A wind storage frequency safety constraint optimization scheduling method considering transformer power limiter

Abstract

The application discloses a wind storage frequency safety constraint optimization scheduling method considering transformer power amplitude limiting, and belongs to the field of frequency safety constraint optimization scheduling, and comprises the following steps: according to the frequency response characteristic difference of grid-connected transformer power and network-constructed transformer resources, a system frequency response model is constructed; power amplitude constraints of multi-source inertia response are derived, and initial frequency change rate constraints are constructed; a region of interest active sampling is adopted to generate a training data set; a composite loss function is designed, a multi-input-single-output deep neural network fitting model is trained, the model is converted into a mixed integer linear constraint through a big M method, and a maximum frequency deviation constraint is constructed; a steady-state frequency deviation constraint is constructed; and a basic optimization scheduling model is constructed, and the model is solved in combination with frequency safety constraints to obtain an optimization scheduling result which takes into account economic efficiency and frequency safety. The application effectively improves the frequency stability level of a high-proportion new energy power system, and provides theoretical support for optimization scheduling decision-making.

Description

Technical Field

[0001] This invention belongs to the field of frequency security constraint optimization scheduling, and specifically relates to a wind power storage and thermal power frequency security constraint optimization scheduling method that considers converter power supply power limitation. Background Technology

[0002] With the advancement of new power system construction, high-proportion renewable energy and high-proportion power electronic equipment are deeply integrated into the grid. The coordinated operation of grid-connected energy storage, grid-connected wind farms, and traditional thermal power units has become the mainstream form of the power grid. Unit combination, as a core link in power system dispatch optimization, is crucial for the safe and efficient operation of the new power system due to its decision-making design that balances operational economy and frequency security and stability. However, renewable energy grid integration significantly reduces the system's inertia support capacity. There are fundamental differences in the frequency response characteristics between grid-connected and grid-connected converter power supplies. Due to the phase-locked loop filtering and delay characteristics, grid-connected converter power supplies cannot achieve instantaneous inertia response. When the system encounters disturbances, it is prone to problems such as excessive initial frequency change rate, exceeding maximum frequency deviation limits, and exceeding steady-state frequency deviation limits, seriously threatening frequency security. At the same time, the strong volatility of renewable energy output further exacerbates the complexity of dynamic frequency response, placing higher demands on the frequency security constraints of optimized dispatch.

[0003] Current research on frequency security-constrained optimal scheduling still has significant shortcomings. Some methods ignore the dynamic characteristics of frequency, some fail to consider the differences in frequency response characteristics between grid-connected and grid-connected power sources, and some use discretization methods to model the power limiting link, which introduces a large number of decision variables, resulting in low model solution efficiency. The optimized scheduling schemes are difficult to balance operation economy and frequency stability, and cannot adapt to the scheduling needs of new power systems.

[0004] How to solve the above problems and construct a frequency security constraint optimization scheduling model that is adaptable to different control types, considers power limiting, and solves efficiently has become a key challenge in the scheduling optimization of new power systems. Summary of the Invention

[0005] This invention addresses the issues of insufficient system inertia support and limited frequency safety under high-proportion renewable energy grid integration. It provides an optimized scheduling method for wind-storage-thermal energy farms that considers converter power limiting. First, it considers the power limiting stage, constructing a system frequency response model by combining the differences in frequency response characteristics between grid-connected and grid-connected converter power supplies. Second, it considers the power limiting of the initial inertia response of grid-connected energy storage and grid-connected wind farms, deriving the power limiting constraint for multi-source inertia response and constructing an initial frequency change rate constraint. Next, it trains a deep neural network classification model and generates a training dataset near the frequency safety threshold using a region-of-interest active sampling method based on the classification probability. Finally, it designs a composite loss function and trains a multi-input-single-output deep neural network fitting model to... This invention fits the nonlinear mapping relationship between scheduling characteristic variables and the maximum frequency deviation, and transforms the trained multi-input single-output deep neural network fitting model into a mixed-integer linear constraint using the Big M method, thus constructing the maximum frequency deviation constraint. Subsequently, leveraging the characteristic that the frequency change rate is zero in steady state, and combining the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, auxiliary variables are introduced to characterize the difference between the unconstrained steady-state frequency response power and the actual steady-state frequency response power, establishing a correlation constraint between steady-state frequency response and steady-state frequency deviation, and constructing a steady-state frequency deviation constraint. Finally, a basic optimized scheduling model is constructed and solved in conjunction with the above frequency security constraints to obtain an optimized scheduling result that balances economy and frequency security. This invention specifically overcomes the shortcomings of existing technologies, achieving efficient integration and solution of frequency security constraint construction and optimized scheduling model, synergistically improving the economy and frequency security of power system scheduling, and supporting the technical requirements of high-proportion renewable energy grid connection.

[0006] According to one aspect of this specification, a wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitations is provided, comprising:

[0007] Based on the power limiting elements of the grid-connected converter power supply and the power limiting elements of the grid-connected converter power supply, and combined with the differences in the frequency response characteristics of the grid-connected converter power supply and the power limiting elements of the grid-connected converter power supply, a system frequency response model is constructed.

[0008] Based on the power limiting of the initial inertial response of grid-connected energy storage and grid-connected wind farm, the power limiting constraints of the initial inertial response of grid-connected wind farm and grid-connected energy storage are derived. Combined with the mechanical inertia of the system, the initial frequency change rate constraint is constructed.

[0009] The unlabeled sample set is obtained based on the input variables in the system frequency response model. The unlabeled sample set is then input into the trained deep neural network classification model to obtain the predicted probability that each unlabeled sample belongs to the frequency safe category and the frequency unsafe category. The region of interest active sampling method is then used to generate the training dataset.

[0010] By combining the constructed composite loss function, the multi-input single-output deep neural network fitting model is trained using the training dataset to obtain the trained multi-input single-output deep neural network fitting model. The trained multi-input single-output deep neural network fitting model is then transformed into mixed integer linear constraints using the Big M method, and the maximum frequency deviation constraint is constructed.

[0011] Combining the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms and grid-connected wind farms, auxiliary variables are introduced to establish the correlation constraints between steady-state frequency response and steady-state frequency deviation, and steady-state frequency deviation constraints are constructed.

[0012] A basic optimization scheduling model is constructed, and a frequency security constraint optimization scheduling model is formed by combining it with frequency security constraints. This model is then transformed into a mixed integer linear programming problem, and the optimization scheduling result is obtained by solving the problem. The frequency security constraints include initial frequency change rate constraints, maximum frequency deviation constraints, and steady-state frequency deviation constraints.

[0013] As a further technical solution, the power limiting circuit of the grid-connected converter power supply includes:

[0014] The unlimited frequency response power of the grid-type converter power supply is limited to its preset frequency regulation reserve range: when the unlimited frequency response power does not exceed the frequency regulation reserve limit, the actual frequency response power is equal to the unlimited frequency response power; when the unlimited frequency response power exceeds the frequency regulation reserve limit, the actual frequency response power is equal to the frequency regulation reserve limit.

[0015] The unlimited frequency response power of the grid converter power supply is limited to a preset frequency regulation reserve range: when the unlimited frequency response power exceeds the frequency regulation reserve limit, the actual frequency response power is equal to the frequency regulation reserve limit; when the unlimited frequency response power does not exceed the frequency regulation reserve limit, the actual frequency response power is equal to the unlimited frequency response power.

[0016] As a further technical solution, the expression for the initial frequency change rate constraint is:

[0017]

[0018] in, Let represent the disturbance power at time t. This represents the initial inertia response power of an unlimited grid-connected wind farm. This represents the initial inertial response power of an unconstrained grid-type energy storage system. Represents the mechanical inertia of the system. This indicates the limit of the rate of change of frequency.

[0019] As a further technical solution, a region-of-interest active sampling method is used to generate a training dataset, including:

[0020] The sampling probability of each unlabeled sample is determined based on the predicted probability of each unlabeled sample belonging to the frequency safe category and the frequency unsafe category. Input samples that meet the frequency safe threshold requirements are then extracted from high to low according to the sampling probability to form a sample set to be labeled.

[0021] Each input sample in the unlabeled sample set is substituted into the system frequency response model one by one to obtain the corresponding actual maximum frequency deviation. Each input sample and the corresponding actual maximum frequency deviation are then combined to form a labeled sample, which constitutes the training dataset.

[0022] As a further technical solution, the training of the deep neural network classification model includes:

[0023] An initial sample set is randomly generated based on the boundary values ​​of the input variables in the system frequency response model.

[0024] Substitute each sample in the initial sample set into the system frequency response model and solve for the actual maximum frequency deviation.

[0025] Based on the relationship between the actual maximum frequency deviation and the preset maximum frequency deviation safety threshold, a frequency safety category label is set to form an initial labeled sample set;

[0026] A deep neural network classification model is constructed, including an input layer for receiving input samples and transmitting them to a hidden layer; at least one hidden layer for processing the input samples using an activation function to obtain hidden layer representation features; and an output layer for receiving the hidden layer representation features and outputting the frequency-safe category and category probability corresponding to the input sample; adjacent network layers are fully connected.

[0027] The deep neural network classification model is trained using the initial labeled sample set to obtain the trained deep neural network classification model.

[0028] As a further technical solution, the trained multi-input single-output deep neural network fitting model is transformed into a mixed-integer linear constraint using the Big M method, including:

[0029] The trained multi-input single-output deep neural network fitting model is decomposed into the linear transformation relationship of each network layer and the ReLU activation operation relationship of each hidden layer.

[0030] Binary variables are set to represent the activation state of each ReLU neuron, and the Big M method is used to construct linear inequality constraints to represent the ReLU activation operation relationship;

[0031] The linear transformation relationship of each network layer is represented by a linear equation. The trained multi-input single-output deep neural network fitting model is represented as a mixed integer linear constraint consisting of continuous variables, binary variables, linear equations and linear inequalities. The output value of the multi-input single-output deep neural network fitting model is a continuous variable.

[0032] As a further technical solution, a steady-state frequency deviation constraint is constructed, including:

[0033] By utilizing the characteristic that the rate of frequency change is zero in steady state, the system swing equation in steady state is constructed;

[0034] Combining the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, auxiliary variables are introduced to characterize the difference between the unconstrained steady-state frequency response power and the actual steady-state frequency response power. Correlation constraints between steady-state frequency response and steady-state frequency deviation are established, and steady-state frequency deviation constraints are constructed.

[0035] As a further technical solution, a basic optimized scheduling model is constructed, including:

[0036] The objective function for optimizing scheduling is to minimize the reserve cost, operating cost, and penalty cost of the optimized scheduling. A basic optimized scheduling model is constructed by combining the general operating constraints of each frequency response resource. The general operating constraints include the start-up and shutdown of thermal power units, upper and lower limits of output, frequency regulation reserve and ramping constraints, energy storage charging and discharging power, charging and discharging state and energy state constraints, and wind farm output, frequency regulation reserve and system power balance constraints.

[0037] As a further technical solution, the constructed composite loss function is expressed as follows:

[0038]

[0039] in, Let represent the composite loss function, and n represent the index of the nth sample in the training dataset. This indicates the number of samples obtained. This represents the weight of the nth sample. This represents the predicted output value. This represents the true output value of the nth sample. Indicates the penalty coefficient. This represents the ReLU function.

[0040] According to one aspect of this specification, a wind-storage-thermal frequency security constraint optimization scheduling system considering converter power supply limitations is provided, for implementing the aforementioned wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitations, comprising:

[0041] The first processing module is used to construct a system frequency response model based on the power limiting circuits of the grid-connected converter power supply and the power limiting circuits of the grid-connected converter power supply, combined with the differences in frequency response characteristics between the grid-connected converter power supply and the power limiting circuits of the grid-connected converter power supply.

[0042] The second processing module is used to derive the power limiting constraints of the initial inertia responses of grid-connected wind farms and grid-connected energy storage based on the power limiting of the initial inertia responses of grid-connected energy storage and grid-connected wind farms, and to construct the initial frequency change rate constraint in combination with the mechanical inertia of the system.

[0043] The third processing module is used to obtain an unlabeled sample set based on the input variables in the system frequency response model, input the unlabeled sample set into the trained deep neural network classification model, obtain the predicted probability of each unlabeled sample belonging to the frequency safe category and the frequency unsafe category, and generate a training dataset using the region interest active sampling method.

[0044] The fourth processing module is used to combine the constructed composite loss function and train the multi-input single-output deep neural network fitting model using the training dataset to obtain the trained multi-input single-output deep neural network fitting model. The trained multi-input single-output deep neural network fitting model is then transformed into mixed integer linear constraints using the Big M method, and the maximum frequency deviation constraint is constructed.

[0045] The fifth processing module combines the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms and grid-connected wind farms, introduces auxiliary variables to establish the correlation constraints between steady-state frequency response and steady-state frequency deviation, and constructs steady-state frequency deviation constraints.

[0046] The sixth processing module is used to construct a basic optimization scheduling model, combine it with frequency security constraints to form a frequency security constraint optimization scheduling model, and transform it into a mixed integer linear programming problem to obtain the optimization scheduling result. The frequency security constraints include initial frequency change rate constraints, maximum frequency deviation constraints, and steady-state frequency deviation constraints.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. This invention constructs a frequency response model that considers the differences in frequency response and power limiting characteristics between grid-connected and grid-connected converter power supplies. This model can accurately characterize the dynamic response and actual power support boundary of wind power, energy storage, and thermal power resources during disturbances. It effectively solves the problem that existing methods ignore the differences in response and power limiting characteristics of different control types, and improves the applicability of frequency dynamic modeling.

[0049] 2. This invention generates a training dataset near a frequency safety threshold using a deep neural network classification model and a region-of-interest (ROI) active sampling method. It then trains a multi-input, single-output deep neural network fitting model using a composite loss function, achieving high-precision, conservative fitting of the maximum frequency deviation and its mixed-integer linearization. Compared to traditional discretization methods, this approach reduces the number of variables and constraints, decreases model size, and improves solution efficiency, thereby achieving a coordinated optimization of system operational economy and frequency safety.

[0050] 3. This invention proposes a multi-dimensional frequency security constraint method that covers the initial stage of disturbance, transient process and steady state stage by constructing initial frequency change rate, maximum frequency deviation and steady state frequency deviation constraints. It can comprehensively evaluate the frequency security level after the expected disturbance, effectively solve the problem of incomplete consideration of frequency security indicators in existing scheduling methods, and improve the security and reliability of optimized scheduling results. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 A flowchart of a wind-storage-thermal frequency safety constraint optimization scheduling method considering converter power supply power limitation provided in an embodiment of the present invention.

[0053] Figure 2 This is a system frequency response model diagram provided in an embodiment of the present invention.

[0054] Figure 3 This is a topology diagram of an IEEE 30-node embodiment provided in this invention.

[0055] Figure 4 This is a schematic diagram of the frequency change rate, maximum frequency deviation, and steady-state frequency deviation of the system provided in the embodiments of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figure 1As shown, this embodiment of the invention proposes a wind-storage-thermal frequency safety constraint optimization scheduling method considering converter power limiting, including: Step 1, constructing a system frequency response model based on the power limiting links of grid-connected and grid-connected converter power supplies, combined with the differences in frequency response characteristics between the two power supplies; Step 2, deriving the power limiting constraints of the initial inertia responses of grid-connected wind farms and grid-connected energy storage based on the power limiting of the initial inertia responses of grid-connected energy storage and grid-connected wind farms, and constructing initial frequency change rate constraints in combination with the mechanical inertia of the system; Step 3, obtaining an unlabeled sample set based on the input variables in the system frequency response model, inputting the unlabeled sample set into a trained deep neural network classification model to obtain the predicted probability of each unlabeled sample belonging to the frequency safe category and the frequency unsafe category, and generating a training dataset using a region of interest active sampling method; Step 4. Combining the constructed composite loss function, the multi-input single-output deep neural network fitting model is trained using the training dataset to obtain the trained multi-input single-output deep neural network fitting model. The trained multi-input single-output deep neural network fitting model is then transformed into mixed-integer linear constraints using the Big M method, and a maximum frequency deviation constraint is constructed. Step 5: Combining the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, auxiliary variables are introduced to establish the correlation constraint between steady-state frequency response and steady-state frequency deviation, and a steady-state frequency deviation constraint is constructed. Step 6: A basic optimization scheduling model is constructed, and a frequency safety constraint optimization scheduling model is formed by combining frequency safety constraints and transforming it into a mixed-integer linear programming problem. The solution yields an optimization scheduling result that balances economy and frequency safety. The frequency safety constraints include initial frequency change rate constraints, maximum frequency deviation constraints, and steady-state frequency deviation constraints.

[0058] Specifically, step 1 includes the following:

[0059] The power system frequency response resources proposed in this embodiment of the invention include thermal power units and converter power supplies. The converter power supply refers to a power source or energy storage device that is connected to the power system and participates in frequency response through a power electronic converter. Depending on the control method adopted, the converter power supply includes grid-connected converter power supplies and grid-linked converter power supplies. Grid-connected converter power supplies include grid-connected energy storage and grid-connected wind farms, while grid-linked converter power supplies include grid-linked wind farms. The terms "grid-connected" and "grid-linked" respectively indicate that the converter power supply adopts a grid-connected control method and a grid-linked control method, and are not independent power system component types.

[0060] 1. The frequency response characteristics of grid-connected converter power supplies differ from those of grid-connected converter power supplies, as detailed below:

[0061] The grid-type converter power supply does not rely on a phase-locked loop in terms of frequency sensing. It can generate its own frequency, actively respond to frequency changes and provide instantaneous inertial support. Its inertial response speed is fast and its frequency response has no inherent lag.

[0062] In terms of frequency sensing, grid-connected converter power supplies need to synchronize with the grid frequency through a phase-locked loop (PLL). They can only passively follow frequency changes. The filtering characteristics and inherent delay of the PLL will cause a significant lag in its frequency response. Furthermore, the inertial support capability needs to be achieved through additional virtual inertial control commands, resulting in a slow inertial response speed.

[0063] 2. The power limiting circuitry for grid-connected converter power supplies is as follows:

[0064] For grid-type converter power supplies, according to Figure 2 The frequency response branches of the grid-connected energy storage and grid-connected wind farms shown are equipped with virtual inertial response and droop response circuits, respectively. The virtual inertial response circuit generates inertial response power based on the system frequency change rate, while the droop response circuit generates primary frequency regulation power based on the system frequency deviation. The dynamic response process of the primary frequency regulation power is characterized by corresponding time constants. The virtual inertial response power and the primary frequency regulation power are superimposed to form the unlimited frequency response power of the grid-connected converter power supply, which is then input to the power limiting circuit. The power limiting circuit restricts the unlimited frequency response power of the grid-connected converter power supply to its preset frequency regulation reserve range. When the unlimited frequency response power does not exceed the frequency regulation reserve upper limit, the actual frequency response power equals the unlimited frequency response power; when it exceeds the frequency regulation reserve upper limit, the actual frequency response power equals the frequency regulation reserve upper limit. The limited actual frequency response power is input to the system power balance circuit to characterize the active support role of the grid-connected converter power supply for the system frequency.

[0065] For grid-connected converter power supplies, according to Figure 2The frequency response branch of the grid-connected wind farm shown in the diagram, based on the virtual inertial response and droop response stages, further incorporates a frequency measurement delay stage and a first-order inertial stage to characterize the filtering and delay characteristics caused by the phase-locked loop, frequency measurement, and control execution processes. Specifically, the virtual inertial response stage generates inertial response power based on the frequency change rate after measurement and delay processing, while the droop response stage generates primary frequency modulation power based on the frequency deviation after measurement and delay processing. Both are characterized by their dynamic response processes through corresponding first-order inertial stages. The inertial response power and the primary frequency modulation power are superimposed to form the unlimited frequency response power of the grid-connected converter power supply, which is then limited to its preset frequency modulation reserve range by a power limiting stage. When the unlimited frequency response power exceeds the frequency modulation reserve upper limit, the actual frequency response power is taken from the upper limit; otherwise, the actual frequency response power equals the unlimited frequency response power. The limited actual frequency response power is input to the system power balance stage to characterize the dynamic characteristics of the grid-connected converter power supply gradually providing frequency support under the influence of frequency measurement and control delay.

[0066] 3. Based on the power limiting components of the grid-connected and grid-based converter power supplies, and considering the differences between them in frequency sensing methods, inertial response speed, and dynamic hysteresis characteristics (grid-based converter power supplies are instantaneous, hysteresis-free, and actively supported, while grid-connected converter power supplies are delayed, gradual, and passively responsive), construct a system frequency response model, such as... Figure 2 As shown in the figure, the grid-type converter power supply can directly respond to system frequency changes without relying on a phase-locked loop (PLL) and can provide inertia support at the initial moment of a disturbance; the grid-type converter power supply needs to obtain the system frequency through a PLL, and its frequency response is affected by filtering characteristics and delay characteristics, thus exhibiting a certain dynamic lag.

[0067] The constructed system frequency response model comprehensively considers the primary frequency regulation response of thermal power units, the inertial response and primary frequency regulation response of grid-connected energy storage, the inertial response and primary frequency regulation response of grid-connected wind farms, and the inertial response and primary frequency regulation response of grid-connected wind farms. Among them, the inertial response process of grid-connected wind farms reflects the filtering and delay effects brought by phase-locked loops through the combination of the aforementioned first-order inertial element and pure delay element. At the same time, power limiting constraints are introduced in the response models of each type of resource to limit their maximum output or absorbed power and ensure that the response process conforms to the equipment operating boundary.

[0068] exist Figure 2 middle, The disturbance power; This is for frequency deviation; , , , These are the installed capacity percentages of thermal power units (g), grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, respectively. , , , These are the droop coefficients for thermal power units (g), grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, respectively. , , , These are the turbine time constant, governor time constant, reheater time constant, and high-pressure cylinder power share of thermal power unit g, respectively. For the mechanical inertia of the system; The system damping coefficient; , , These are the inertia coefficients for grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, respectively. , , These are the time constants of the first-order inertial elements for primary frequency regulation in grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, respectively. The time constant of the first-order inertial element in the inertial response of a grid-type wind farm; The measurement inherent time delay is due to the grid-connected wind farm.

[0069] Specifically, step 2 includes the following:

[0070] 1. The process for determining the power limit of the initial inertia response of grid-connected energy storage and grid-connected wind farms is as follows:

[0071] The expression for the system's swing equation is:

[0072]

[0073] In the formula, The rate of change of frequency; For thermal power unit indexes and sets; , , , denoted as g for thermal power unit, grid-connected energy storage, grid-connected wind farm, and grid-connected wind farm, respectively; t is the frequency response time.

[0074] At the initial moment of the power disturbance, i.e., t=0, the frequency deviation is 0. The thermal power unit and the grid-connected converter power supply will undergo an instantaneous inertial response. The system swing equation at the initial moment of the disturbance can be expressed as:

[0075]

[0076] Limited by the frequency regulation reserves of grid-connected energy storage and grid-connected wind farms, the initial inertial response power of grid-connected energy storage and grid-connected wind farms can be expressed as:

[0077]

[0078]

[0079] Due to the presence of min() in formulas (3) and (4), (3) and (4) are nonlinear. Therefore, the above formulas will be linearized next.

[0080] 2. The power limiting constraints for the initial inertia response of grid-connected wind farms and grid-connected energy storage are derived as follows:

[0081] Assuming that the inertial response power of each grid-type energy storage is unlimited, we can obtain:

[0082]

[0083] In the formula, The initial inertial response power of the grid-type energy storage without amplitude limiting.

[0084] Combining equation (5) and equation (2), we can eliminate... We can obtain:

[0085]

[0086] Similarly, for grid-type wind farms, the following formula can be obtained:

[0087]

[0088] Thus far, it has been observed It is a weighted sum of binary variables, and Since they are continuous variables, the Big M method can be used for linearization and equivalent transformation. Similarly, `min()` can also be linearized and equivalently transformed using the Big M method. This introduces the decision variables involved in the embodiments of this invention, namely... , , , , , , , , , , , We can obtain the following formula:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107] In the formula, , These are the auxiliary variables that were introduced; This is a binary variable representing the start / stop state of generator g, where 0 indicates shutdown and 1 indicates startup. , These represent the initial inertial response power of the unconstrained grid-type energy storage and the constrained initial inertial response power of the grid-type energy storage, respectively; M is a sufficiently large constant. , These represent the initial inertial response power of an unlimited grid-connected wind farm and the initial inertial response power of a limited grid-connected wind farm, respectively. Let g be the inertia of the generator; , These are frequency regulation backups for grid-connected energy storage and grid-connected wind farms, respectively. , These are binary variables introduced to linearize the min step. In the above equation, (9)-(12) are the linearization of (8), and (13) is the substitution of variables in (6). , , , The formulas after (15)-(18) are linearizations of (14), and (19) is the result of substituting variables into (7). , , , The formulas are as follows: (21) and (22) are linearizations of (20), and (24) and (25) are linearizations of (23).

[0108] 3. Construct an initial frequency change rate constraint based on the mechanical inertia of the system.

[0109] The system's mechanical inertia is provided by the online thermal power unit, that is:

[0110]

[0111] The initial rate of change of the system frequency occurs at the initial moment of the disturbance. Under normal large disturbances, it is the maximum rate of change of the system frequency after the disturbance, and it needs to be constrained, i.e., an initial rate of change constraint is constructed, as shown in the following expression:

[0112]

[0113] In the formula, This is the limit for the rate of change of frequency.

[0114] Specifically, step 3 includes the following:

[0115] 1. The input variable x in the system frequency response model can be represented by a vector using the following formula:

[0116]

[0117] In the formula, It is a frequency regulation backup for thermal power unit g. It is intended as a backup for frequency regulation in grid-connected wind farms.

[0118] An initial sample set is randomly generated based on the boundaries of the input variable values, and the number of samples is... The initial sample set is substituted into the system frequency response model to obtain the actual maximum frequency deviation. Based on the relationship between the actual maximum frequency deviation of each input sample and the preset maximum frequency deviation safety threshold, a corresponding frequency safety category label is assigned to each input sample. Each input sample and its corresponding frequency safety category label are then combined to form a labeled sample set, creating an initial labeled sample set. This initial labeled sample set is used to train a deep neural network classification model A, resulting in a trained deep neural network classification model A. The deep neural network classification model A can be a fully connected feedforward deep neural network classification model, comprising: an input layer for receiving input samples and transmitting them to a hidden layer; at least one hidden layer for processing the input samples using an activation function to obtain hidden layer representation features; and an output layer for receiving the hidden layer representation features and outputting the frequency safety category and category probability corresponding to the input sample. Adjacent network layers are fully connected, and the hidden layers use the ReLU function as the activation function. The output of the deep neural network classification model A is a label and its corresponding probability. The label is set to 1 if the maximum frequency deviation is within the safety threshold and 0 if it is outside the safety threshold.

[0119] 2. Obtain an unlabeled sample set based on the input variables in the system frequency response model. Input the unlabeled sample set into the trained deep neural network classification model to obtain the predicted probability of each unlabeled sample belonging to the frequency-safe and frequency-unsafe categories. Then, generate a training dataset using the Region of Interest (ROI) active sampling method. The process of generating the training dataset using the ROI active sampling method is as follows:

[0120] Based on the boundaries of the input variable values, a larger set of candidate input samples is randomly generated to form an unlabeled sample set. The sample size is The unlabeled sample set only contains the values ​​of the input variables and does not include the corresponding actual maximum frequency deviation. The unlabeled sample set is input into the trained deep neural network classification model A to obtain the predicted probability that each unlabeled sample belongs to the frequency-safe category or the frequency-unsafe category. Specifically, the initial sample set and the unlabeled sample set are not the same sample set. Both are generated based on the same input variable value boundaries, but their uses are different: the initial sample set is used to construct the initial labeled sample set and train the deep neural network classification model; its samples need to have their actual maximum frequency deviation calculated by the system frequency response model and frequency-safe category labels set. The unlabeled sample set is a candidate sample set regenerated after the deep neural network classification model is trained; it only contains the values ​​of the input variables and is used to filter samples near the frequency-safe threshold using the region-of-interest active sampling method.

[0121] A region-of-interest (ROI) active sampling method is employed to determine the sampling probability of each unlabeled sample based on the proximity of the two types of predicted probabilities. The closer the two types of predicted probabilities are, the higher the sampling probability of the corresponding sample. Input samples that meet the frequency safety threshold (near the frequency safety threshold) are preferentially extracted according to the sampling probability to form the unlabeled sample set. The formula is as follows:

[0122]

[0123] In the formula, s is the number of samples obtained; a is the active sampling function; Indicates the unlabeled sample set Find the element that maximizes the subsequent expression; k is the category index, which takes the value 0 or 1, where 0 indicates that the maximum frequency deviation is within the safe threshold and 1 indicates that it is not within the safe threshold. In a given sample Under the given conditions, the deep neural network classification model A predicts the posterior probability that the sample belongs to class k.

[0124] Substitute each input sample in the sample set to be labeled one by one. Figure 2 The system frequency response model shown is used to solve for the actual maximum frequency deviation corresponding to each input sample. Each input sample and its corresponding actual maximum frequency deviation are then combined to form a labeled sample, which constitutes the training dataset.

[0125] Specifically, step 4 includes the following:

[0126] 1. To improve the prediction accuracy of the network model near the frequency safety threshold and reduce the frequency safety risk caused by the underestimation of the maximum frequency deviation, a composite loss function that integrates the weighted mean square error term and the conservative penalty term is constructed. The process is as follows:

[0127] A Gaussian function is used to assign higher weights to samples near the frequency safety threshold. The weights w are calculated as follows:

[0128]

[0129] Where Q is the weight magnitude coefficient; c is the maximum frequency deviation critical value; σ is the Gaussian function width coefficient; and K is the true output value of the sample. A basic weighted mean square error loss term is constructed based on the weight w:

[0130]

[0131] Where Y is the model The predicted output value.

[0132] To ensure forecast conservatism, a ReLU quadratic penalty term is applied to the bias where Y > K:

[0133]

[0134] In the formula, This is the penalty coefficient.

[0135] The resulting composite loss function is shown below:

[0136]

[0137] In the formula, Let represent the composite loss function, and n represent the index of the nth sample in the training dataset. This indicates the number of samples obtained. This represents the weight of the nth sample. This represents the predicted output value. This represents the true output value of the nth sample. Indicates the penalty coefficient. This represents the ReLU function.

[0138] 2. Using the training dataset to fit the model to a multi-input single-output deep neural network. Training is performed to obtain a trained multi-input single-output deep neural network fitting model, wherein the multi-input single-output deep neural network fitting model... A fully connected feedforward deep neural network is adopted, including an input layer, at least one hidden layer and an output layer. Adjacent network layers are fully connected. The hidden layer uses the ReLU function as the activation function, and the output layer uses a linear activation function and sets an output neuron to output the maximum frequency deviation.

[0139] 3. Fit the trained multi-input single-output deep neural network model using the Big M method. The constraints are transformed into mixed-integer linear constraints that can be handled by the solver, as follows:

[0140] 3.1 The trained multi-input single-output deep neural network fitting model is decomposed into the linear transformation relationship of each network layer and the ReLU activation operation relationship of each hidden layer.

[0141] The general form of an L-layer deep neural network with ReLU activation is given, consisting of stacked linear transformations and ReLU activation:

[0142]

[0143]

[0144] In the formula, This represents the output of the i-th neuron in the l-th layer. For the i-th neuron in the l-th layer and the i-th neuron The connection weights of the j-th neuron in the layer; This is the bias of the i-th neuron in the l-th layer; Let be the set of indices of all neurons in layer l; y is the final output of the neural network. , The bias of the output layer L and the bias of the Lth layer and the Lth layer are respectively. The connection weights of the j-th neuron in the layer.

[0145] 3.2 Set binary variables to represent the activation state of each ReLU neuron, and use the Big M method to construct linear inequality constraints to equivalently represent the ReLU activation operation relationship.

[0146] For the ReLU function (ReLU(x) = max{0,x}), the Big M method is used to transform it into a linear constraint:

[0147]

[0148]

[0149]

[0150]

[0151]

[0152] Based on the above constraints, the i-th neuron in hidden layer l can be represented as:

[0153]

[0154] In the formula, The linear transformation output of the i-th neuron in the l-th layer; This is a binary indicator variable corresponding to ReLU activation.

[0155] 3.3. Linear equations are used to represent the linear transformation relationships of each network layer. The trained multi-input single-output deep neural network fitting model is represented as a mixed integer linear constraint composed of continuous variables, binary variables, linear equations and linear inequalities.

[0156] The output layer L does not have a ReLU activation function, only retaining the linear transformation shown in equation (35). After the ReLU activation functions of each hidden layer are equivalently transformed into mixed-integer linear constraints using the Big M method through equations (36) to (41), the entire multi-input single-output deep neural network fitting model is represented as a set of mixed-integer linear constraints consisting of continuous variables, binary variables, linear equations, and linear inequalities. The output value y of the multi-input single-output deep neural network is a continuous variable in the mixed-integer linear constraints, used to represent the multi-input single-output deep neural network fitting model. The maximum frequency deviation obtained from the fitting is used to construct a maximum frequency deviation constraint by limiting the output y to not exceed a preset maximum frequency deviation safety threshold, and the constraint is then embedded into the basic optimization scheduling model.

[0157] The feature vectors of each scheduling period, i.e. the feature vectors shown in equation (28), are used to fit the trained multi-input single-output deep neural network model according to equations (36)-(41). Transforming it into a mixed-integer linear constraint form yields the maximum frequency deviation constraint after the anticipated power disturbance in each scheduling period. The maximum frequency deviation constraint is expressed as follows:

[0158]

[0159] in, This is the value for the maximum allowable frequency deviation.

[0160] Specifically, the deep neural network classification model and the multi-input single-output deep neural network fitting model proposed in this embodiment of the invention both include an input layer, at least one hidden layer and an output layer, with fully connected layers between adjacent network layers, and the hidden layer using the ReLU function as the activation function.

[0161] For the deep neural network classification model, the number of neurons in its input layer is the same as the dimension of the input feature vector, and the output layer is used to output the frequency safety category and category probability corresponding to the input sample, so as to filter the input samples near the frequency safety threshold according to the category probability.

[0162] For the multi-input single-output deep neural network fitting model, the number of neurons in the input layer is the same as the dimension of the feature vector shown in equation (28). The output layer uses a linear activation function and has one output neuron to output the maximum frequency deviation, thereby characterizing the nonlinear relationship between the input features and the maximum frequency deviation. As an example, the multi-input single-output deep neural network fitting model has two hidden layers. The first hidden layer and the second hidden layer include 32 neurons and 16 neurons, respectively. The number of hidden layers and the number of neurons in each hidden layer can be adjusted according to the training sample size and fitting accuracy requirements, and are not limited to the specific values ​​mentioned above.

[0163] Specifically, step 5 includes the following:

[0164] 1. Utilize the property that the steady-state frequency change rate is zero to construct the swing equation of the steady-state system.

[0165] When the frequency modulation reaches steady state, the rate of frequency change of the system is 0. Therefore, the system swing equation in steady state can be expressed in the form of constraints as follows:

[0166]

[0167] In the formula, Let t be the frequency deviation when the frequency modulation reaches steady state after the expected disturbance; , , , These represent the steady-state response power of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms after the anticipated disturbance at time t.

[0168] 2. Based on the primary frequency regulation characteristics of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-linked wind farms, the unlimited steady-state frequency response power of each frequency response resource is determined as follows:

[0169]

[0170]

[0171]

[0172]

[0173] 3. Combining the power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-linked wind farms, auxiliary variables are introduced to establish the correlation constraints between the unlimited steady-state frequency response power and the actual steady-state frequency response power, as detailed below:

[0174]

[0175]

[0176]

[0177]

[0178] In the formula, , , , The response power of unlimited thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms are respectively. , , , These are auxiliary variables that act on the power limiting links of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, respectively. All of them are non-negative numbers. The above auxiliary variables correspond to thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, respectively, and are used to represent the portion of the unconstrained steady-state response power of each resource that exceeds its actual steady-state response power.

[0179] 4. Based on the frequency regulation reserves that thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-linked wind farms can provide, the actual steady-state frequency response power is limited, and the following constraints are established:

[0180]

[0181]

[0182]

[0183]

[0184] 5. Construct steady-state frequency deviation constraints, the expression of which is as follows:

[0185]

[0186] In the formula, This is the steady-state frequency deviation limit.

[0187] Specifically, step 6 includes the following:

[0188] The basic optimization scheduling model consists of the optimization scheduling objective function and the general operating constraints of each frequency response resource. Specifically, Equations (57) to (60) are the optimization scheduling objective functions, used to minimize the operating cost of thermal power units, the frequency regulation reserve cost of various resources, and related penalty costs; Equations (61) to (78) are the general operating constraints, including thermal power unit start-up and shutdown, output upper and lower limits, frequency regulation reserve and ramping constraints, energy storage charging and discharging power, charging and discharging state and energy state constraints, and wind farm output, frequency regulation reserve and system power balance constraints.

[0189] The objective function for optimizing scheduling is to minimize the reserve cost, operating cost, and penalty cost of the optimized scheduling, as expressed below:

[0190]

[0191]

[0192]

[0193]

[0194] In the formula, , , These are the start-up variables, shutdown variables, and power generation of thermal power unit g, respectively. , , Calculate the penalty cost for operating costs, reserve costs, and steady-state frequency deviation, respectively. , , , These are the start-up cost, shutdown cost, no-load cost, and unit power cost of thermal power unit g, respectively. , , , denoted as g for thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-following wind farms, respectively; m is a sufficiently small constant.

[0195] The general constraints of the basic optimization scheduling model are as follows:

[0196]

[0197]

[0198]

[0199]

[0200]

[0201]

[0202]

[0203]

[0204]

[0205]

[0206]

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214] In the formula, , These are the maximum and minimum power of thermal power unit g, respectively; This is the maximum allowable frequency regulation reserve for thermal power units. , These are the upward and downward ramp rate limits for thermal power unit g, respectively. , These represent the charging and discharging power of grid-type energy storage; , These represent the maximum charge / discharge rates of grid-type energy storage. This is a binary variable for charging and discharging energy storage. A value of 1 indicates charging, and a value of 0 indicates discharging. The energy stored for energy storage; , The charging and discharging efficiency of energy storage; The scheduling time step; , The maximum and minimum energy stored for energy storage; , These are the maximum power tracking points for grid-type wind farms and grid-following wind farms, respectively. , These are the actual power outputs of grid-type wind farms and integrated grid-type wind farms, respectively. This represents the maximum allowable load shedding rate for the wind farm. Let be the load power of node d during time period t. In the above formula, equation (61) is the unit start-stop state transition constraint; equation (62) is the unit minimum continuous operating time constraint; equation (63) is the unit minimum continuous shutdown time constraint; equation (64) is the unit maximum output and frequency regulation reserve capacity constraint; equation (65) is the unit minimum stable output constraint; equation (66) is the unit frequency regulation reserve capacity upper and lower limit constraint; equation (67) is the unit ramp / downhill rate constraint; equation (68) is the energy storage charging power constraint; equation (69) is the energy storage discharging power constraint; equation (70) is... Equation (71) is the constraint for the dynamic change of the state of charge of energy storage; Equation (72) is the constraint for the safe operation of the state of charge of energy storage; Equation (73) is the constraint for the consistency of the state of charge of energy storage during daily cycles; Equation (74) is the constraint for the frequency regulation reserve capacity of energy storage; Equation (75) is the constraint for the frequency regulation reserve capacity of grid-connected wind farms; Equation (76) is the constraint for the frequency regulation reserve capacity of grid-connected wind farms; Equation (77) is the constraint for the frequency regulation reserve capacity of grid-connected wind farms; Equation (78) is the constraint for the active power balance of the system.

[0215] Based on the constructed basic optimization scheduling model, the initial frequency change rate constraint, maximum frequency deviation constraint, and steady-state frequency deviation constraint are added to the basic optimization scheduling model to form a frequency safety constraint optimization scheduling model. After the above linearization process, the frequency safety constraint optimization scheduling model is transformed into a mixed integer linear programming problem, which can be solved by calling commercial mixed integer linear programming solvers such as the Gurobi solver. The optimized scheduling results are composed of the start-up and shutdown status and active power output of thermal power units, frequency regulation reserve of each frequency response resource, energy storage charging and discharging power and energy status, and active power output of wind farms in each scheduling period.

[0216] Specifically, to demonstrate the effectiveness of this invention in ensuring frequency stability, an IEEE 30-node transmission network comprising six thermal power units, loads, a grid-controlled energy storage power station, a grid-controlled wind farm, and a grid-following wind farm is used as an embodiment of this invention. The model and method proposed in this invention are analyzed, and the topology of the IEEE 30-node network is as follows: Figure 3 As shown, Figure 3 The numbers 1-30 represent node numbers. The safety thresholds for the system's rate of frequency change, maximum frequency deviation, and steady-state frequency deviation are -0.5Hz / s, -0.5Hz, and -0.2Hz, respectively. The system is solved using the gurobi solver on the MATLAB platform. The obtained values ​​for the system's rate of frequency change, maximum frequency deviation, and steady-state frequency deviation after power disturbances during each scheduling period are as follows: Figure 4 As shown, due to the application of frequency security constraints, all frequency indicators are within the safe range, thus proving the effectiveness of the present invention.

[0217] The implementation of the various embodiments of the present invention is based on programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a wind-storage-thermal frequency safety constraint optimization scheduling system that considers converter power supply power limitation. This system is used to execute a wind-storage-thermal frequency safety constraint optimization scheduling method that considers converter power supply power limitation from the above method embodiments.

[0218] The system includes: a first processing module, used to construct a system frequency response model based on the power limiting elements of the grid-connected converter power supply and the power supply of the grid-connected converter, combined with the differences in frequency response characteristics between the two power supplies; a second processing module, used to derive the power limiting constraints of the initial inertia responses of the grid-connected wind farm and the grid-connected energy storage based on the power limiting of the initial inertia responses of the grid-connected energy storage and the grid-connected wind farm, and construct initial frequency change rate constraints in combination with the mechanical inertia of the system; a third processing module, used to obtain an unlabeled sample set based on the input variables in the system frequency response model, input the unlabeled sample set into a trained deep neural network classification model, obtain the predicted probability of each unlabeled sample belonging to the frequency safe category and the frequency unsafe category, and generate a training dataset using a region of interest active sampling method; and a fourth processing module, used to combine the constructed composite loss The first processing module trains a multi-input single-output (MIMO) deep neural network fitting model using a training dataset. This trained model is then transformed into a mixed-integer linear constraint using the Big M method, and a maximum frequency deviation constraint is constructed. The second processing module combines the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms. It introduces auxiliary variables to establish the correlation constraint between steady-state frequency response and steady-state frequency deviation, and constructs a steady-state frequency deviation constraint. The third processing module constructs a basic optimization scheduling model, combines it with frequency security constraints to form a frequency security constraint optimization scheduling model, and transforms it into a mixed-integer linear programming problem. Solving this problem yields the optimization scheduling result. The frequency security constraints include initial frequency change rate constraints, maximum frequency deviation constraints, and steady-state frequency deviation constraints.

[0219] This invention provides a wind-storage-thermal frequency safety constraint optimization scheduling system considering converter power limit. Addressing the issues of insufficient system inertia support and limited frequency safety under high-proportion renewable energy grid connection, this method offers a wind-storage-thermal frequency safety constraint optimization scheduling approach. First, considering the power limit stage, a system frequency response model is constructed by combining the frequency response characteristics differences between grid-connected and grid-connected converter power supplies. Second, considering the power limit of the initial inertia response of grid-connected energy storage and grid-connected wind farms, the power limit constraint of the multi-source inertia response is derived, and an initial frequency change rate constraint is constructed. Next, the ReLU function is selected as the activation function, and a loss function that balances conservatism with high fitting accuracy near the maximum frequency deviation safety threshold is designed. A training dataset is generated using a region-of-interest active sampling method, and a deep neural network is used to achieve accurate fitting of the maximum frequency deviation. The maximum frequency deviation is then transformed into a mixed-integer linear constraint using the Big M method. Subsequently, a steady-state frequency deviation constraint is constructed by introducing auxiliary variables. Finally, a basic optimization scheduling model is constructed and solved in conjunction with the aforementioned frequency safety constraints to obtain an optimized scheduling result that balances economy and frequency safety.

[0220] Finally, it should be noted that the above specific embodiments are merely representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and many variations are possible. Any simple modifications, equivalent changes, and alterations made to the above specific embodiments based on the technical essence of the present invention should be considered within the protection scope of the present invention.

Claims

1. A wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply power limitation, characterized in that, include: Based on the power limiting elements of the grid-connected converter power supply and the power limiting elements of the grid-connected converter power supply, and combined with the differences in the frequency response characteristics of the grid-connected converter power supply and the power limiting elements of the grid-connected converter power supply, a system frequency response model is constructed. Based on the power limiting of the initial inertial response of grid-connected energy storage and grid-connected wind farm, the power limiting constraints of the initial inertial response of grid-connected wind farm and grid-connected energy storage are derived. Combined with the mechanical inertia of the system, the initial frequency change rate constraint is constructed. The unlabeled sample set is obtained based on the input variables in the system frequency response model. The unlabeled sample set is then input into the trained deep neural network classification model to obtain the predicted probability that each unlabeled sample belongs to the frequency safe category and the frequency unsafe category. The region of interest active sampling method is then used to generate the training dataset. By combining the constructed composite loss function, the multi-input single-output deep neural network fitting model is trained using the training dataset to obtain the trained multi-input single-output deep neural network fitting model. The trained multi-input single-output deep neural network fitting model is then transformed into mixed integer linear constraints using the Big M method, and the maximum frequency deviation constraint is constructed. Combining the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms and grid-connected wind farms, auxiliary variables are introduced to establish the correlation constraints between steady-state frequency response and steady-state frequency deviation, and steady-state frequency deviation constraints are constructed. A basic optimization scheduling model is constructed, and a frequency security constraint optimization scheduling model is formed by combining it with frequency security constraints. This model is then transformed into a mixed integer linear programming problem, and the optimization scheduling result is obtained by solving the problem. The frequency security constraints include initial frequency change rate constraints, maximum frequency deviation constraints, and steady-state frequency deviation constraints.

2. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, The power limiting circuitry for grid-connected converter power supplies includes: The unlimited frequency response power of the grid-type converter power supply is limited to its preset frequency regulation reserve range: when the unlimited frequency response power does not exceed the frequency regulation reserve limit, the actual frequency response power is equal to the unlimited frequency response power; when the unlimited frequency response power exceeds the frequency regulation reserve limit, the actual frequency response power is equal to the frequency regulation reserve limit. The unlimited frequency response power of the grid converter power supply is limited to a preset frequency regulation reserve range: when the unlimited frequency response power exceeds the frequency regulation reserve limit, the actual frequency response power is equal to the frequency regulation reserve limit; when the unlimited frequency response power does not exceed the frequency regulation reserve limit, the actual frequency response power is equal to the unlimited frequency response power.

3. The wind-storage-thermal frequency safety constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, The expression for the initial frequency change rate constraint is: , in, Let represent the disturbance power at time t. This represents the initial inertia response power of an unlimited grid-connected wind farm. This represents the initial inertial response power of an unconstrained grid-type energy storage system. Represents the mechanical inertia of the system. This indicates the limit of the rate of change of frequency.

4. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, The training dataset is generated using a region-of-interest active sampling method, including: The sampling probability of each unlabeled sample is determined based on the predicted probability of each unlabeled sample belonging to the frequency safe category and the frequency unsafe category. Input samples that meet the frequency safe threshold requirements are then extracted from high to low according to the sampling probability to form a sample set to be labeled. Each input sample in the unlabeled sample set is substituted into the system frequency response model one by one to obtain the corresponding actual maximum frequency deviation. Each input sample and the corresponding actual maximum frequency deviation are then combined to form a labeled sample, which constitutes the training dataset.

5. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, The training of the deep neural network classification model includes: An initial sample set is randomly generated based on the boundary values ​​of the input variables in the system frequency response model. Each sample in the initial sample set is substituted into the system frequency response model to solve for the actual maximum frequency deviation. A frequency safety category label is set based on the relationship between the actual maximum frequency deviation and the preset maximum frequency deviation safety threshold to form an initial labeled sample set. A deep neural network classification model is constructed, including an input layer for receiving input samples and transmitting them to a hidden layer; at least one hidden layer for processing the input samples using an activation function to obtain hidden layer representation features; and an output layer for receiving the hidden layer representation features and outputting the frequency-safe category and category probability corresponding to the input sample; adjacent network layers are fully connected. The deep neural network classification model is trained using the initial labeled sample set to obtain the trained deep neural network classification model.

6. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, The Big M method is used to transform the trained multi-input single-output deep neural network fitting model into a mixed-integer linear constraint, including: The trained multi-input single-output deep neural network fitting model is decomposed into the linear transformation relationship of each network layer and the ReLU activation operation relationship of each hidden layer. Binary variables are set to represent the activation state of each ReLU neuron, and the Big M method is used to construct linear inequality constraints to represent the ReLU activation operation relationship; The linear transformation relationship of each network layer is represented by a linear equation. The trained multi-input single-output deep neural network fitting model is represented as a mixed integer linear constraint consisting of continuous variables, binary variables, linear equations and linear inequalities. The output value of the multi-input single-output deep neural network fitting model is a continuous variable.

7. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, Constructing steady-state frequency deviation constraints includes: By utilizing the characteristic that the rate of frequency change is zero in steady state, the system swing equation in steady state is constructed; Combining the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms, and grid-connected wind farms, auxiliary variables are introduced to characterize the difference between the unconstrained steady-state frequency response power and the actual steady-state frequency response power. Correlation constraints between steady-state frequency response and steady-state frequency deviation are established, and steady-state frequency deviation constraints are constructed.

8. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, Construct a basic optimized scheduling model, including: The objective function for optimizing scheduling is to minimize the reserve cost, operating cost, and penalty cost of the optimized scheduling. A basic optimized scheduling model is constructed by combining the general operating constraints of each frequency response resource. The general operating constraints include the start-up and shutdown of thermal power units, upper and lower limits of output, frequency regulation reserve and ramping constraints, energy storage charging and discharging power, charging and discharging state and energy state constraints, and wind farm output, frequency regulation reserve and system power balance constraints.

9. The wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in claim 1, characterized in that, The constructed composite loss function is expressed as follows: , in, Let represent the composite loss function, and n represent the index of the nth sample in the training dataset. This indicates the number of samples obtained. This represents the weight of the nth sample. This represents the predicted output value. This represents the true output value of the nth sample. Indicates the penalty coefficient. This represents the ReLU function.

10. A wind-storage-thermal frequency safety constraint optimization scheduling system considering converter power supply power limitation, characterized in that, A wind-storage-thermal frequency security constraint optimization scheduling method considering converter power supply limitation as described in any one of claims 1-9 includes: The first processing module is used to construct a system frequency response model based on the power limiting circuits of the grid-connected converter power supply and the power limiting circuits of the grid-connected converter power supply, combined with the differences in frequency response characteristics between the grid-connected converter power supply and the power limiting circuits of the grid-connected converter power supply. The second processing module is used to derive the power limiting constraints of the initial inertia responses of grid-connected wind farms and grid-connected energy storage based on the power limiting of the initial inertia responses of grid-connected energy storage and grid-connected wind farms, and to construct the initial frequency change rate constraint in combination with the mechanical inertia of the system. The third processing module is used to obtain an unlabeled sample set based on the input variables in the system frequency response model, input the unlabeled sample set into the trained deep neural network classification model, obtain the predicted probability of each unlabeled sample belonging to the frequency safe category and the frequency unsafe category, and generate a training dataset using the region interest active sampling method. The fourth processing module is used to combine the constructed composite loss function and train the multi-input single-output deep neural network fitting model using the training dataset to obtain the trained multi-input single-output deep neural network fitting model. The trained multi-input single-output deep neural network fitting model is then transformed into mixed integer linear constraints using the Big M method, and the maximum frequency deviation constraint is constructed. The fifth processing module combines the primary frequency regulation characteristics and power limiting of thermal power units, grid-connected energy storage, grid-connected wind farms and grid-connected wind farms, introduces auxiliary variables to establish the correlation constraints between steady-state frequency response and steady-state frequency deviation, and constructs steady-state frequency deviation constraints. The sixth processing module is used to construct a basic optimization scheduling model, combine it with frequency security constraints to form a frequency security constraint optimization scheduling model, and transform it into a mixed integer linear programming problem to obtain the optimization scheduling result. The frequency security constraints include initial frequency change rate constraints, maximum frequency deviation constraints, and steady-state frequency deviation constraints.

Images