A network-structured energy storage frequency and system strength dual security constraint planning method

Abstract

The application discloses a network-constructing type energy storage frequency and system strength double safety constraint planning method, relates to the technical field of power system energy storage planning, and comprises the following steps: introducing an asymmetric penalty term based on a safety constraint sample set, obtaining a frequency-strength constraint proxy model through dynamic pruning growth, converting a model constraint form by using the frequency-strength constraint proxy model by adopting a Big-M method, constructing a GFM-ESS locating and sizing model, and obtaining a network-constructing type energy storage frequency and system strength double safety constraint planning result according to the GFM-ESS locating and sizing model combined with a system locating strategy. The application realizes the collaborative optimization of locating and sizing and operation strategy, accurately evaluates the marginal contribution of a candidate node to the system strength and eliminates redundant nodes by designing a collaborative locating strategy integrating node strength gain and redundant node set removal, effectively compresses the planning scale, and significantly improves the solving efficiency while guaranteeing the supporting effect.

Description

Technical Field

[0001] This invention relates to the field of power system energy storage planning technology, and in particular to a planning method for grid-type energy storage with dual security constraints of frequency and system strength. Background Technology

[0002] Against the backdrop of the low-carbon transformation of the power system, the penetration rate of wind and solar power generation continues to rise. By the end of 2025, my country's cumulative installed capacity of wind and solar power generation had reached 1.84 billion kilowatts, accounting for approximately 47% of the country's total installed power generation capacity. However, renewable energy systems (RES) connected to the grid via power electronic converters lack the inherent mechanical rotor inertia and voltage / frequency support capabilities of synchronous generators (SG), resulting in a grid exhibiting low inertia and weak support characteristics, making it difficult to withstand severe disturbances. Typical examples include the 2019 UK blackout and the 2025 Iberian Peninsula blackout.

[0003] New energy storage technologies, such as grid-forming energy storage systems (GFM-ESS), can effectively simulate the rotor motion equations and damping characteristics of SG (Power Storage System) under control strategies such as Virtual Synchronous Machines (VSMs), actively providing inertia, frequency, and voltage support strength for the system. Therefore, to meet the operational requirements of high-proportion power electronic systems, a dual-safety constraint planning method for grid-forming energy storage, encompassing both frequency and system strength, is urgently needed to address the shortcomings of existing technologies. Summary of the Invention

[0004] The purpose of this invention is to propose a planning method for grid-type energy storage with dual security constraints of frequency and system strength, so as to solve the problem that existing energy storage planning technologies do not have the ability to simultaneously consider frequency and strength support.

[0005] To achieve the above objectives, this invention provides a planning method for grid-type energy storage with dual security constraints of frequency and system strength, comprising:

[0006] S1. Based on the safety constraint sample set, an asymmetric penalty term is introduced, and a frequency-intensity constraint surrogate model is obtained through dynamic pruning and growth.

[0007] S2. Using the frequency-intensity constraint proxy model, the Big-M method is used to transform the model constraint form and construct the GFM-ESS addressing and dimensionality model.

[0008] S3. Based on the GFM-ESS location and capacity model and the system location strategy, obtain the planning results of dual security constraints of network-type energy storage frequency and system strength.

[0009] Optionally, S1, based on the safety constraint sample set, introduces an asymmetric penalty term and obtains a frequency-intensity constraint surrogate model through dynamic pruning growth, including:

[0010] Collect power grid frequency and system strength operation data to obtain a safety constraint sample set;

[0011] A network loss function is established using the aforementioned security constraint sample set, and a composite loss function is obtained by combining it with an asymmetric penalty term.

[0012] Based on the composite loss function, dynamic pruning and growth are performed using a weight mask to obtain a frequency-intensity constrained surrogate model.

[0013] Optionally, the formula for calculating the composite loss function is:

[0014]

[0015] in, Let N be the total loss function, N be the number of training samples, and i be the i-th training sample. It is a hyperbolic cosine function. It is a logarithmic function. For the predicted value, y i For the true value, Here, is the conservative penalty coefficient, RegressionLoss is the regression loss, and Safety Penalty is the safety penalty term.

[0016] Optionally, based on the composite loss function, dynamic pruning and growth are performed using a weight mask to obtain a frequency-intensity constrained surrogate model, including:

[0017] Based on the composite loss function and weight mask, the active connection set and baseline gradient are determined;

[0018] The weight magnitudes are read from the active connection set, and the least weighted connections are filtered for weight pruning to extract gradient information.

[0019] The gradient information is combined with the baseline gradient to activate the weights, and the network connection is grown by adjusting the weight mask to obtain the updated network connection structure.

[0020] Based on the updated network connection structure, the Adam optimizer is used to update the network parameters, and the learning rate is dynamically adjusted in combination with the cosine annealing strategy until the preset training termination condition is reached, thereby obtaining the frequency-intensity constrained surrogate model.

[0021] Optionally, the formula for calculating the reference gradient is:

[0022]

[0023] in, Let G be the baseline gradient, |Z| be the number of active parameters, j be the j-th training sample, Z be the set of currently active connections, and G be the set of active connections. j Let M be the gradient value of the j-th connection, and M be the mask matrix. j Let be the mask matrix for the j-th connection.

[0024] Optionally, S2, using the frequency-intensity constraint surrogate model, the Big-M method is employed to transform the model constraint form, constructing a GFM-ESS addressing and dimensionality model, including:

[0025] A GFM-ESS planning optimization model is established using the total system cost coefficient, which includes the energy storage investment cost coefficient and the system operation cost coefficient.

[0026] Based on the frequency-intensity constraint surrogate model, the Big-M method is used to linearize the activation function, and the GFM-ESS planning optimization model is used to obtain the GFM-ESS planning optimization model with embedded mixed integer linear constraints.

[0027] By embedding the system's dynamic safety constraints into the GFM-ESS planning and optimization model with the mixed integer linear constraints, and combining the system planning and steady-state operational constraints, the GFM-ESS location and occupancy model is obtained.

[0028] Optionally, the system dynamic safety constraints include frequency change rate constraints, frequency minimum point constraints, and generalized short-circuit ratio constraints.

[0029] Optionally, the system planning and steady-state constraints include energy storage planning constraints, power network constraints, conventional unit operation constraints, energy storage operation constraints, new energy constraints, and carbon emission constraints.

[0030] Optionally, S3, based on the GFM-ESS location and capacity model combined with the system location strategy, obtain the planning results of the grid-type energy storage with dual security constraints of frequency and system strength, including:

[0031] Using the GFM-ESS addressing and sizing model, an initial set of candidate nodes is extracted;

[0032] The marginal increase in the generalized short-circuit ratio of the scenario is calculated based on the initial candidate node set, and the node strength gain index is obtained.

[0033] The node scene vector is calculated based on the marginal improvement of the generalized short-circuit ratio of the scenario, and the node similarity is combined with the threshold recognition sensitivity similarity to construct redundancy suppression constraints.

[0034] Based on the node strength gain index and the redundancy suppression constraint, the final candidate nodes are determined, and the planning results of the grid-type energy storage with dual security constraints of frequency and system strength are obtained.

[0035] The formula for calculating the node strength gain index is:

[0036]

[0037] In the formula,

[0038]

[0039]

[0040] in, Let be the intensity gain exponent of node i. For scene set, As scene weight, SI represents the marginal improvement in gSCR for node i in scenario w. i,w This is the system strength improvement index for node i in scenario w. It is a very small number. To meet the system requirements for the generalized short-circuit ratio, For the gSCR of node i in scenario w when GFM-ESS is not configured, To configure gSCR for node i in scenario w when using GFM-ESS, Unit test capacity.

[0041] Compared with the closest existing technology, the present invention has the following advantages:

[0042] This invention first proposes a conservative dynamic neural network surrogate model for dual security constraints, introducing an asymmetric penalty term to ensure the conservatism of the security boundary, and embedding it into a mixed-integer programming framework using the Big-M method to achieve synergistic optimization of location sizing and operation strategy. Second, it designs a synergistic location strategy that integrates node strength gain and removal of redundant node sets, accurately evaluating the marginal contribution of candidate nodes to system strength and eliminating redundant nodes, effectively compressing the planning scale and significantly improving solution efficiency while ensuring support effect. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a planning method for grid-type energy storage with dual security constraints of frequency and system strength, according to an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the iterative solution process of the energy storage planning model proposed in this embodiment of the invention;

[0046] Figure 3 The diagram shows the training process of learning the minimum frequency point and gSCR constraint proposed in the embodiment of the present invention. (a) is the gSCR training process, and (b) is the minimum frequency point training process.

[0047] Figure 4 This is a schematic diagram illustrating the approximate errors of different models in the box diagram proposed in the embodiments of the present invention;

[0048] Figure 5 This is a schematic diagram showing the minimum frequency point and typical daily results of RoCoF under different calculation examples proposed in the embodiments of the present invention;

[0049] Figure 6 These are schematic diagrams illustrating the frequency fitting of different models proposed in embodiments of the present invention;

[0050] Figure 7 This is a comparison chart of gSCR under different calculation examples proposed in the embodiments of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] The terminology used in the embodiments section of this invention is for the purpose of explaining specific embodiments of the invention only, and is not intended to limit the invention.

[0053] like Figure 1-2 As shown, this embodiment of the invention provides a planning method for grid-type energy storage with dual security constraints of frequency and system strength, including:

[0054] S1. Based on the safety constraint sample set, an asymmetric penalty term is introduced, and a frequency-intensity constraint surrogate model is obtained through dynamic pruning and growth.

[0055] Based on a sample set of security constraints related to power grid frequency and system strength, an asymmetric penalty term is introduced during network training to enhance conservatism. Simultaneously, a dynamic pruning and growth network optimization method is employed for model training, ultimately resulting in a frequency-strength constraint surrogate model that combines accuracy and lightweight design—a Convolutional Deep Neural Network (CDNN). The specific implementation process is as follows:

[0056] S1-1. Collect power grid frequency and system strength operation data to obtain a safety constraint sample set;

[0057] We collect various raw operational data, including grid frequency and system strength related dimensions under different grid operating conditions. We then organize, classify, and divide the raw data to construct a safety constraint sample set that can be used for neural network training, providing basic data support for subsequent loss function construction and model training.

[0058] S1-2. Establish a network loss function using the security constraint sample set, and obtain a composite loss function by combining it with an asymmetric penalty term;

[0059] Conservative neural network construction is carried out based on the obtained safety-constrained sample set: a composite loss function consisting of a robust regression term and an exponential safety penalty term is designed, namely the total loss function. The first term employs the Log-Cosh loss function, which is robust to outliers. The second term introduces an exponential penalty, imposing an exponential penalty on positive prediction errors, forcing the network to converge towards negative bias and achieving conservative prediction. The formula for the composite loss function is as follows:

[0060]

[0061] Where N is the size of the training set, and i is the i-th training sample. It is a hyperbolic cosine function. It is a logarithmic function. For the predicted value, y i For the true value, Here, is the conservative penalty coefficient, Regression Loss is the regression loss, and SafetyPenalty is the safety penalty term.

[0062] S1-3. Based on the composite loss function, dynamic pruning and growth are performed using a weight mask to obtain a frequency-intensity constrained surrogate model;

[0063] Based on a conservative neural network with a composite loss function, a dynamic neural network design is implemented: by introducing a weight mask, inactive connections are dynamically pruned and critical paths are activated during training. Specifically, this includes: determining the set of active connections and the baseline gradient; removing connections based on their weight magnitudes, selecting the connection with the smallest weight in each iteration and setting it to zero; reactivating potential critical connections using gradient information, selecting inactive connections with gradient magnitudes greater than a threshold for activation; updating model parameters using the Adam optimizer and dynamically adjusting the learning rate using a cosine annealing strategy. The specific implementation process is as follows:

[0064] S1-3-1. Based on the composite loss function and weight mask, determine the active connection set and the baseline gradient;

[0065] By combining the constructed composite loss function and weight masking mechanism, the neural network training initialization work is completed, and the active connection set and baseline gradient are uniformly identified and determined. This serves as the benchmark for growth assessment. The formula for calculating the benchmark gradient is as follows:

[0066]

[0067] Where |Z| is the number of active parameters, j is the j-th training sample, Z is the set of currently active (non-zero) connections, and G j Let M be the gradient value of the j-th connection, and M be the mask matrix (1 for reserved, 0 for zero). j Let be the mask matrix for the j-th connection.

[0068] S1-3-2. Read the weight magnitudes based on the active connection set, filter the least weighted connections, perform weight pruning, and extract gradient information;

[0069] Weighted pruning: Removes connections based on their weight magnitude. Specifically:

[0070] Using the initially determined set of active connections as the traversal object, network connections are filtered and eliminated based on their weight magnitude. In each iteration, the k connections with the smallest weight values ​​are uniformly selected and their weights are reset to zero. The calculation formula is as follows:

[0071]

[0072]

[0073]

[0074] Where k is the number of pruning branches. Let W be the pruning ratio, and W be the weight matrix of the neural network. Let P be the total number of elements in the weights of this layer, and let P be the set of the k connections with the smallest magnitudes. The absolute value of the weight of the j-th connection. It is the k-th smallest value after sorting by weight in ascending order. Let p be the p-th data mask matrix. Precise weight pruning is performed using the weight mask, while simultaneously extracting gradient information from each connection in the current network to provide data support for subsequent network growth operations.

[0075] S1-3-3. Use the gradient information combined with the baseline gradient to activate the weights, and grow the network connection by adjusting the weight mask to obtain the updated network connection structure.

[0076] Weight activation: Reactivating latent key connections using gradient information. Specifically:

[0077] By combining the collected gradient information with the baseline gradient obtained during the initialization phase, inactive connections after pruning are screened and selected based on their gradient magnitudes. Inactive connections exceeding a preset threshold are reactivated by adjusting the weight mask state, as calculated below:

[0078]

[0079]

[0080]

[0081] Where R is the set of reactivated connections. The absolute gradient value of the pruned connection p. Weighted growth threshold factor, M r Let M be the r-th data mask matrix. t Let t be the data mask matrix. This is the set of pruned but not reactivated connections, whose mask is updated to 0. Finally, network connection growth is completed, generating the updated network connection structure.

[0082] S1-3-4. Based on the updated network connection structure, the Adam optimizer is used to update the network parameters, and the learning rate is dynamically adjusted in combination with the cosine annealing strategy until the preset training termination condition is reached, thereby obtaining the frequency-intensity constrained surrogate model.

[0083] Parameter Update: Based on the updated network connection structure, the model parameters are updated using the Adam optimizer with the composite loss function as the error calculation basis. A cosine annealing strategy is introduced to dynamically adjust the learning rate to avoid local optima and improve model convergence. Learning Rate The decay occurs during iteration according to the following formula:

[0084]

[0085] in, Let T be the initial learning rate, t be the current iteration number, and T be the current learning rate. max The maximum iteration period is defined as follows. After completing a single iteration, the entire process of pruning, gradient verification, connection growth, and parameter updating is continuously executed in a loop until the preset training termination condition is met, ultimately resulting in a frequency-intensity constrained surrogate model.

[0086] S2. Using the frequency-intensity constraint proxy model, the Big-M method is used to transform the model constraint form and construct the GFM-ESS addressing and dimensionality model.

[0087] Based on the trained frequency-intensity constraint surrogate model, a GFM-ESS site selection and capacity determination model is developed to meet the planning requirements of grid-based energy storage while considering both frequency and system intensity. The Big-M method is used to linearly transform the safety boundary represented by the CDNN, converting it into a Mixed-Integer Linear Programming (MILP) constraint embedded within the model to achieve coordinated optimization of site selection, capacity determination, and system operation. The specific implementation process is as follows:

[0088] S2-1. Establish a GFM-ESS planning and optimization model using the total system cost coefficient;

[0089] A GFM-ESS planning and optimization model is established with the objective function of minimizing the total system cost coefficient, where the total system cost coefficient C... total Including energy storage investment cost coefficient C inv And system operating cost coefficient C op .

[0090]

[0091]

[0092]

[0093] Where r is the discount rate, Y is the lifespan of the energy storage project in years, i is the energy storage node, and c om , , c curt c c These represent the unit cost coefficients for operation and maintenance, fuel, start-up, curtailment, and carbon emission penalties, respectively. For the unit power cost of energy storage, P i For energy storage power, For the unit capacity cost of energy storage, E i Let be the energy storage capacity, s be the s-th typical scenario, S be the set of typical scenarios, g be the synchronous machine node, and P be the energy storage capacity. g,t,sFor the output of the synchronizer g at time t in the s-th typical scenario, For the abandoned power of new energy sources, Let be the charging power of the i-th energy storage unit at time t in the s-th typical scenario. Let be the charging power of the i-th energy storage unit at time t in the s-th typical scenario. Let g be the carbon emissions of the synchronizer g at time t in the s-th typical scenario. Indicates the start-up and shutdown status of the unit; the start-up and shutdown cost coefficient is obtained through... The implementation of the difference logic, , Let g be the start / stop state of the synchronizer g at time t in the s-th typical scenario. Let x be the start / stop state of synchronizer g at time t-1 in the s-th typical scenario, and let x be... .

[0094] S2-2. Based on the frequency-intensity constraint surrogate model, the Big-M method is used to linearize the activation function, and the GFM-ESS planning optimization model is used to obtain the GFM-ESS planning optimization model with embedded mixed integer linear constraints.

[0095] Based on the frequency-intensity constrained conservative dynamic neural network surrogate model that has been trained, the Big-M method is used to linearize the activation function of the CDNN for the nonlinear activation function inside the model, and the entire neural network is equivalently transformed into a mixed integer linear constraint. This is then embedded into the planning model to form a GFM-ESS planning optimization model with embedded mixed integer linear constraints.

[0096]

[0097]

[0098]

[0099]

[0100]

[0101] Wherein, Neural Network refers to a neural network. Let i be the output of neuron i in the l-th layer of the neural network, and ReLU be the activation function. The bias of neuron i in layer l. The weights between neurons i and j are... This represents the output of neuron i in the (l-1)th layer of the neural network. h is the neural network prediction value. L For the output of the neural network output layer, b L For the bias of the output layer, Let the weights be the weights of the j-th neuron in the output layer. The second to last layer (the L The output of the j-th neuron in layer -1 (serving as the input to the output layer), h represents the output of the ReLU function, x represents the input to the ReLU function, K is a sufficiently large positive constant, and a is a binary variable. The neural network predicts the value at the lowest frequency point. For system disturbance, M is the overall system inertia constant, D is the overall system damping constant, R is the overall system droop coefficient, and F is the overall system maximum reserve capacity for primary frequency regulation (upper limit of frequency regulation output). The minimum requirement for the lowest frequency point. For, U g For synchronous machine status (1 online, 0 offline), P g For the synchronous machine output, U ess In energy storage mode (1 online, 0 out of service), S ess For energy storage capacity, This is the minimum requirement for the system gSCR.

[0102] S2-3. Embed the system dynamic safety constraints into the GFM-ESS planning and optimization model with the mixed integer linear constraints, and combine the system planning and steady-state operation constraints to obtain the GFM-ESS location and capacity model;

[0103] The upgraded planning model incorporates system dynamic security constraints, including constraints on the rate of change of frequency, the minimum frequency point, and the generalized short-circuit ratio (gSCR), while also overlaying system planning and steady-state operational constraints. This results in a complete and solvable MILP model, namely the GFM-ESS site selection and capacity optimization model. System planning and steady-state operational constraints include energy storage planning constraints, power network constraints, conventional unit operation constraints, energy storage operation constraints, renewable energy constraints, and carbon emission constraints. Specifically:

[0104] (1) Energy storage planning constraints include configuration quantity limits and upper and lower limits for power and capacity configuration. The decision variables for ESS planning include energy storage location, rated power, and capacity. The calculation formula for energy storage planning constraints is as follows:

[0105]

[0106]

[0107]

[0108]

[0109] in, For the decision variable set of ESS planning, l i To determine whether to construct an ESS for node i, P i E represents the planned power of the ESS. i Indicates the capacity of the ESS. Let N be the set of candidate energy storage nodes. max L represents the number of candidate energy storage nodes. i In energy storage state, P i,s E represents the power of energy storage i in a typical scenario s. i,s For the capacity of energy storage i in a typical scenario s, , The upper and lower limits of energy storage charge and discharge, respectively. , The upper and lower limits of energy storage capacity, T min This is the minimum discharge time for energy storage. For energy storage discharge power, For energy storage capacity, T max This represents the maximum discharge time of the stored energy.

[0110] (2) Power network constraints include node power balance equations and line transmission power limits, and the calculation formulas are as follows:

[0111]

[0112]

[0113]

[0114] Among them, P g,t,s Let P be the output of the synchronizer g at time t in the s-th typical scenario, and res be the set of wind and solar turbine units. res,t,s The output of the wind turbine generator g at time t in the s-th typical scenario. Let P be the energy storage discharge power at time t in the s-th typical scenario, and l be the l-th line. l,t,s Let d be the line power at time t in the s-th typical scenario, and d be the load node. Let d be the load power at time t in the s-th typical scenario. For the charging power of energy storage i at time t in the s-th typical scenario, P l For the transmission power of the l-th line, , The phase angles of the nodes at both ends of line l are x and x, respectively. l Let l be the reactance of line l. This represents the maximum transmission power of line l.

[0115] (3) Conventional unit operation constraints include the upper and lower limits of unit output, ramp-up, and start-up / shutdown constraints, and the calculation formula is as follows:

[0116]

[0117]

[0118]

[0119]

[0120] in, The synchronous machine's operating state is a 0-1 variable. , These represent the minimum and maximum outputs of the synchronous machine, R. g This refers to the ramp rate of the synchronous machine. For the output of the synchronizer g at time t in the s-th typical scenario, For the output of the synchronizer g at time t-1 in the s-th typical scenario, , These are the minimum start and stop times, u g,t Let u be the state of synchronizer g at time t. g,t-1 Let g be the state of the synchronizer at time t-1.

[0121] (4) Energy storage operation constraints include charging and discharging power limits, complementary charging and discharging states constraints, SOC dynamic equations, and intraday energy balance constraints, and the calculation formulas are as follows:

[0122]

[0123]

[0124]

[0125]

[0126]

[0127] in, In energy storage charging state, In the energy storage and discharge state, E i,t,s Let E be the capacity of energy storage i at time t in the s-th typical scenario. i,t-1,s Let i be the capacity of energy storage at time t-1 in the s-th typical scenario. For energy storage self-discharge efficiency, , These are the energy storage charging and discharging efficiencies, Step size, SOC max SOC min These are the upper and lower limits of the stored charge state, E i For energy storage capacity, E i,T,sLet i be the capacity of energy storage at time T(24) in the s-th typical scenario.

[0128] (5) The formula for calculating the new energy constraint is as follows:

[0129]

[0130] Among them, P res,t,s This represents the actual output power received by the renewable energy units during time period t under scenario s. This represents the acceptable output of renewable energy units during time period t under scenario s. This represents the curtailment power of renewable energy units during time period t under scenario s, which is the difference between the acceptable output and the actual accepted output of renewable energy units.

[0131] (6) The formula for calculating carbon emission constraints is as follows:

[0132]

[0133] Among them, EF g Synchronous machine carbon emission factor This refers to the carbon emissions of the synchronizing machine.

[0134] S3. Based on the GFM-ESS location and capacity model and the system location strategy, obtain the planning results of dual security constraints of network-type energy storage frequency and system strength;

[0135] like Figure 2 As shown, a collaborative site selection method is proposed for the GFM-ESS site selection and capacity grading model, which integrates multi-scenario strength gain and redundant node set removal (i.e., redundancy suppression). This method combines node strength gain calculation and redundant node suppression logic under multiple scenarios. First, it quantitatively evaluates the marginal support contribution of each candidate node to the generalized short-circuit ratio (gSCR) after energy storage is connected. Then, based on the node support characteristics, redundant candidate nodes with overlapping effects are eliminated, narrowing the solution scope of the model before solving the planning model. The result outputs a complete planning result with dual security constraints of frequency and system strength for grid-type energy storage. The specific implementation process is as follows:

[0136] S3-1. Using the GFM-ESS addressing and sizing model, candidate nodes are extracted;

[0137] The MILP form of the GFM-ESS location and capacity model, which has been linearized by the CDNN neural network using the Big-M method and has embedded the minimum frequency point constraint and gSCR system strength constraint, is retrieved. All grid nodes that can be deployed for grid-type energy storage are read from the model and organized into an initial candidate node set L, which serves as the basic data carrier for subsequent node strength calculation and redundancy screening.

[0138] S3-2. Calculate the marginal increase of the generalized short-circuit ratio of the scenario based on the initial candidate node set, and obtain the node strength gain index;

[0139] Based on scene set The marginal increase in the generalized short-circuit ratio (SNR) for each scenario after connecting candidate nodes to grid-connected devices of unit capacity is calculated. Then, after scenario weighting and index standardization, the Node Strength Gain Index (SGI) is calculated, which is the system strength improvement index of the node. This quantifies the support capability of each node for grid strength after installing energy storage. Specifically:

[0140] Define the Strength Gain Index (SGI) for node i. The weighted enhancement capability of nodes to system strength in all typical scenarios is comprehensively measured, and the calculation formula is as follows:

[0141]

[0142] in, It consists of the normalized multi-scenario gSCR marginal improvement and scenario weights. Let g be the marginal improvement in SCR for node i in scenario w. For scene weights.

[0143]

[0144]

[0145]

[0146] in, Let be the intensity gain exponent of node i. For scene set, As scene weight, SI represents the marginal improvement in gSCR for node i in scenario w. i,w This is the system strength improvement index for node i in scenario w. It is a very small number. To meet the system requirements for the generalized short-circuit ratio, For the gSCR of node i in scenario w when GFM-ESS is not configured, To configure gSCR for node i in scenario w when using GFM-ESS, Unit test capacity.

[0147] S3-3. Calculate the node scene vector based on the marginal improvement of the generalized short-circuit ratio of the scenario, and combine the node similarity judgment threshold recognition sensitivity similarity to construct redundancy suppression constraints.

[0148] Based on the marginal improvement of gSCR for all initial candidate nodes in various typical scenarios, the scenario vector of node i is defined. The sensitivity similarity between nodes is calculated using the following formula:

[0149]

[0150]

[0151] Where i and j are node numbers, r ij Let be the sensitivity similarity between node i and node j. Transpose the scene sensitivity vector for node i. Let j be the scene sensitivity vector. Let the scene sensitivity vector of node i be the 2-norm. Let be the 2-norm of the scene sensitivity vector for node j.

[0152] The recognition sensitivity similarity exceeds the preset threshold (i.e., the node similarity judgment threshold r). max Redundant nodes are identified by using similarity criteria to construct redundancy suppression constraints, which are used to constrain the repeated selection of energy storage installation sites with similar functions during the optimization process.

[0153] S3-4. Based on the node strength gain index and the redundancy suppression constraint, determine the final candidate nodes and obtain the planning results of dual security constraints on network-type energy storage frequency and system strength.

[0154] All initial nodes are prioritized according to their Strength Gain Index (SGI) from highest to lowest. Then, redundancy suppression constraints are applied to eliminate overlapping redundant nodes, resulting in a final candidate node. This final candidate node will simultaneously meet the dual safety constraints of the model's rate of change, minimum frequency point, and gSCR. The final candidate node is determined based on the Strength Gain Index and redundancy suppression constraints, calculated as follows:

[0155]

[0156]

[0157] Among them, L i L j Let nodes i and j be candidate states, and S be the set of redundant nodes. For the final candidate node combination, k is the node number, and L\S is the new candidate node set formed after removing redundant nodes from the initial candidate node set.

[0158] After the final candidate nodes are selected, they are substituted into the MILP model for iterative solution. The energy storage planning scheme is continuously updated until the total system cost is minimized, and the network-type energy storage site selection-capacity planning result that meets the dual safety constraints of frequency and system strength is output.

[0159] In summary, the core work of this invention is to construct a GFM-ESS addressing and capacity optimization model that balances frequency stability and system strength support. First, a conservative dynamic neural network surrogate model oriented towards dual security constraints is proposed, introducing an asymmetric penalty term to ensure the conservatism of the security boundary. This model is then embedded into a mixed-integer programming framework using the Big-M method to achieve coordinated optimization of addressing and capacity sizing and operational strategies. Second, a collaborative addressing strategy that integrates node strength gain and the removal of redundant node sets is designed. This strategy accurately evaluates the marginal contribution of candidate nodes to system strength and eliminates redundant nodes, effectively compressing the planning scale and significantly improving solution efficiency while ensuring support effectiveness.

[0160] To verify the effectiveness of the proposed planning method, a high-proportion renewable energy test system was constructed based on the improved IEEE 39-node system. By adjusting some conventional unit parameters, three thermal power units (total capacity 1200MW) were retained, and three wind farms (total capacity 900MW) and four photovoltaic farms (total capacity 900MW) were connected. In the planning model, the maximum configuration capacity of a single ESS was set at 800MW. Relevant technical parameters are shown in Tables 1-3, where Table 1 is a summary table of energy storage-related parameters, Table 2 is a summary table of neural network model parameters, and Table 3 is a summary table of parameters for generalized short-circuit ratio calculation. Considering the seasonal fluctuations in wind and photovoltaic output and load demand, typical daily operating scenarios were extracted based on historical data using the K-means clustering algorithm; the specific clustering process is not detailed here. Furthermore, the following safety constraints were set: disturbance power of 10% of the load, frequency change rate threshold of 0.5Hz / s, frequency deviation threshold of 0.5Hz, and minimum gSCR requirement of 3. In Table 4, X... g X is the equivalent additional reactance of the synchronous machine. n For the equivalent additional reactance of grid-type energy storage, S g For the capacity of the synchronous machine, S n This refers to grid-type energy storage capacity.

[0161] Table 1

[0162]

[0163] Table 2

[0164]

[0165] Table 3

[0166]

[0167] Table 4

[0168]

[0169] (1) Training process

[0170] To verify the conservatism and high accuracy of CDNN, its fitting performance was compared with that of neural networks (NN), conservative neural networks (C-NN), and deep neural networks (D-NN) under unified conditions. Figure 3 (a)- Figure 3 As shown in (b), the convergence process of each model to the generalized short-circuit ratio and the minimum frequency safety boundary is illustrated. The results show that C-NN suffers from insufficient accuracy due to its high MSE; D-NN, with its dynamic structure adjustment mechanism, outperforms NN in accuracy. In contrast, CDNN, with its unique dynamic structure adjustment mechanism and conservative design, not only achieves a lower MSE than C-NN to ensure high fitting accuracy, but also exhibits conservatism in its error distribution. Therefore, CDNN can safely and accurately approximate the double safety boundary, highlighting its core advantage in characterizing safety constraints.

[0171] like Figure 4 As shown, the approximation error distributions of the four models are further illustrated. The error distributions of NN and D-NN are close to zero mean, with both positive and negative errors being significant, posing a risk of overestimating the safety margin. While C-NN exhibits strong conservatism, its error dispersion is large, limiting its fitting accuracy. In contrast, CDNN's errors are mainly concentrated in the negative region and have a more compact distribution, indicating that it can maintain high approximation accuracy while suppressing non-conservative predictions.

[0172] (2) Energy storage planning results

[0173] pass Figure 2 The iterative solution method shown yielded ESS configuration schemes for different examples, with specific results illustrated in Table 5. Example 1 did not consider safety constraints, resulting in a relatively small energy storage configuration. Example 2, by introducing frequency constraints, significantly increased energy storage power and capacity, leading to a 20.11% increase in investment cost compared to Example 1, but a 7.25% decrease in operating cost. Example 3 increased the ESS configuration scale compared to considering frequency safety, to meet the strength support requirements in weak system scenarios. Example 4, considering the dual support of ESS for both frequency safety and system strength, achieved the largest ESS configuration capacity. Compared to Example 2, its investment cost increased by 28.96%, but its operating cost decreased by 1.52%, resulting in a total cost increase of only 3.25%. The results demonstrate that the proposed model can meet dual safety constraints with a moderate investment increment and reduce operating costs by minimizing conventional unit operation.

[0174] Table 5

[0175]

[0176] To analyze the impact of frequency security constraints on ESS planning results, we compare Case 1 and Case 2 in Table 5. After introducing frequency constraints, the ESS capacity and investment demand increase significantly, indicating that frequency security has a significant constraining effect on energy storage planning. Figure 5 As shown, the results for typical days with high frequency limit exceedance risk are presented. It can be seen that only in Example 2, which considers frequency constraints, and Example 4, which has dual safety constraints, both the minimum frequency point and RoCoF meet the safety criteria. However, Examples 1 and 3 still exhibit frequency limit exceedance risk. This indicates that system strength support cannot replace dynamic frequency support; the two differ in their mechanisms of action and time scales. Therefore, it is necessary to embed both frequency and system strength constraints into the ESS planning model to achieve coordinated protection of system operational safety.

[0177] To verify the robustness of the proposed CDNN surrogate model, a comparative analysis was conducted on typical operating conditions where the frequency safety index approaches the safety threshold. For example... Figure 6 As shown, under the constraint of a frequency safety lower limit of 49.5Hz: when the CDNN predicts the lowest frequency at 49.50Hz, the corresponding analytical model's true value is 49.53Hz, indicating that the scheme is actually within the safety margin; however, when the NN predicts the same value of 49.50Hz, the corresponding analytical true value is only 49.47Hz, indicating that the scheme has actually exceeded the frequency limit. The above comparison results show that conventional NNs are prone to making non-conservative decisions due to overestimating safety indicators, misjudging potentially limit-exceeding schemes as feasible; while CDNNs can effectively suppress the overestimation of safety margins, accurately identify and eliminate risky schemes, thereby significantly improving the robustness of ESS planning results.

[0178] like Figure 7 As shown, the gSCR changes within a typical day are compared. Examples 1 and 2 show periods where the gSCR is below the safety threshold, indicating that constraining only the frequency response is insufficient to guarantee the strength security of a weak power grid. In contrast, Example 4 satisfies both the minimum frequency point and the gSCR constraint. These results validate the effectiveness of the proposed dual-security-constraint planning model in real-world power grids with high RES penetration, demonstrating its ability to synergistically improve frequency dynamic security and system strength margin.

[0179] 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.

[0180] 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 this application. 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.

[0181] 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.

[0182] 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.

[0183] 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 planning method for grid-type energy storage with dual security constraints of frequency and system strength, characterized in that, include: S1. Based on the safety constraint sample set, an asymmetric penalty term is introduced, and a frequency-intensity constraint surrogate model is obtained through dynamic pruning and growth. S2. Using the frequency-intensity constraint proxy model, the Big-M method is used to transform the model constraint form and construct the GFM-ESS addressing and dimensionality model. S3. Based on the GFM-ESS location and capacity model and the system location strategy, obtain the planning results of dual security constraints of network-type energy storage frequency and system strength.

2. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 1, characterized in that, S1. Based on the safety constraint sample set, an asymmetric penalty term is introduced, and a frequency-intensity constraint surrogate model is obtained through dynamic pruning and growth, including: Collect power grid frequency and system strength operation data to obtain a safety constraint sample set; A network loss function is established using the aforementioned security constraint sample set, and a composite loss function is obtained by combining it with an asymmetric penalty term. Based on the composite loss function, dynamic pruning and growth are performed using a weight mask to obtain a frequency-intensity constrained surrogate model.

3. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 2, characterized in that, The formula for calculating the composite loss function is: ， in, Let N be the total loss function, N be the number of training samples, and i be the i-th training sample. It is a hyperbolic cosine function. It is a logarithmic function. For the predicted value, y i For the true value, The conservative penalty coefficient is denoted as Regression Loss, and Safety Penalty is denoted as the safety penalty term.

4. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 2, characterized in that, Based on the aforementioned composite loss function, dynamic pruning and growth are performed using a weight mask to obtain a frequency-intensity constrained surrogate model, including: Based on the composite loss function and weight mask, the active connection set and baseline gradient are determined; The weight magnitudes are read from the active connection set, and the least weighted connections are filtered for weight pruning to extract gradient information. The gradient information is combined with the baseline gradient to activate the weights, and the network connection is grown by adjusting the weight mask to obtain the updated network connection structure. Based on the updated network connection structure, the Adam optimizer is used to update the network parameters, and the learning rate is dynamically adjusted in combination with the cosine annealing strategy until the preset training termination condition is reached, thereby obtaining the frequency-intensity constrained surrogate model.

5. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 4, characterized in that, The formula for calculating the reference gradient is: ， in, Let G be the baseline gradient, |Z| be the number of active parameters, j be the j-th training sample, Z be the set of currently active connections, and G be the set of active connections. j Let M be the gradient value of the j-th connection, and M be the mask matrix. j Let be the mask matrix for the j-th connection.

6. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 1, characterized in that, S2. Using the frequency-intensity constraint surrogate model, the Big-M method is employed to transform the model constraint form, constructing a GFM-ESS addressing and dimensionality model, including: A GFM-ESS planning optimization model is established using the total system cost coefficient, which includes the energy storage investment cost coefficient and the system operation cost coefficient. Based on the frequency-intensity constraint surrogate model, the Big-M method is used to linearize the activation function, and the GFM-ESS planning optimization model is used to obtain the GFM-ESS planning optimization model with embedded mixed integer linear constraints. By embedding the system's dynamic safety constraints into the GFM-ESS planning and optimization model with the mixed integer linear constraints, and combining the system planning and steady-state operational constraints, the GFM-ESS location and occupancy model is obtained.

7. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 6, characterized in that, The system's dynamic security constraints include frequency change rate constraints, frequency minimum point constraints, and generalized short-circuit ratio constraints.

8. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 6, characterized in that, The system planning and steady-state constraints include energy storage planning constraints, power network constraints, conventional unit operation constraints, energy storage operation constraints, new energy constraints, and carbon emission constraints.

9. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 1, characterized in that, S3. Based on the GFM-ESS location and capacity model and the system location strategy, obtain the planning results of the grid-type energy storage with dual security constraints of frequency and system strength, including: Using the GFM-ESS addressing and sizing model, an initial set of candidate nodes is extracted; The marginal increase in the generalized short-circuit ratio of the scenario is calculated based on the initial candidate node set, and the node strength gain index is obtained. The node scene vector is calculated based on the marginal improvement of the generalized short-circuit ratio of the scenario, and the node similarity is combined with the threshold recognition sensitivity similarity to construct redundancy suppression constraints. Based on the node strength gain index and the redundancy suppression constraint, the final candidate nodes are determined, and the planning results of the grid-type energy storage with dual security constraints of frequency and system strength are obtained.

10. The planning method for grid-type energy storage with dual security constraints of frequency and system strength according to claim 9, characterized in that, The formula for calculating the node strength gain index is: ， In the formula, , ， ， in, Let be the intensity gain exponent of node i. For scene set, As scene weight, SI represents the marginal improvement in gSCR for node i in scenario w. i,w This is the system strength improvement index for node i in scenario w. It is a very small number. To meet the system requirements for the generalized short-circuit ratio, For the gSCR of node i in scenario w when GFM-ESS is not configured, To configure gSCR for node i in scenario w when using GFM-ESS, Unit test capacity.

Images