the node's limit-exceeding state; , and These are the weighting coefficients for the three parts of the loss.

[0080] This invention uses a data-driven neural network model to replace the high-frequency AC optimal power flow calculation in the iterative process, which significantly improves the computational efficiency of the power flow verification stage. At the same time, the safety and reliability of the final result are guaranteed by a multi-objective training mechanism oriented towards feasible solution recovery and a final physical model verification as a fallback.

[0081] Step 3: Solve the problem using a modular dual-drive two-stage iterative solution method.

[0082] In one specific implementation, the master problem of unit combination under DC safety constraints in multiple scenarios is first solved to obtain the initial unit start-up and shutdown scheme and output plan. Then, a trained neural network surrogate model is used to quickly predict the voltage over-limit situation of each node in multiple scenarios in the initial unit start-up and shutdown scheme, and the prediction results are verified to obtain the final unit start-up and shutdown scheme.

[0083] The modular dual-drive two-stage iterative solution method adopts the following approach: Figure 3The solution framework shown includes: a physical model-driven main problem module based on multi-scenario DC safety constraints for unit combinations; a data-driven sub-problem module based on neural networks for multi-scenario AC power flow recovery; a sensitivity calculation module; a feedback constraint generation module; and a physical verification module based on accurate AC power flow. The main problem module and sub-problem modules iterate in a closed loop using unit start-up / shutdown schemes and voltage over-limit feedback information. The main problem module outputs unit start-up / shutdown schemes to the sub-problem modules; the sub-problem modules quickly predict voltage over-limits using a neural network surrogate model. If no over-limit occurs, the process is transferred to the physical verification module; if an over-limit occurs, over-limit information is generated and sent to the sensitivity calculation module. The sensitivity calculation module calculates the sensitivity matrix of node voltage to unit reactive power output, and the feedback constraint generation module constructs linear feedback constraints. These feedback constraints are added to the main problem module to guide the next round of main problem solving.

[0084] Specifically, the following steps are included: Step 3.1: Solve the master problem of unit combination under DC safety constraints in multiple scenarios to obtain the initial unit start-up and shutdown scheme and output plan.

[0085] Step 3.1.1: First, initialize the parameters.

[0086] Specifically, the example information is imported, and the node voltage over-limit information, unit status information, and sensitivity feedback information are initialized to provide initial parameters for the first round of solving the main problem.

[0087] Step 3.1.2: Next, based on the active load of multiple scenarios, the CPLEX solver is used to directly solve the main problem of DC-SCUC (DC-SCUC) combination of multiple scenarios through algorithms such as branch and bound.

[0088] Specifically, the main program calls the SCUC subproblem solver for DC power flow constraints to obtain the current turbine start-up and shutdown scheme and corresponding output plan based on the active power load values ​​in multiple scenarios and the existing turbine start-up and shutdown feedback constraints.

[0089] Step 3.2: Input the unit start-up and shutdown scheme into the trained neural network proxy model to quickly predict voltage over-limit situations at each node under multiple scenarios.

[0090] Step 3.2.1: Read the offline trained neural network proxy model.

[0091] Specifically, after obtaining the current turbine unit combination, the program loads 24 time-sharing neural network proxy models formed by offline training, input and output standardized parameters, and metadata required for topological feature recovery.

[0092] Step 3.2.2: By constructing online inputs for each time period, the decision results of the combined DC safety constraint units in multiple scenarios are reconstructed.

[0093] Specifically, the basic inputs include the system load level and the start-up and shutdown status of the units during the period. If the offline training adopts the topology enhancement mode, the online phase synchronously recovers the cooperative online characteristics, load coverage characteristics, spatial dispersion characteristics, and online unit quantity characteristics of the units, and finally forms an input vector that is completely isomorphic to the offline training.

[0094] Step 3.2.3: Calculate the power flow exceedance status for each time period based on the trained neural network proxy model.

[0095] Specifically, the program completes input standardization, neural network feedforward inference, and output destandardization one by one according to time period to obtain the full-time node voltage over-limit prediction results under the current turbine unit combination, and performs threshold filtering on small numerical disturbances near zero.

[0096] Step 3.3: Perform data-driven power flow feasibility verification based on the voltage over-limit situation of each node.

[0097] Specifically, based on the undervoltage and limit exceedance results output by the neural network surrogate model, it is determined whether the current unit combination has essentially eliminated undervoltage at all nodes throughout the entire time period under neural network verification. If the undervoltage and limit exceedance results show exceedances, the maximum exceedance value of each node in all scenarios is updated. If the undervoltage and limit exceedance results show no exceedances, a physical verification model based on precise AC power flow is invoked for final verification. If the physical verification model passes the verification, the final SCUC scheme that satisfies the AC power flow constraints is output; if the physical verification model fails the verification, the exceedance information obtained from the actual physical verification is used to replace the surrogate result, the feedback information is updated again, and the process returns to the main problem to continue iteration. The above iteration is repeated until the main problem outputs a unit combination scheme that satisfies the precise AC power flow constraints.

[0098] Specifically, the node voltage prediction value output by the neural network surrogate model is compared with the preset upper and lower voltage limits to determine whether there is a limit violation.

[0099] Step 3.3.1: If the undervoltage limit result is not exceeded, then call the physical verification model based on precise AC power flow for final verification.

[0100] Specifically, if the proxy verification passes, the process proceeds to the physical precision verification stage. That is, in this embodiment, if the neural network determines that the unit combination scheme in the current loop is basically feasible, the program calls the complete physical verification module and uses the physical verification model of precise AC power flow to perform voltage safety verification of the final scheme based on precise AC power flow.

[0101] The physical verification model aims to minimize the voltage over-limit penalty value, and the objective function is... for: (twenty three).

[0102] In the formula: Indicates that unit i is at time... The unproductive output of scenario k and Indicates that node d at time t The node voltage magnitude and phase angle of scenario k; and Relaxation variables representing undervoltage and overvoltage; and This represents the voltage over-limit penalty coefficient. Constraints include: soft constraints on upper and lower voltage limits, active power constraints of generating units, upper and lower limits on reactive power output of generating units, active / reactive power balance at nodes, AC power flow equation constraints, and branch thermal stability constraints.

[0103] Step 3.3.2: If the undervoltage limit exceeds the limit, update the maximum limit value of each node in all scenarios.

[0104] Specifically, when the proxy verification fails, the program extracts and saves the maximum limit exceedance information for each node and time period in all scenarios within the current loop. It then updates the unit status variables. The program saves the unit start / stop status and reactive power output information obtained in the current round, making it a reference point for the next round of feedback correction.

[0105] If voltage limits are exceeded, feedback constraints are generated and added to the main problem based on the voltage limit exceedance information and the sensitivity matrix of node voltage to unit reactive power output. Constraints are added to the main problem based on the limit exceedance values. The main program calls the sensitivity calculation module to recalculate the sensitivity matrix of node voltage to unit regulation behavior and resolves the main problem.

[0106] The sensitivity matrix is ​​obtained by recalculating the AC power flow after applying a small reactive power disturbance. It is used to quantify the impact of changes in reactive power output of each unit on the voltage of each node, and to construct linear feedback constraints accordingly.

[0107] Sensitivity matrix for: (twenty four).

[0108] In the formula: The sensitivity of the voltage at node m to the reactive power output of generator unit i at time t is expressed by the following formula: (25).

[0109] In the formula: This represents the voltage at node m after a reactive disturbance is applied to unit i at time t. This represents the voltage at node m of unit i before the disturbance at time t; This represents the amplitude of reactive power disturbance applied by unit i. The calculation formula is: .

[0110] in, It is a function in MATPOWER used to solve AC power flow, and is used to calculate the system voltage and power flow distribution under given unit output and load conditions; This represents the system data in MATPOWER format after applying reactive power disturbances to the generator unit.

[0111] Based on the current round's over-limit results and the sensitivity matrix, generate new linear feedback correction constraints: (26).

[0112] In the formula: Let be the set of over-limit voltages of each node at time t; To maximize the reactive power output of each unit, Represents the set of real numbers; and Let t be the unit start-up and shutdown values ​​during the current and previous iterations. The calculated power flow over-limit matrix, This is the sensitivity matrix for reactive power and voltage of the generator unit.

[0113] These updated feedback messages are sent back to the SCUC master problem to guide the next round of unit combination adjustments. After completing the update of the over-limit information and sensitivity feedback, the outer iteration counter is incremented, and the main program calls the DC-SCUC model again to enter the next round of master problem solving.

[0114] When both the neural network and physical model verification results indicate that the unit combination decision scheme meets the voltage safety requirements based on the AC power flow equation, the iteration terminates, and the program outputs the final unit start-up, shutdown, and output scheme that satisfies the AC power flow equation.

[0115] This invention, through a closed-loop iterative structure of "main problem optimization - data-driven verification - sensitivity feedback correction", can converge to a feasible solution that satisfies the power flow constraints of multiple scenarios in a relatively small number of iterations. It has both high precision and high efficiency and has good prospects for engineering applications.

[0116] To verify the accuracy and effectiveness of the proposed method, this invention uses the IEEE 30-bus system as a test case. The system comprises 30 nodes, 41 branches, and 6 generator units. To simulate uncertainty, five active and reactive load scenarios are generated based on the ellipsoidal uncertainty set, with the load fluctuation coefficient set to 0.05. The neural network employs three hidden layers, each with 64 neurons, using ReLU activation, and the output layer uses linear activation. The training data consists of 5000 randomly generated generator unit combination schemes and their corresponding precise AC optimal power flow calculation results.

[0117] Process simulation (1) Consider the effectiveness verification of reactive power in multiple scenarios.

[0118] To verify the effectiveness of considering multiple reactive power scenarios, the node voltage results of different methods under the same time period and the same specific reactive power scenario are as follows: Figure 4 As shown. Method 1's AC-SCUC uses a PFR oriented towards reactive power expectation, while Method 2's AC-SCUC uses a PFR oriented towards multiple scenarios. For example... Figure 4 As shown, the overall shape of the obtained node voltage distribution is relatively consistent, which indicates that the start-stop decision after considering multiple reactive power scenarios will not significantly disrupt the original power flow distribution characteristics.

[0119] Further observation of the magnified area reveals that Method 1 has reached a critical undervoltage point in this scenario. The node voltage marked with red dots in the figure falls near the lower voltage limit, reflecting that although the start-up and shutdown scheme formed under the expected reactive power in a single scenario has a certain adaptability to the expected scenario, its reactive power support margin is insufficient when the system enters a specific disturbance scenario, and the voltage safety boundary is relatively fragile. In contrast, Method 2 forms a unit start-up and shutdown scheme based on five reactive power scenarios, and then introduces this scheme into the current scenario for verification. It can be seen that the voltage of the weak node that was originally near the lower limit is raised, the critical voltage over-limit phenomenon is eliminated, and the voltage of most other nodes changes very little. This indicates that multi-scenario decision-making does not simply increase the output of all units, but rather preserves more sufficient reactive power regulation capacity for the system through a more reasonable online combination of units, so that weak nodes can still be maintained within the safe range under scenario disturbances.

[0120] In summary, the unit combination decision considering multiple reactive power scenarios is essentially reserving a voltage safety margin in advance for uncertain reactive power distribution during the day-ahead scheduling phase, thereby making the unit start-up and shutdown schemes more robust and feasible when implemented in actual scenarios.

[0121] (2) Verification of the effectiveness of the neural network.

[0122] In the modular dual-drive solution framework of this invention, the data-driven PFR model mainly undertakes the task of rapid voltage verification in the outer unit combination iteration. Its function is not to completely replace the physical model, but to quickly identify potential undervoltage risks at the high-frequency call level and provide prior judgments for subsequent main problem correction and end-point physical verification. Therefore, the evaluation of the effectiveness of this model should not only focus on the regression error in a general sense, but should focus on whether it meets the following three engineering requirements: first, it can accurately approximate the node undervoltage distribution given by the physical model; second, it can avoid underestimation and underestimation of the actual undervoltage as much as possible; and third, it allows for moderate conservatism when necessary to obtain higher stability and safety of the outer iteration.

[0123] Overall, the constructed time-sharing neural network model can approximate the physical PFR output quite well. Simulation results show that the model's average RMSE on the IEEE30 system is only 6×10⁻⁶ pu, and the average R0 is... 2 A score of 0.99 indicates that its numerical fitting of the node undervoltage distribution under the most unfavorable scenario has achieved high accuracy. Combined with... Figure 5 The scatter plot further shows that the predicted total over-limit and the actual total over-limit are distributed almost along the diagonal, indicating that the model accurately portrays the overall severity of the sample and there is no systematic distortion.

[0124] However, for the unit combination outer-layer iterative problem studied in this invention, simply having a high R... 2 Low RMSE is not the only effectiveness criterion. Since the surrogate model directly participates in determining the direction of the main problem's correction, an overly optimistic prediction of the actual undervoltage could lead to the failure to promptly add units with reactive power support capabilities, thus weakening the feasible region recovery effect. Therefore, more critical indicators are the ability to recall units exceeding limits and the underestimation of control capabilities, with values ​​such as... Figure 6 As shown in the figure, this invention defines nodes with voltage amplitudes below 0.98 pu as positive samples, and the recall rate is the proportion of correctly identified over-limit nodes to the total number of actual over-limit nodes. Overall, the model's recall rate reaches 0.9271, indicating that it can identify approximately 92.71% of the true undervoltage points; simultaneously, the underestimation rate is only 0.0500, indicating that the model's missed detection rate and underestimation of true undervoltage are both small. In other words, in most cases, data-driven PFR can promptly expose potential voltage-weak nodes without misleading outer unit combination adjustments due to a significant underestimation of the undervoltage level.

[0125] comprehensive Figure 5 and Figure 6 The results suggest that the data-driven PFR model constructed in this invention has largely met the three core requirements of the modular dual-drive solution for the surrogate model: First, it can accurately approximate the physical model output at the numerical level and has high undervoltage amplitude regression accuracy; Second, it maintains a high out-of-limit recall rate and a low underestimation probability in key safety indicators, and can provide a reliable direction for feedback correction of the main problem.

[0126] To verify the effectiveness of the proposed modular dual-drive solution method, the present invention sets up the following two methods to conduct comparative simulations of this example: Method 1: Consider the AC-SCUC for multiple scenarios of recovering feasible solutions through communication.

[0127] Method 2: The multi-scenario AC-SCUC decision method driven by both analog and digital methods, which is the method of this invention.

[0128] Method 1 introduces multi-scenario PFR on the basis of the traditional multi-scenario DC-SCUC model and implements multi-scenario AC-SCUC based on the solution framework in Chapter 2. Method 2 is a modular dual-drive multi-scenario AC-SCUC based on Method 1, which considers data-driven replacement of multi-scenario PFR calculation. To analyze the effectiveness of the proposed dual-drive solution method in terms of AC power flow feasibility, solution efficiency, and decision result quality, comparisons are made in Table 1. Figure 7 and Figure 8 The calculation results of the two methods shown.

[0129] Table 1. Comparison of solution results between Method 1 and Method 2

[0130] From Table 1, Figure 7 and Figure 8 It can be seen that Method 1 and Method 2 maintain consistency in the quality of the final solution. The feasible solution cost obtained by both methods is 122.9904 million yuan, the unit start-up and shutdown schemes are exactly the same, and neither method results in voltage over-limit, with both calculations showing power flow feasibility. This indicates that replacing the physical PFR in the iterative process with data-driven PFR does not weaken the AC feasibility of the final solution, nor does it reduce the economic efficiency of the unit combination result.

[0131] From a computational efficiency perspective, Method 2's total computation time is 145.103 s, lower than Method 1's 176.872 s, a reduction of 31.769 s. Combined with the iteration data in Table 1, it can be seen that both methods have 3 iterations of the main function and 5 verification scenarios per iteration, indicating that the efficiency improvement of the method in this invention does not stem from a reduction in the number of iterations, but mainly from a decrease in the computational cost of the AC power flow recovery stage in a single iteration. Further examining the breakdown of time components, the two methods have similar times for the initial solution and sensitivity calculation stages, with the main difference concentrated in the AC power flow recovery stage. Method 1's AC power flow recovery time is 49.379 s, while Method 2's is only 0.371 s. Although Method 2 adds an extra 16.510 s of physical model verification time, this overhead is far less than the time saved in the AC power flow recovery stage, thus still demonstrating higher overall solution efficiency.

[0132] The above results indicate that the advantage of the method of this invention lies not in reducing the number of iterations, but in reducing the computational efficiency of each iteration. In other words, under the condition that the iteration framework, verification scale, and termination criteria are basically the same, data-driven PFR can significantly compress the computational burden of the AC power flow recovery step, thereby achieving an overall speedup. However, the high solution efficiency of Method 2 is conditional; that is, its number of iterations must be kept at a similar level to Method 1. Although data-driven PFR has the advantage of fast solution speed, the approximation error it introduces cannot be completely eliminated. Only when this error remains controllable during the iteration process, without causing additional significant rework or an increase in the number of iterations, can the time gain brought by data-driven substitution truly be transformed into an overall solution efficiency advantage.

[0133] To further analyze the effectiveness of Method 2 in maintaining the number of iterations, Figure 9 , Figure 10 and Figure 11 The iterative changes in the objective function value, the total number of operating units, and the maximum voltage deviation are presented. Figure 9 and Figure 10 It can be seen that the two methods show basically the same trend in terms of the objective function value and the total number of operating units, and both converge to the same feasible solution after the third iteration. This indicates that the data-driven PFR's judgment on the direction of unit start-up and shutdown correction in the intermediate iteration stage is generally consistent with the physical model-driven method. Figure 11 It can be seen that the estimation of the maximum voltage deviation by Method 2 in the intermediate iteration stage is somewhat different from that of Method 1, especially the deviation is slightly larger in the second iteration. However, the maximum voltage deviation of both methods gradually decreases with iteration and eventually converges to 0 pu, indicating that the error does not destroy the overall direction of the iterative correction, nor does it affect the acquisition of the final feasible solution of AC power flow.

[0134] The fundamental reason lies in the fact that, within the model-driven and data-driven dual-drive framework, the role of the surrogate model is not to precisely reconstruct the voltage of all nodes point by point, but rather to quickly identify whether power flow exceeds limits and provide a reasonable estimate of the degree of exceedance. For unit combination iterative correction, the more crucial aspect is identifying which scenarios, time periods, and nodes pose voltage safety risks and driving corrections accordingly, rather than completing a full physical power flow recovery at a high cost in each iteration. Therefore, even if the voltage estimation of data-driven PFR has some approximation error, as long as its judgment on the direction of limit identification and correction remains correct, it can ensure the stable progress of the iterative process and return to a feasible solution consistent with the physical model-driven method in the final physical model verification stage.

[0135] In summary, the dual-drive solution method proposed in this invention, while ensuring consistency between the final AC power flow feasibility, unit combination results, and operational economy and the physical model-driven method, replaces the high-cost physical PFR calculations in the iterative process with data-driven PFR, significantly reducing the computational overhead of the AC power flow recovery stage and improving the overall solution efficiency. Simultaneously, with the help of the final physical model verification mechanism, this method can ensure the correct direction of unit combination correction, the stability of the iterative process, and the consistency between the final result and the physical model solution result, thus verifying its effectiveness in solving SCUC in multiple scenarios.

[0136] Example 2: Embodiment 2 of the present invention provides a model-data dual-driven safety constraint unit combination decision system, comprising: The physical model module is configured to generate active and reactive load scenarios based on ellipsoidal uncertainty sets. Based on the active and reactive load scenarios, a multi-scenario AC power flow constraint unit combination model is constructed as the main problem of multi-scenario DC safety constraint unit combination. The historical data-driven module is configured to build a neural network proxy model with unit start-up and shutdown status and topology enhancement features as inputs and node voltage overruns as outputs. Through multi-objective training, the neural network proxy model is made to approximate the optimal AC power flow mapping. The two-stage solution module is configured to use a two-stage iterative solution method driven by both analog and digital methods. First, it solves the master problem of unit combination under DC safety constraints in multiple scenarios to obtain the initial unit start-up and shutdown scheme and output plan. Then, it uses a trained neural network surrogate model to quickly predict the voltage over-limit situation of each node in multiple scenarios in the initial unit start-up and shutdown scheme, and verifies the prediction results to obtain the final unit start-up and shutdown scheme.

[0137] Example 3: Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed the steps in the model data dual-driven safety constraint unit combination decision method as described in Embodiment 1 of the present invention.

[0138] Example 4: Embodiment 4 of the present invention provides a computer device, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps in the model data dual-driven safety constraint unit combination decision-making method as described in Embodiment 1 of the present invention.

[0139] Example 5: Embodiment 5 of the present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the model data dual-driven safety constraint unit combination decision-making method as described in Embodiment 1 of the present invention.

[0140] The steps and methods involved in Examples 2, 3, 4 and 5 above correspond to those in Example 1. For specific implementation methods, please refer to the relevant description section of Example 1.

[0141] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0142] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0143] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A model-data-driven, dual-driven method for safety-constrained unit combination decision-making, characterized in that, Includes the following steps: Based on the active and reactive load scenarios generated by the ellipsoidal uncertainty set, a multi-scenario AC power flow constraint unit combination model is constructed according to the active and reactive load scenarios as the main problem of multi-scenario DC safety constraint unit combination. A neural network surrogate model is constructed with unit start-up and shutdown status and topology enhancement features as inputs and node voltage over-limit as output. Through multi-objective training, the neural network surrogate model is made to approximate the optimal AC power flow mapping. The solution is obtained by a two-stage iterative method with analog and digital dual drive. First, the main problem of unit combination under DC safety constraints in multiple scenarios is solved to obtain the initial unit start-up and shutdown scheme and output plan. Then, a trained neural network surrogate model is used to quickly predict the voltage over-limit situation of each node in multiple scenarios in the initial unit start-up and shutdown scheme, and the prediction results are verified to obtain the final unit start-up and shutdown scheme.

2. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, Active and reactive load scenario sets are generated based on the ellipsoidal uncertainty set. The covariance matrix is ​​used to characterize the intensity and correlation of load fluctuations at multiple nodes. The overall amplitude of the disturbance is controlled by the ellipsoidal radius to generate a finite number of active and reactive load scenarios.

3. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The main problem of multi-scenario DC safety constraint unit combination aims to minimize the expected operating cost in multiple scenarios. The constraints include power balance constraints, DC power flow constraints, upper and lower limits of unit output constraints, minimum start-up and shutdown time constraints, unit spinning reserve constraints, and unit ramping constraints.

4. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, Topology enhancement features include the start-up and shutdown status of units in each time period, system load level, and topology enhancement features consisting of online unit pair coordination features, load coverage features, spatial dispersion features, and the number of online units.

5. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The specific steps for training a neural network surrogate model to approximate the optimal power flow mapping through multi-objective training are as follows: The neural network proxy model adopts a multi-layer feedforward neural network structure, and through offline training, the neural network approximates the nonlinear mapping relationship of AC optimal power flow calculation. The neural network agent model training process employs multi-objective loss functions, including Huber loss, quantile loss, and binary cross-entropy loss.

6. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The neural network proxy model adopts a time-sharing training strategy, that is, it trains independent neural network models for each of the 24 time periods of the day, and the network for each time period is trained only for the load level and operating characteristics of that time period.

7. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The analog-digital dual-drive two-stage iterative solution method includes a physical model-driven main problem module based on multi-scenario DC safety constraint unit combination, a data-driven sub-problem module based on neural network multi-scenario AC power flow recovery, a sensitivity calculation module, a feedback constraint generation module, and a physical verification module based on accurate AC power flow. The main problem module and sub-problem modules iterate in a closed loop using unit start-up / shutdown schemes and voltage over-limit feedback information. The main problem module outputs the unit start-up / shutdown scheme to the sub-problem module. The sub-problem module quickly predicts voltage over-limits using a neural network surrogate model; if no over-limit occurs, it proceeds to the physical verification module; if an over-limit occurs, it generates over-limit information and sends it to the sensitivity calculation module. The sensitivity calculation module calculates the sensitivity matrix of node voltage to unit reactive power output, and the feedback constraint generation module constructs linear feedback constraints. These feedback constraints are added to the main problem module to guide the next round of main problem solving.

8. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The specific steps for solving the master problem of DC safety-constrained unit combination in multiple scenarios are as follows: First, initialize the parameters; Secondly, based on active loads in multiple scenarios, CPLEX is used to solve the master problem of unit combination under DC safety constraints in multiple scenarios.

9. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The specific steps for quickly predicting voltage over-limit situations at various nodes under multiple scenarios in the initial unit start-up and shutdown scheme using a trained neural network surrogate model are as follows: Read the offline trained neural network agent model; By constructing online inputs for each time period, the decision-making results of unit combinations under DC safety constraints in multiple scenarios are reconstructed. The power flow exceedance status at each time period is calculated based on the trained neural network proxy model.

10. The model data dual-driven safety constraint unit combination decision-making method as described in claim 1, characterized in that, The prediction results are verified by the undervoltage limit results output by the neural network proxy model to determine whether the current unit combination has basically eliminated the undervoltage at all times under the neural network verification.

11. The model data dual-driven safety constraint unit combination decision-making method as described in claim 10, characterized in that, If the undervoltage limit violation result shows a limit violation, update the maximum limit violation value of each node in all scenarios. If the undervoltage limit violation result shows no limit violation, call the physical verification model based on accurate AC power flow for final verification. If the physical verification model passes the verification, output the final SCUC scheme that satisfies the AC power flow constraints. If the physical verification model fails the verification, replace the proxy result with the limit violation information obtained from the actual physical verification, update the feedback information again, and return to the main problem to continue iterating.

12. A model-data dual-driven safety-constrained unit combination decision-making system, characterized in that, include: The physical model module is configured to generate active and reactive load scenarios based on ellipsoidal uncertainty sets. Based on the active and reactive load scenarios, a multi-scenario AC power flow constraint unit combination model is constructed as the main problem of multi-scenario DC safety constraint unit combination. The historical data-driven module is configured to build a neural network proxy model with unit start-up and shutdown status and topology enhancement features as inputs and node voltage overruns as outputs. Through multi-objective training, the neural network proxy model is made to approximate the optimal AC power flow mapping. The two-stage solution module is configured to use a two-stage iterative solution method driven by both analog and digital methods. First, it solves the master problem of unit combination under DC safety constraints in multiple scenarios to obtain the initial unit start-up and shutdown scheme and output plan. Then, it uses a trained neural network surrogate model to quickly predict the voltage over-limit situation of each node in multiple scenarios in the initial unit start-up and shutdown scheme, and verifies the prediction results to obtain the final unit start-up and shutdown scheme.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the model data dual-driven safety constraint unit combination decision-making method as described in any one of claims 1-11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed by the model data dual-driven safety constraint unit combination decision method as described in any one of claims 1-11.

15. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the model data dual-driven safety constraint unit combination decision method as described in any one of claims 1-11.

Images