Low Voltage Ride-Through Control Method for Wind Farms Based on Reactive Power Guidance and Constraint Priority Playback
By optimizing the low-voltage ride-through control of wind farms using a deep reinforcement learning strategy model, the contradiction between reactive power demand and active power retention is resolved, achieving the effectiveness and stability of reactive power allocation across the turbine cluster and meeting physical constraints.
Patent Information
- Application Number
- CN202611320690.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
During low-voltage ride-through, existing wind farms struggle to simultaneously address reactive power demand, turbine group differences, and active power maintenance. This results in insufficient reactive power allocation for some turbine groups and increases the complexity of states and actions by directly controlling individual turbines.
A low-voltage ride-through control method for wind farms, employing reactive power guidance and constraint priority playback, is proposed. This method generates active power regulation and reactive power allocation actions of the turbine cluster through a deep reinforcement learning strategy model. Combined with the reactive power capacity of the turbine cluster and the reactive power sensitivity of the PCC voltage, the active power control commands and reactive power control commands are optimized.
It effectively reduces the dimensionality of state and action space, ensures that reactive power demand corresponds to fault state, reduces reactive power deficit and excess, improves control stability, and meets physical constraints.
Smart Images

Figure CN123121679A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power systems, specifically relating to a low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback. Background Technology
[0002] As the scale of wind power grid connection expands, wind farms need to maintain grid-connected operation and provide reactive power support during grid voltage dips. Existing fixed-ratio, equal-capacity ratio, or single-sensitivity allocation methods struggle to simultaneously address reactive power demand, turbine cluster differences, and active power maintenance. Furthermore, some turbine clusters may experience surplus reactive power that cannot be fully allocated after reaching their capacity limits. Directly controlling individual turbines increases the dimensionality of states and actions. Deep reinforcement learning can be used to generate continuous control actions, but conventional experience replay cannot prioritize key samples such as reactive power deficits, reactive power surpluses, and active power declines. Therefore, a low-voltage ride-through control method for wind farms is needed, combining reactive power demand guidance, constraint violation priority experience replay, and coordinated active and reactive power control across the turbine cluster. Summary of the Invention
[0003] To address the aforementioned problems, this invention provides a low-voltage ride-through control method for wind farms that prioritizes reactive power guidance and constraint playback, resolving the conflict between reactive power support and active power maintenance during low-voltage ride-through. This method groups wind turbines based on initial operating information and network topology information, determines effective reactive power demand using the PCC fault voltage under conditions without additional reactive power support, generates active power adjustment actions, reactive power allocation actions, and total reactive power correction actions for the turbine cluster through a trained deep reinforcement learning strategy model, and combines the cluster's reactive power capacity, PCC voltage reactive power sensitivity, and capacity circle constraints to form the final active power control command and the final reactive power control command.
[0004] The technical solution adopted in this invention is as follows: A low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback, comprising the following steps:
[0005] S1: Obtain initial operating information and network topology information of wind turbine units, group wind turbine units according to initial active power, electrical distance to the grid connection point PCC and initial terminal voltage, and determine the operating status, available active power, reactive power and PCC voltage reactive sensitivity of each group.
[0006] S2: Construct the state space, action space, and training scenario for low voltage ride-through control of wind farms. Determine the priority of experience samples based on the temporal difference error, constraint violation degree, and scenario difficulty of the experience samples, and train the deep reinforcement learning strategy model based on the priority experience playback.
[0007] S3: After detecting that the wind farm has entered the low voltage ride-through state, acquire the PCC voltage, wind farm power and the operating status of each turbine group, determine the effective reactive power demand of the wind farm based on the PCC fault voltage under the condition of no additional reactive power support, and input the relevant status into the trained deep reinforcement learning strategy model.
[0008] S4: Generate active power instructions for each turbine group and total reactive power target for the wind farm based on the output of the deep reinforcement learning strategy model. Combine the reactive power capacity of the turbine group, the reactive power sensitivity of the PCC voltage and the reactive power allocation action to determine the reactive power allocation ratio of the turbine group, and redistribute the remaining reactive power for the turbine group that has reached the upper limit of reactive power capacity.
[0009] S5: Based on the power boundary and capacity circle constraints of each generator group, perform safe projection of active and reactive power commands to obtain the final active and reactive power control commands for each generator group, and issue them to the corresponding generator group.
[0010] Furthermore, in step S1, the first The electrical distance from the typhoon generator to the PCC is the sum of the impedance magnitudes of all branches along its connection path:
[0011]
[0012] In the formula: Indicates the first Electrical distance between the typhoon generator and the PCC; Indicates the first The set of access path branches from the typhoon generator to the PCC; Indicates a branch index; Indicates the first The complex impedance of a branch; Indicates the first The impedance magnitude of each branch;
[0013] For the The first typhoon turbine Standardize the class grouping features:
[0014]
[0015] In the formula: Indicates the first Typhoon turbine unit The original values of the class grouping features; This represents the corresponding standardized value; , They represent the first The mean and standard deviation of the class grouping features, and ;
[0016] The clustering characteristics include initial active power, electrical distance, and initial terminal voltage;
[0017] Let the first The set of wind turbine units included in a single turbine cluster is Then the first The equivalent rated apparent power capacity, initial active power, and equivalent initial terminal voltage of each generator group are as follows:
[0018]
[0019]
[0020]
[0021] In the formula: Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates the first The initial active power of each generator group; Indicates the first The equivalent initial terminal voltage of each cluster; , , They represent the first Rated apparent power capacity, initial active power, and initial terminal voltage of typhoon generator units; PCC voltage reactive power sensitivity. By examining the first A reactive power disturbance is applied to each generator group, and the amount of reactive power disturbance is determined by the ratio of the PCC voltage change to the reactive power disturbance.
[0022] Furthermore, in step S2, the deep reinforcement learning policy model adopts a continuous action space, and the state vector of the policy model is represented as:
[0023]
[0024] In the formula: Represents the state vector of the policy model; This represents the actual PCC voltage per unit value after applying cluster control. Indicates the fault severity label; , These represent the total active power and total reactive power of the wind farm, respectively, after being normalized using the system power reference value. Indicates effective reactive power demand; Indicates the first Individual cluster status; Indicates the number of aircraft in the fleet;
[0025] No. The status of a cluster is represented as follows:
[0026]
[0027] The respective proportional states are as follows:
[0028]
[0029]
[0030]
[0031]
[0032]
[0033] In the formula: Indicates the first Active load rate of individual machine clusters; Indicates the first Reactive load rate of individual generator groups; Indicates the first Terminal voltage of each unit group; Indicates the proportion of remaining reactive power margin; Indicates the available active power ratio; , , They represent the first The current active power, current reactive power, and available active power of each generator group; Indicates the first The equivalent rated apparent power capacity of each unit group; It represents the residual reactive power margin determined by the capacity circle;
[0034] The action vector of the policy model is represented as:
[0035]
[0036] In the formula: Represents the action vector of the policy model; Indicates the first Active power regulation actions of individual generator groups; Indicates the first Reactive power distribution actions of individual generator groups; This represents the total reactive power correction action; each action is a continuous variable and is restricted to... .
[0037] Furthermore, in step S2, regarding the first... Based on empirical samples, the proportions of reactive power deficit, reactive power surplus, and active power reduction were calculated respectively:
[0038]
[0039]
[0040]
[0041] In the formula: , , They represent the first The proportions of reactive power deficit, reactive power surplus, and active power decline corresponding to each empirical sample; Indicates the first Effective reactive power demand corresponding to one empirical sample; This indicates the actual total reactive power output of the wind farm after control. This represents the total active power of the wind farm before the fault. This represents the actual total active power of the wind farm after control. This represents the allowable reactive power support margin coefficient, which is not less than 1. This indicates a positive number that prevents the denominator from being zero;
[0042] After deducting the corresponding numerical tolerance or scenario soft threshold, the effective constraint violation amount is obtained. Let the first... The effective reactive power deficit, effective reactive power surplus, and effective reactive power decline for each sample are as follows: , and The expression is as follows:
[0043]
[0044]
[0045]
[0046] In the formula, , , They represent the first The corresponding proportions of reactive power deficit, reactive power surplus, and active power reduction for each empirical sample. , These represent the numerical tolerances for reactive power deficit and reactive power surplus, respectively. This indicates the soft threshold for the allowable decrease in active power in the current wind speed scenario;
[0047] Constraint violations of the original scores are determined in stratification according to the following formula:
[0048]
[0049] The constraint violation score is obtained by using a monotonic compression function:
[0050]
[0051] In the formula: Indicates the first Constraint violation score after compression of an empirical sample; Indicates the first The constraints of the empirical sample violate the original score; This represents a compression scale parameter that is greater than zero.
[0052] Scene difficulty score Determined by both wind speed level and fault intensity label, under the same conditions, the higher the wind speed level and the deeper the PCC voltage drop represented by the fault intensity label, the higher the corresponding scenario difficulty score.
[0053] No. The priority and sampling probability of each empirical sample are as follows:
[0054]
[0055]
[0056] In the formula: Indicates the first Priority of each empirical sample; This represents the temporal difference error of the empirical sample; This indicates that the constraint violates the fraction; Indicates the difficulty score of the scene; This represents the non-negative weight used to adjust the contribution of time-series difference error to the priority of empirical samples; This represents the non-negative weight used to adjust the contribution of constraint violation scores to the priority of empirical samples; This represents the non-negative weight used to adjust the contribution of scene difficulty scores to the priority of experience samples; This represents a constant that guarantees a positive priority. Indicates the first The sampling probability of an empirical sample; Indicates as training progresses Adjusted priority sampling index; This indicates the current number of samples in the experience replay pool; Represents the index for summing empirical samples; where, The training progresses and the difficulty gradually decreases. , and The sampling distribution is adjusted according to the training process, so that the sampling probability of key constraint samples and difficult scenario samples is gradually increased in the early stage of training.
[0057] Furthermore, in step S3, the PCC fault voltage under the condition of no additional reactive power support is assumed to be... The reactive current demand is The maximum reactive current is The direct proportionality coefficient between voltage drop and reactive current is: The preset reactive power support start-up voltage threshold is The preset depth low voltage threshold is The reactive current demand is then expressed as:
[0058]
[0059] Determine the effective reactive power demand based on the aforementioned reactive current demand:
[0060]
[0061] In the formula: This indicates the effective reactive power demand of a wind farm; This represents the reactive power margin factor that is greater than zero. This indicates the PCC fault voltage under conditions of no additional reactive power support. Indicates reactive current demand; Indicates the maximum reactive current; This represents the proportionality coefficient between voltage sag and reactive current. This indicates the preset reactive power support start-up voltage threshold. This indicates the preset deep low voltage threshold.
[0062] Furthermore, in step S4, the first... Active power regulation actions of individual generator groups Transformed according to the linear mapping relationship into the first Active power instructions for individual machine clusters The active power command range of the linear mapping is jointly determined by the pre-fault active power reference value and the current available active power; the active power adjustment actions of each generator group are mapped to the active power range allowed during the fault period, let the first... The lower and upper limits of active power commands for a single cluster are respectively: and The active power instruction is:
[0063]
[0064] In the formula: Indicates the first Active commands generated by individual machine clusters; and These represent the lower and upper limits of active power commands allowed during a fault in the generator group, respectively. Indicates the first The active power regulation actions of each generator group, and The upper and lower limits of the active power command are jointly determined by the pre-fault active power reference value and the current available active power, and satisfy the following conditions: ;
[0065] Determine the total reactive power target for the wind farm based on the effective reactive power demand and total reactive power correction action:
[0066]
[0067] In the formula, This indicates the total reactive power target of the wind farm; Indicates effective reactive power demand; Indicates the total reactive power correction action; This represents the non-negative relative correction range, and The total reactive power target is therefore limited to nearby.
[0068] Furthermore, in step S4, the first The maximum positive reactive power that a single generator group can provide under the current active power command is:
[0069]
[0070] In the formula: Indicates the first The maximum positive reactive power that a single generator group can provide under the current active power command; Indicates the first The upper limit of the absolute value of reactive power allowed for a single generator group; Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates the first Active commands for individual machine groups;
[0071] Based on the forward reactive power of the generator group, the reactive power sensitivity of the PCC voltage, and the reactive power distribution action, an unnormalized reactive power distribution coefficient is constructed:
[0072]
[0073] In the formula: Indicates the first Unnormalized reactive power distribution coefficient of individual generator groups; Indicates the first PCC voltage reactive power sensitivity of individual generator groups; , These represent the non-negative weight indices for the non-functionality term and the sensitivity term, respectively. Indicates the intensity of reactive power distribution adjustment; Indicates the first Reactive power distribution actions of individual generator groups;
[0074] Let the first The set of clusters that still participate in the allocation is The initial round of cumulative reactive power commands for each machine group Initially unassigned reactive power targets Then the first Wheel of Life The normalized reactive power allocation weights for each unsaturated generator group are:
[0075]
[0076] No. Round assigned to the first The reactive power increment of each generator group is:
[0077]
[0078] Complete the first After the round of allocation, update the cumulative reactive power instructions for each cluster and the reactive power target for the next round that has not yet been allocated:
[0079]
[0080]
[0081] In the formula: Indicates the first Wheel of Life Normalized reactive power allocation weights for individual generator groups; Indicates the first The reactive power targets that have not yet been allocated at the start of the round; , They represent the first Before and after the round of allocation Cumulative reactive power command for each generator group; This represents the cluster summation index; when the cluster's cumulative reactive power instruction reaches... When the time comes, the generator group will be removed from the next round of generator group allocation until the remaining reactive power target is less than the preset tolerance or all generator groups reach the upper limit of reactive power capacity. The preset tolerance is a preset small positive number greater than 0.
[0082] Furthermore, in step S5, the first... The active power command of each generator group is limited to between zero and the available active power of that group, resulting in the boundary-limited active power command. ; will the first The reactive power command for each generator group is limited to the absolute value of the allowable reactive power of that generator group, resulting in the reactive power command after boundary limiting. The upper limit of the absolute value of reactive power allowed within the capacity circle is determined based on the active power command after boundary limiting.
[0083]
[0084] By limiting the reactive power command after boundary limiting to the reactive power range allowed by the capacity circle, the final reactive power control command is obtained:
[0085]
[0086] In the formula: Indicates the first Individual clusters of machines receive active power commands after boundary limiting. The upper limit of the absolute value of reactive power is determined by the capacity circle; Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates reactive power command after boundary limiting; This indicates the final reactive power control command after the safety projection; the final active power control command remains at [value missing]. .
[0087] The present invention also provides a wind farm low voltage ride-through control system with reactive power guidance and constraint priority playback, used to implement the wind farm low voltage ride-through control method with reactive power guidance and constraint priority playback as described above, including: a turbine clustering and parameter identification module, a deep reinforcement learning model training module, a fault detection and reactive power demand calculation module, a power command generation and reactive power allocation module, and a power safety projection and command issuance module.
[0088] The wind turbine clustering and parameter identification module is used to acquire the initial operating information and network topology information of the wind turbines, and to group the wind turbines according to the initial active power, the electrical distance to the grid connection point PCC and the initial terminal voltage, and to determine the operating status, available active power, reactive power and PCC voltage reactive sensitivity of each cluster.
[0089] The deep reinforcement learning model training module is used to construct the state space, action space and training scenario of low voltage ride-through control in wind farms. It determines the priority of experience samples based on the temporal difference error, constraint violation degree and scenario difficulty of experience samples, and trains the deep reinforcement learning strategy model based on the priority experience playback.
[0090] The fault detection and reactive power demand calculation module is used to detect whether the wind farm has entered the low voltage ride-through state. When the low voltage ride-through state is detected, the PCC voltage, wind farm power and the operating status of each turbine group are obtained. The effective reactive power demand of the wind farm is determined based on the PCC fault voltage under the condition of no additional reactive power support, and the relevant status is input into the trained deep reinforcement learning strategy model.
[0091] The power command generation and reactive power allocation module is used to receive the output of the deep reinforcement learning strategy model, generate active power commands for each group and the total reactive power target of the wind farm, determine the reactive power allocation ratio of the group by combining the reactive power capacity of the group, the reactive power sensitivity of the PCC voltage and the reactive power allocation action, and redistribute the remaining reactive power of the group that has reached the upper limit of reactive power capacity.
[0092] The power safety projection and command issuance module is used to safely project active and reactive power commands based on the power boundary and capacity circle constraints of each generator group, obtain the final active power control command and the final reactive power control command of each generator group, and issue them to the corresponding generator group.
[0093] The present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described above.
[0094] The present invention has the following beneficial effects:
[0095] 1) By grouping wind turbines, the control of a single wind turbine is transformed into group control, reducing the dimensionality of the state space and action space, while retaining the initial output, electrical distance and terminal voltage differences.
[0096] 2) Effective reactive power demand is determined by the fault power flow voltage under the condition of no additional reactive power support, so that the demand target corresponds to the fault state and network equation.
[0097] 3) The strategy model only makes limited adjustments to the total reactive power demand and learns the cluster allocation relationship, reducing the reactive power deficit and reactive power surplus caused by directly searching all cluster reactive power instructions.
[0098] 4) Reactive power allocation takes into account the reactive power capacity of the generator group, the reactive power sensitivity of the PCC voltage, and the strategy action, and continues to allocate the remaining reactive power after the generator group is saturated.
[0099] 5) Replaying the constraint violation priority experience improves the utilization rate of key constraints and difficult scenario samples, which is conducive to improving control stability under different wind speeds and fault intensities.
[0100] 6) The safety projection maintains the active power command and restricts the reactive power command within the capacity circle to ensure that the final control command meets the physical constraints. Attached Figure Description
[0101] Figure 1 This is the overall flowchart of the wind farm low voltage ride-through active and reactive power coordinated control method of the present invention;
[0102] Figure 2 This is a wind farm topology diagram from the embodiment;
[0103] Figure 3 A comparison chart of rewards for different methods in multiple low-voltage ride-through scenarios;
[0104] Figure 4 A comparison chart of PCC voltage response for representative scenarios;
[0105] Figure 5A comparison chart of active power response for representative scenarios;
[0106] Figure 6 A comparison chart of reactive power response for representative scenarios. Detailed Implementation
[0107] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following description is provided in conjunction with the accompanying drawings and embodiments. These embodiments are used to illustrate the implementation process of this invention and are not intended to limit the scope of protection of this invention.
[0108] Example 1
[0109] like Figure 1 As shown, a low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback includes the following steps:
[0110] S1: Obtain initial operating information and network topology information of wind turbines. Group the wind turbines according to initial active power, electrical distance to the grid connection point (PCC), and initial terminal voltage. Determine the operating status, available active power, reactive power, and PCC voltage reactive power sensitivity of each group. Among these, the first... The electrical distance from the typhoon generator to the PCC is expressed as the sum of the impedance magnitudes of each branch along its connection path:
[0111]
[0112] In the formula, Indicates the first Electrical distance between the typhoon generator and the PCC; This represents the set of branches from the wind turbine to the PCC; Indicates a branch index; Indicates the first The complex impedance of a branch; This indicates its impedance magnitude.
[0113] In a radial network, The only access path from the corresponding wind turbine to the PCC.
[0114] To avoid dimensional differences between initial active power, electrical distance, and terminal voltage, a standardized approach is adopted to handle the first... Class characteristics:
[0115]
[0116] In the formula, and They represent the first Typhoon turbine unit Original and standardized values of class features; and They represent the first The mean and standard deviation of the class features, and .
[0117] The standardized initial active power, electrical distance, and initial terminal voltage are used to form a clustering input vector with preset weights. Clustering aims to reduce the differences in characteristics between wind turbine units within the same turbine group. After clustering, each turbine group is equalized: the group capacity is the sum of the capacities of its constituent wind turbine units, the initial active power of the group is the sum of the initial active power of its constituent wind turbine units, and the terminal voltage of the group is a capacity-weighted average.
[0118] The sensitivity of the PCC voltage to the reactive power of the generator group was obtained through power flow perturbation calculations. For the first... A small reactive power disturbance is applied to each generator group, and the change in PCC voltage before and after the disturbance is recorded. The ratio of the voltage change to the reactive power disturbance is used as the sensitivity.
[0119] S2: Construct the state space, action space, and training scenarios for low-voltage ride-through control in wind farms. Prioritize empirical samples based on their temporal difference error, constraint violation degree, and scenario difficulty. Train a deep reinforcement learning policy model based on the prioritized empirical samples. To reduce redundant variables, fault detection signals are only used to start the controller and not as continuous states in the policy model. Fault intensity labels represent the fault categories in the training scenarios and are used as state inputs to the policy model, while also determining the scenario difficulty score. The actual voltage state is represented by the PCC voltage.
[0120] The state vector of the policy model is represented as:
[0121]
[0122] In the formula, Represents the state vector of the policy model; This represents the actual PCC voltage per unit value after applying cluster control. Indicates the fault severity label; , These represent the total active power and total reactive power of the wind farm, respectively, normalized using the system power reference value. Indicates effective reactive power demand; Indicates the first Individual cluster status; subscript Indicates the number of aircraft in the fleet.
[0123] No. The status of a cluster is represented as follows:
[0124]
[0125] In the formula, Indicates the active power load rate of the generator group; Indicates the reactive load rate of the generator group; Indicates the terminal voltage of the generator set; Indicates the proportion of remaining reactive power margin; This indicates the available active power ratio.
[0126] The respective proportional states are defined as follows:
[0127]
[0128]
[0129]
[0130]
[0131]
[0132] In the formula: , They represent the first Current active power and current reactive power of each generator group; Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates the first The upper limit of the absolute value of reactive power allowed for a single generator group; This represents the remaining reactive power margin obtained after simultaneously considering the upper limit of the absolute value of the allowable reactive power of the generator group and the capacity circle constraint. Indicates the first Available active power of each generator group; , , , These represent the active load rate, reactive load rate, remaining reactive margin ratio, and available active power ratio, respectively.
[0133] Power state divided by This is used to eliminate capacity differences between different generator groups, allowing state variables to represent active load rate, reactive load rate, and power margin ratio, respectively. Since the generator group terminal voltage is already in per-unit value, it is no longer divided by the generator group capacity.
[0134] The action vector is represented as:
[0135]
[0136] In the formula, Represents the action vector of the policy model; Indicates the first Individual generator group active power regulation action; Indicates the first Reactive power distribution actions of individual generator groups; Indicates total reactive power correction action; subscript Indicates the number of aircraft in the group; each action is limited to .
[0137] The immediate rewards consist of benefits for meritorious service and various types of penalties:
[0138]
[0139] In the formula, Indicates the first Instant rewards for each experience sample; This indicates that the effort has been worthwhile and the benefits have been maintained. Indicates PCC voltage deviation penalty; This indicates a penalty for ineffective performance. This indicates a punishment for ineffective or excessive work. This indicates a reduction in punishment for meritorious service; This indicates additional penalties for exceeding capacity limits, changes in actions, and alterations in safety projections. All items are calculated based on control results from the same empirical sample.
[0140] Reactive power deficit and reactive power surplus are respectively:
[0141]
[0142]
[0143] In the formula, Indicates a lack of reactive power; This indicates that there is excess work done without effort; Indicates effective reactive power demand; This indicates the actual total reactive power output of the wind farm after control. This represents the allowable reactive power support margin coefficient, which is not less than 1.
[0144] The percentage decrease in active power is expressed as follows:
[0145]
[0146] In the formula, Indicates the percentage decrease in active power; This indicates that the wind farm had a total active power before the fault occurred; This indicates the actual total active power of the wind farm after control. This indicates a positive number that prevents the denominator from being zero.
[0147] PCC voltage deviation is determined by the difference between the actual PCC voltage and the ideal power flow target voltage. The ideal target voltage is obtained through power flow calculations from a reference command that satisfies effective reactive power demand without violating cluster capacity constraints.
[0148] The constraint violation priority experience replay transforms reactive power deficit, reactive power surplus, and active power decline into constraint violation scores. First, the numerical tolerance or soft threshold of each indicator is deducted, and then the indicators are stratified and combined according to the control priority relationship of reactive power deficit, reactive power surplus, and active power decline.
[0149] When the sample has an effective reactive power deficit, the reactive power deficit is the primary factor; when there is no effective reactive power deficit but there is reactive power surplus, the reactive power surplus is the primary factor; when neither exists, the degree of constraint violation is determined based on the active power decrease exceeding the soft threshold.
[0150] Constraints violate original scores The layers are determined according to the following formula:
[0151]
[0152] In the formula: Indicates the first The constraints of the empirical sample violate the original score; Indicates effective reactive power deficit; This indicates that there is excess effective power consumption; This indicates a decrease in effective work.
[0153] Constraints violating the original score are addressed using a monotonically compressed function:
[0154]
[0155] In the formula: Indicates the first Constraint violation score after compression of an empirical sample; Indicates the first The constraints of the empirical sample violate the original score; This represents a compression scale parameter that is greater than zero. It represents the natural logarithm.
[0156] Scene difficulty score Determined by both wind speed level and fault intensity label, under the same conditions, the higher the wind speed level and the deeper the PCC voltage drop represented by the fault intensity label, the higher the corresponding scenario difficulty score.
[0157] No. The priority of each empirical sample is represented as follows:
[0158]
[0159] In the formula, Indicates the first Priority of each empirical sample; This represents the temporal difference error of the sample; This indicates that the constraint violates the fraction; Indicates the difficulty score of the scene; This represents the non-negative weight used to adjust the contribution of time-series difference error to the priority of empirical samples; This represents the non-negative weight used to adjust the contribution of constraint violation scores to the priority of empirical samples; This represents the non-negative weight used to adjust the contribution of scene difficulty scores to the priority of experience samples; This represents a constant that guarantees a positive priority.
[0160] The empirical sample sampling probability is:
[0161]
[0162] In the formula, Indicates the first The sampling probability of an empirical sample; This indicates the priority of the sample; Indicates as training progresses Adjusted priority sampling index; This indicates the current number of samples in the experience replay pool; This represents the summation index of the empirical samples; the sum of all sample probabilities is 1.
[0163] To correct for biases caused by non-uniform sampling, importance sampling weights are used in value network updates. New samples are assigned a base priority plus constraint violation and scene difficulty scores. After the value network update, the sample priorities are further updated based on the new temporal difference error. The priority sampling index and the corresponding weights for temporal difference error, constraint violation score, and scene difficulty score are adjusted as the training progresses. During training... Gradually decrease as training progresses; , and The sampling distribution is adjusted according to the training process, so that the sampling probability of key constraint samples and difficult scenario samples is gradually increased in the early stage of training.
[0164] The continuous action policy model employs a dual-delay deep deterministic policy gradient structure, comprising a policy network, two value networks, and a corresponding target network. The target policy action is superimposed with constrained noise, and the target value is the smaller of the outputs of the two target value networks.
[0165]
[0166] In the formula, Indicates the first The target value of an empirical sample; Indicates an immediate reward; Indicates the discount factor; This indicates the termination flag; it is set to 1 when the termination occurs and 0 when the termination has not occurred. Indicates the next state; Indicates the target action corresponding to the next state; , These represent two target value networks. The smaller of their outputs is used to reduce overestimation of value.
[0167] Two value networks are updated in each training update, the policy network is updated with a preset delay, and the target network tracks the online network using a soft update method. Training scenarios are formed by combinations of different wind speeds and fault intensities. During the experience pool pre-filling phase, stable rule actions and actions with perturbations are written into the experience pool together; during the formal training phase, scenarios are extracted cyclically according to weighted scenarios, and exploration noise is gradually reduced. Model evaluation considers average reward, worst-case scenario reward, maximum reactive power deficit, maximum reactive power surplus, active power decrease, and power flow convergence.
[0168] Among them, the equivalent parameters of the turbine group are determined by the capacity and operating data of the wind turbine units included in the turbine group. Let the first... A single wind turbine cluster comprises the following sets of wind turbine units: Then the cluster capacity and initial active power are respectively:
[0169]
[0170]
[0171] In the formula, , They represent the first The equivalent rated apparent power capacity and initial active power of each generator group; Indicates the first A group of wind turbines is a collection of wind turbine units. , They represent the first The rated apparent power capacity and initial active power of the typhoon generator unit. This represents the summation index for wind turbine units.
[0172] The equivalent terminal voltage of the generator group is determined using a capacity-weighted average method.
[0173]
[0174] In the formula, Indicates the first Initial terminal voltage of the typhoon generator unit; Indicates the first The equivalent initial terminal voltage of each cluster; Indicates the first Rated apparent power capacity of typhoon generator units; Indicates the first A collection of wind turbine units in a cluster. Using a capacity-weighted average can avoid the excessive impact of small-capacity units on the equivalent terminal voltage.
[0175] The remaining reactive power margin of the power plant cluster is determined by the current active and reactive power operating points, the upper limit of the absolute value of the allowable reactive power of the cluster, and the capacity circle constraint, and can be expressed as:
[0176]
[0177] In the formula: Indicates the first The remaining reactive power margin of a group of generators is obtained by simultaneously considering the upper limit of the absolute value of the allowable reactive power of the group and the capacity circle constraint at the current operating point. Indicates the first The upper limit of the absolute value of reactive power allowed for a single generator group; This indicates the equivalent rated apparent power capacity of the unit group; and These represent the current active power and reactive power of the generator group, respectively. This represents the absolute value of the current reactive power.
[0178] The station-level and cluster-level states undergo range verification before being input into the strategy model. For voltage and power quantities exceeding the physical range, they are limited according to preset upper and lower limits; for missing states, the most recent valid value or a substitute value calculated by the power flow model is used to avoid non-finite values in the strategy output due to measurement anomalies.
[0179] Fault intensity label This is used to differentiate the difficulty of training and computational scenarios, but the actual reactive power demand is still driven by unsupported fault voltage. Generation. Therefore, scene labels are responsible for describing the fault category. It is responsible for reflecting the actual voltage drop results under the combined effect of fault resistance and network topology.
[0180] After the strategy model outputs actions, the dimension and numerical validity of the actions are first verified, and then the actions are restricted to a preset range of continuous actions. If the strategy model output contains non-finite values, zero-correction actions and rule allocation results based on cluster capacity and sensitivity are used as alternative instructions. The rule allocation results take the total reactive power target as input, form the basic weights according to the product of the cluster reactive power capacity and the PCC voltage reactive power sensitivity, and adopt the same saturated cluster iterative process as the strategy allocation. This alternative method does not change the safety projection constraints and can maintain basic reactive power support capacity when the strategy output is abnormal.
[0181] In the empirical sample, reactive power deficit, reactive power surplus, and active power degradation are first subtracted from numerical tolerances or scenario soft thresholds to obtain the effective constraint violation quantities. Let the first... The effective reactive power deficit, effective reactive power surplus, and effective reactive power decline for each sample are as follows: , and The calculation relationship is as follows:
[0182]
[0183]
[0184]
[0185] In the formula, , , They represent the first The corresponding proportions of reactive power deficit, reactive power surplus, and active power reduction for each empirical sample. , These represent the numerical tolerances for reactive power deficit and reactive power surplus, respectively. This indicates the soft threshold for the allowable decrease in active power in the current wind speed scenario.
[0186] The constraint violation score is constructed using a hierarchical approach. When the sample has an effective reactive power deficit, the effective reactive power deficit is the primary term; when there is no effective reactive power deficit but there is an effective reactive power surplus, the effective reactive power surplus is the primary term; when neither of the first two terms exists, the degree of constraint violation is determined based on the decrease in effective reactive power exceeding the soft threshold.
[0187] The scenario difficulty score is determined by both wind speed level and fault intensity label, and is used to distinguish between capacity-constrained scenarios, boundary fault scenarios, and general scenarios. All other things being equal, the higher the wind speed level and the deeper the PCC voltage drop represented by the fault intensity label, the higher the corresponding scenario difficulty score.
[0188] During training, the priority sampling index, as well as the weights corresponding to temporal difference error, constraint violation, and scene difficulty, are gradually adjusted according to the training progress. A relatively flat sampling distribution is maintained in the early stages of training, while the sampling frequency of key samples is increased in the later stages to balance experience coverage and constraint learning efficiency.
[0189] After the value network update is completed, the temporal difference error of the sampled experience is recalculated, and the new experience priority is written back to the experience replay pool. When new samples enter the experience replay pool, the basic priority is superimposed with constraint violation and scenario difficulty scores to avoid all new samples inheriting the same maximum priority, which would weaken the sample discrimination.
[0190] Candidate strategy models are comprehensively screened based on average reward, worst-case reward, upper bound of reactive power deficit, upper bound of reactive power surplus, and active power degradation index. Evaluation penalties are added to models that do not meet the reactive power deficit or reactive power surplus constraints to prevent models with high average rewards but poor performance in boundary scenarios from being selected.
[0191] S3: After detecting that the wind farm has entered the low voltage ride-through state, acquire the PCC voltage, wind farm power and the operating status of each turbine group, determine the effective reactive power demand of the wind farm based on the PCC fault voltage under the condition of no additional reactive power support, and input the relevant status into the trained deep reinforcement learning strategy model.
[0192] The fault model uses a PCC three-phase ground fault, and the fault resistance is calibrated based on a preset fault intensity label. The fault intensity label is used for training scenario organization and scenario difficulty identification, while the actual effective reactive power demand is determined by the PCC fault voltage under the condition of no additional reactive power support obtained from power flow calculation.
[0193] When the reactive current demand is not lower than the reactive power support start-up voltage threshold, the reactive current demand is zero; when it is between the reactive power support start-up voltage threshold and the deep low voltage threshold, the reactive current demand increases with the degree of voltage drop; when it is lower than the deep low voltage threshold, the reactive current demand reaches the maximum reactive current.
[0194] Reactive current demand can be expressed as:
[0195]
[0196] In the formula: This indicates the per-unit value of reactive current demand; This represents the per-unit value of the PCC fault voltage under conditions of no additional reactive power support. This represents the per-unit value of the maximum reactive current; This represents the proportionality coefficient between voltage sag and reactive current. This indicates the preset reactive power support start-up voltage threshold. This represents a preset low-voltage threshold. This formula is used to determine the reactive current demand based on the PCC fault voltage under conditions of no additional reactive power support. The effective reactive current demand determined based on this reactive current demand is expressed as:
[0197]
[0198] In the formula: This represents the per-unit value of the effective reactive power demand of a wind farm; This represents the reactive power margin factor that is greater than zero. This represents the per-unit value of the PCC fault voltage under conditions of no additional reactive power support. This represents the per-unit value of reactive current demand. Calculating effective reactive demand using the PCC fault voltage under conditions of no additional reactive power support avoids the PCC voltage after control from reversely altering the demand target.
[0199] S4: Generate active power instructions for each turbine cluster and total reactive power target for the wind farm based on the output of the deep reinforcement learning strategy model. Combine the reactive power capacity of the turbine cluster, the reactive power sensitivity of the PCC voltage, and the reactive power allocation action to determine the reactive power allocation ratio of the turbine cluster. Then, redistribute the remaining reactive power for the turbine cluster that has reached the upper limit of reactive power capacity.
[0200] In this context, the active power regulation actions of each generator group are mapped to the permissible active power range during a fault. Let the first... The lower and upper limits of active power commands for a single cluster are respectively: and The active power instruction is:
[0201]
[0202] In the formula: Indicates the first Active commands generated by individual machine clusters; and These represent the lower and upper limits of active power commands allowed during a fault in the generator group, respectively. Indicates the first The active power regulation actions of each generator group, and The upper and lower limits of the active power command are jointly determined by the pre-fault active power reference value and the current available active power, and must satisfy the following conditions: .
[0203] The total reactive power target is determined by the effective reactive power demand and the total reactive power correction action:
[0204]
[0205] In the formula, This indicates the total reactive power target of the wind farm. Indicates effective reactive power demand. Indicates the total reactive power correction action. This represents the non-negative relative correction range, and The total reactive power target is therefore limited to nearby.
[0206] No. The positive reactive power of a cluster under the current active power command is represented as follows:
[0207]
[0208] In the formula, Indicates the first The upper limit of positive reactive power that a cluster can provide under the current active power command. This indicates the upper limit of the absolute value of reactive power allowed for this generator group; Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates the first Individual cluster active power instructions. (The square root contains...) To avoid negative values due to numerical errors.
[0209] Unnormalized allocation coefficients are constructed based on the reactive power capacity of the fleet, the reactive power sensitivity of the PCC voltage, and the strategy actions:
[0210]
[0211] In the formula: Indicates the first Unnormalized reactive power distribution coefficient of individual generator groups; Indicates the first The maximum positive reactive power that a single generator group can provide under the current active power command; Indicates the first PCC voltage reactive power sensitivity of individual generator groups; , These represent the non-negative weight indices for the non-functionality term and the sensitivity term, respectively. Indicates the intensity of reactive power distribution adjustment; Indicates the first Reactive power distribution actions of individual generator groups; This represents a preset small positive number greater than 0, used to prevent the allocation coefficient from failing when the reactive power or PCC voltage reactive power sensitivity is zero or close to zero. Let represent an exponential function. Let the _i_ be the exponential function. The set of clusters that still participate in the allocation is The initial round of cumulative reactive power commands for each machine group Initially unassigned reactive power targets Then the first Wheel of Life The allocation weights for each unsaturated cluster are:
[0212]
[0213] In the formula, Indicates the first Wheel of Life Normalized reactive power allocation weights for an unsaturated cluster of generators Indicates the first The set of machines that still participate in the allocation in the round, , This represents the unnormalized allocation coefficient of the corresponding cluster. This represents the cluster summation index. If the denominator is less than the preset tolerance, the allocation is performed according to the proportion of remaining non-functional capacity in each cluster.
[0214] No. Round assigned to the first The reactive power increment of a single generator group is expressed as:
[0215]
[0216] In the formula, Indicates the first Round assigned to the first The reactive power increment of a single generator group Indicates the first The reactive power targets that have not yet been assigned at the start of the round, This indicates the weight allocation for this round. Indicates the first round at the beginning of the round. The reactive power already allocated to each generator group This indicates the maximum reactive power limit for the generator group. Minimum value calculation ensures that the cumulative reactive power of the generator group does not exceed [a certain value]. .
[0217] After this round of allocation is completed, the reactive power command and remaining reactive power target of the cluster are updated as follows:
[0218]
[0219]
[0220] In the formula, Indicates the first At the start of the round The reactive power already allocated to each generator group; Indicates the first Round assigned to the first The reactive power increase of a single generator group; Indicates completion of the first After the first round of allocation The cumulative reactive power obtained by each generator group; This indicates the reactive power targets that have not yet been allocated at the start of the next round; This indicates the total reactive power target of the wind farm; Indicates the number of generator sets. When the cumulative reactive power of the generator set reaches the corresponding... When the time comes, remove it from the next round of allocation of the generator set until the remaining reactive power target is less than the preset tolerance or all generator sets reach the upper limit of reactive power capacity. The preset tolerance is a preset small positive number greater than 0.
[0221] S5: Based on the power boundaries and capacity circle constraints of each generator group, perform safe projection on the active and reactive power commands to obtain the final active and reactive power control commands for each generator group, and then issue them to the corresponding generator groups. The safe projection first... The active power command of each generator group is limited to between zero and the available active power of that group, resulting in the boundary-limited active power command. Then the first The reactive power command for each generator group is limited to the absolute value of the allowable reactive power of that generator group, resulting in the reactive power command after boundary limiting. .
[0222] No. The maximum reactive power capacity of a single generator group is:
[0223]
[0224] The final reactive power instruction is represented as:
[0225]
[0226] In the formula: Indicates the active power command after boundary limiting; Indicates reactive power command after boundary limiting; Indicates the first Individual clusters of machines receive active power commands after boundary limiting. The upper limit of the absolute value of reactive power is determined by the capacity circle; Indicates the first The equivalent rated apparent power capacity of each unit group; This indicates the final reactive power control command after the safety projection; the final active power control command remains at [value missing]. .
[0227] Example 2
[0228] Based on the method in Example 1, this embodiment further elaborates on the entire process of implementing the present invention, from training preparation, strategy model training, priority processing of experience samples, candidate model screening and evaluation to online fault control, in conjunction with a simulation testing system. It focuses on providing detailed explanations of deep reinforcement learning training, priority experience playback mechanism, and online anomaly handling logic. This embodiment provides a wind farm low-voltage ride-through active and reactive power coordination control method based on reactive power demand guidance and constraint violation of priority experience playback, including wind turbine grouping, strategy model training, fault state acquisition, generation of active and reactive power commands for the turbine group, and safe projection and issuance of control commands.
[0229] First, the wind farm topology, the active power of the wind turbines during normal operation, and the terminal voltage are read. The electrical distance is calculated based on the line and transformer impedances along the path from the wind turbines to the PCC, and the initial active power, electrical distance, and terminal voltage are standardized.
[0230] Multiple clusters are formed using a clustering method. Post-fault response results are not used during the clustering process to avoid cluster partitioning depending on specific fault scenarios. After clustering, the cluster capacity, available active power, equivalent terminal voltage, and PCC voltage reactive power sensitivity are calculated.
[0231] A power flow model is established, with the active and reactive power of wind turbine units and the turbine cluster as nodal injection quantities. A three-phase ground fault is set in the PCC. For each fault intensity label, the fault resistance is adjusted so that the PCC voltage under the condition of no additional reactive power support corresponds to that label.
[0232] The controller generates effective reactive power demand based on the PCC fault voltage under conditions of no additional reactive power support. The controller collects the actual PCC voltage, total active power, total reactive power, effective reactive power demand of the wind farm, and the status of each turbine group. The fault detection signal is used to trigger the controller and is not repeatedly used as the strategy state; the turbine group power is normalized based on the turbine group capacity.
[0233] The state vector is input into the trained policy network to obtain the active power adjustment actions, reactive power allocation actions, and total reactive power correction actions for each cluster. The cluster's active power commands are generated within the range jointly determined by the pre-fault active power reference value and available active power. The total reactive power correction action generates a total reactive power target near the effective reactive power demand, enabling the policy model to make limited adjustments based on the cluster's capacity strain, without freely searching beyond the fault reactive power demand.
[0234] The reactive power capacity is calculated based on the active power command of the generator group, and an allocation coefficient is formed by combining the reactive power capacity of the generator group, the reactive power sensitivity of the PCC voltage, and the reactive power allocation action. First, all unsaturated generator groups are allocated reactive power, and then generator groups that have reached their capacity limits are removed from the subsequent allocation set. When some generator groups become saturated, the remaining reactive power target and the weights of unsaturated generator groups are recalculated. If all allocation coefficients are too small, allocation is performed according to the proportion of remaining reactive power capacity. After a finite number of iterations, the reactive power command for each generator group is obtained.
[0235] The safety projection first limits the absolute values of active and reactive power, then calculates the allowable reactive power range within the capacity circle based on the active power command. Reactive power commands exceeding the capacity range are compressed to the capacity circle boundary. After the safety projection is completed, the final active and reactive power commands are sent to the cluster control interface, and the PCC voltage, wind farm active and reactive power are calculated using the power flow model. Samples with unconverged power flow are marked as abnormal.
[0236] The training scenario consists of combinations of different wind speeds and fault intensities. Each experience sample records the state, action, reward, next state, termination flag, reactive power deficit, reactive power surplus, active power decrease, wind speed level, and fault intensity label.
[0237] The experience priority reflects timing difference error, constraint violation degree, and scenario difficulty. Reactive power deficit is the primary constraint, while reactive power excess and active power degradation are used to evaluate coordination performance.
[0238] In the initial training phase, a relatively flat priority sampling distribution is used to avoid a small number of extreme samples dominating the training. As training progresses, the sampling role of constraint violation samples and difficult scenario samples is gradually increased. Target policy smoothing is achieved by superimposing restricted noise on the target action; the smaller value output from the two target value networks is used to calculate the target value; the policy network uses delayed updates, while the target network uses soft updates. After training, the candidate policy models are evaluated in a scenario-based manner. Evaluation metrics include average reward, worst-case reward, reactive power deficit, reactive power surplus, active power decline, and power flow convergence.
[0239] During the training preparation phase, the initial state of each turbine group is determined using the power flow results of normal wind farm operation, and the available active power is calculated under different wind speed conditions. The active power reference value before the fault changes with wind speed, thus enabling high capacity occupancy in high wind speed scenarios and reflecting the coupling relationship between active power maintenance and reactive power support.
[0240] The experience pool is pre-filled using rule-based action allocation and actionable actions with random perturbations. Rule-based actions ensure the availability of samples that meet basic reactive power requirements in the early stages of training, while random perturbations expand the action coverage, enabling the value network to identify the performance differences between different active and reactive power combinations. During formal training, the policy network outputs continuous actions, and the environment performs action mapping, reactive power iterative allocation, safety projection, and power flow calculation, writing reward and constraint indicators into the experience samples. The value network updates according to priority sampling batches, while the policy network updates according to delay intervals. Policy exploration noise is gradually reduced during training. A larger exploration range in the early stages of training is used to search for different active and reactive power combinations, while a smaller exploration range in the later stages of training is used to refine the cluster allocation relationships within the feasible control area.
[0241] During the model evaluation phase, the policy network parameters are fixed, and the reward, PCC voltage, active power reduction, reactive power deficit, and reactive power surplus are calculated in all test scenarios. No further exploration noise is added during the evaluation process to ensure the comparability of results between different candidate models.
[0242] When multiple candidate models all meet the reactive power deficit constraint, the model with the higher worst-case reward and smaller average active power decrease is selected first; when there is a difference in reactive power surplus, the model with the lower reactive power surplus is further selected.
[0243] The trained policy model is deployed together with the wind turbine clustering results, state standardization rules, and safety projection rules. During the online execution phase, only state acquisition, policy reasoning, action mapping, turbine cluster allocation, and safety projection are performed; the policy network parameters are not updated.
[0244] Furthermore, to verify the effectiveness of the above control methods, combined with Figure 2 The wind farm topology shown is validated using an equivalent system containing 36 wind turbines. Each wind turbine is connected to a step-up substation via collector lines and then to the external power system at the PCC. The wind turbines are divided into four groups based on their initial active power, electrical distance to the PCC, and initial terminal voltage.
[0245] The test scenarios consist of 24 low-voltage ride-through scenarios, formed by combinations of 4 wind speed categories and 6 fault intensities. Comparison methods include the method of this invention, the DDPG method, the method for removing reactive power demand guidance, and the method for removing constraint violation priority experience replay.
[0246] Each method was tested under the same wind farm topology and scenario set. Evaluation metrics included average reward, worst-case reward, average active power degradation rate, average PCC voltage, and average reactive power excess. The overall results are shown in Table 1.
[0247] Table 1 Performance comparison in 24 LVRT scenarios
[0248]
[0249] As shown in Table 1, the average reward of the method of the present invention is 9.5213, the average active power reduction ratio is 0.0336, and the average reactive power excess is 0.0024 in 24 test scenarios. Compared with the DDPG method, the method of the present invention reduces active power reduction and reactive power excess while maintaining the PCC voltage support level.
[0250] After removing reactive power demand guidance, the strategy model needs to search for the total reactive power level on its own. In some scenarios, the reward is low, indicating that the reactive power demand generated without supporting fault flow can provide effective physical priors for action search.
[0251] After removing the constraint violation of priority experience replay, the average active power decrease ratio and the average reactive power excess both increased, indicating that improving the utilization rate of key constraint samples and difficult scenario samples is conducive to improving control coordination.
[0252] Figure 3 The rewards for 24 scenarios are compared. Figures 4 to 6 The PCC voltage, active power, and reactive power responses are presented under representative scenarios. Overall results show that the method of this invention can maintain stable PCC voltage support under different wind speeds and fault intensities.
[0253] The number of turbine clusters, cluster weights, reactive power demand coefficients, network structure, and training parameters in the above embodiments can be adjusted according to wind farm capacity, converter capacity, grid connection specifications, and network topology. Any equivalent substitutions or modifications made without altering the relationships between reactive power demand guidance, turbine cluster coordination and allocation, constraints on violation of priority experience playback, and safety projection techniques are within the scope of protection of this invention.
Claims
1. A wind farm low voltage ride through control method of reactive guidance and constraint priority playback, characterized in that, Includes the following steps: S1: Obtain initial operating information and network topology information of wind turbine units, group wind turbine units according to initial active power, electrical distance to the grid connection point PCC and initial terminal voltage, and determine the operating status, available active power, reactive power and PCC voltage reactive sensitivity of each group. S2: Construct the state space, action space, and training scenario for low voltage ride-through control of wind farms. Determine the priority of experience samples based on the temporal difference error, constraint violation degree, and scenario difficulty of the experience samples, and train the deep reinforcement learning strategy model based on the priority experience playback. S3: After detecting that the wind farm has entered the low voltage ride-through state, acquire the PCC voltage, wind farm power and the operating status of each turbine group, determine the effective reactive power demand of the wind farm based on the PCC fault voltage under the condition of no additional reactive power support, and input the relevant status into the trained deep reinforcement learning strategy model. S4: Generate active power instructions for each turbine group and total reactive power target for the wind farm based on the output of the deep reinforcement learning strategy model. Combine the reactive power capacity of the turbine group, the reactive power sensitivity of the PCC voltage and the reactive power allocation action to determine the reactive power allocation ratio of the turbine group, and redistribute the remaining reactive power for the turbine group that has reached the upper limit of reactive power capacity. S5: Based on the power boundary and capacity circle constraints of each generator group, perform safe projection of active and reactive power commands to obtain the final active and reactive power control commands for each generator group, and issue them to the corresponding generator group.
2. The low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in claim 1, characterized in that, In step S1, the first The electrical distance from the wind turbine generator set to the PCC is the sum of the impedance modulus of each branch on its access path: In the formula: Indicates the first Electrical distance between the typhoon generator and the PCC; Indicates the first The set of access path branches from the typhoon generator to the PCC; Indicates a branch index; Indicates the first The complex impedance of a branch; Indicates the first The impedance magnitude of each branch; For the The first typhoon turbine Standardize the class grouping features: In the formula: Indicates the first Typhoon generator set The original values of the class grouping features; This represents the corresponding standardized value; , They represent the first The mean and standard deviation of the class grouping features, and ; The clustering characteristics include initial active power, electrical distance, and initial terminal voltage; Let the first The set of wind turbine units included in a single turbine cluster is Then the first The equivalent rated apparent power capacity, initial active power, and equivalent initial terminal voltage of each generator group are as follows: In the formula: Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates the first The initial active power of each generator group; Indicates the first The equivalent initial terminal voltage of each cluster; , , They represent the first Rated apparent power capacity, initial active power, and initial terminal voltage of typhoon generator units; PCC voltage reactive power sensitivity. By examining the first A reactive power disturbance is applied to each generator group, and the amount of reactive power disturbance is determined by the ratio of the PCC voltage change to the reactive power disturbance.
3. The wind farm low-voltage ride-through control method with reactive power guidance and constraint priority playback as described in claim 2, characterized in that, In step S2, the deep reinforcement learning policy model adopts a continuous action space, and the state vector of the policy model is represented as: In the formula: Represents the state vector of the policy model; This represents the actual PCC voltage per unit value after applying cluster control; Indicates the fault severity label; , These represent the total active power and total reactive power of the wind farm, respectively, after being normalized using the system power reference value. Indicates effective reactive power demand; Indicates the first Individual cluster status; Indicates the number of aircraft in the fleet; No. The status of a cluster is represented as follows: The respective proportional states are as follows: In the formula: Indicates the first Active load rate of individual machine clusters; Indicates the first Reactive load rate of individual generator groups; Indicates the first Terminal voltage of each unit group; Indicates the proportion of remaining reactive power margin; Indicates the available active power ratio; , , They represent the first The current active power, current reactive power, and available active power of each generator group; Indicates the first The equivalent rated apparent power capacity of each unit group; It represents the residual reactive power margin determined by the capacity circle; The action vector of the policy model is represented as: In the formula: Represents the action vector of the policy model; Indicates the first Active power regulation actions of individual generator groups; Indicates the first Reactive power distribution actions of individual generator groups; This represents the total reactive power correction action; each action is a continuous variable and is restricted to... .
4. The low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in claim 3, characterized in that, In step S2, for the first Based on empirical samples, the proportions of reactive power deficit, reactive power surplus, and active power reduction were calculated respectively: In the formula: , , They represent the first The proportions of reactive power deficit, reactive power surplus, and active power decline corresponding to each empirical sample; Indicates the first Effective reactive power demand corresponding to one empirical sample; This indicates the actual total reactive power output of the wind farm after control. This represents the total active power of the wind farm before the fault. This represents the actual total active power of the wind farm after control. This represents the allowable reactive power support margin coefficient, which is not less than 1. This indicates a positive number that prevents the denominator from being zero; After deducting the corresponding numerical tolerance or scenario soft threshold, the effective constraint violation amount is obtained. Let the first... The effective reactive power deficit, effective reactive power surplus, and effective reactive power decline for each sample are as follows: , and The expression is as follows: In the formula, , , They represent the first The corresponding proportions of reactive power deficit, reactive power surplus, and active power reduction for each empirical sample. , These represent the numerical tolerances for reactive power deficit and reactive power surplus, respectively. This indicates the soft threshold for the allowable decrease in active power in the current wind speed scenario; Constraint violations of the original scores are determined in stratification according to the following formula: The constraint violation score is obtained by using a monotonic compression function: In the formula: Indicates the first Constraint violation score after compression of an empirical sample; Indicates the first The constraints of the empirical sample violate the original score; This represents a compression scale parameter that is greater than zero. Scene difficulty score Determined by both wind speed level and fault intensity label, under the same conditions, the higher the wind speed level and the deeper the PCC voltage drop represented by the fault intensity label, the higher the corresponding scenario difficulty score. No. The priority and sampling probability of each empirical sample are as follows: In the formula: Indicates the first Priority of each empirical sample; This represents the temporal difference error of the empirical sample; This indicates that the constraint violates the fraction; Indicates the difficulty score of the scene; This represents the non-negative weight used to adjust the contribution of time-series difference error to the priority of empirical samples; This represents the non-negative weight used to adjust the contribution of constraint violation scores to the priority of empirical samples; This represents the non-negative weight used to adjust the contribution of scene difficulty scores to the priority of experience samples; This represents a constant that guarantees a positive priority. Indicates the first The sampling probability of an empirical sample; Indicates as training progresses Adjusted priority sampling index; This indicates the current number of samples in the experience replay pool; Represents the index for summing empirical samples; where, The training progresses and the difficulty gradually decreases. , and The sampling distribution is adjusted according to the training process, so that the sampling probability of key constraint samples and difficult scenario samples is gradually increased in the early stage of training.
5. The low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in claim 1, characterized in that, In step S3, the PCC fault voltage under the condition of no additional reactive power support is assumed to be... The reactive current demand is The maximum reactive current is The direct proportionality coefficient between voltage drop and reactive current is: The preset reactive power support start-up voltage threshold is The preset depth low voltage threshold is The reactive current demand is then expressed as: Determine the effective reactive power demand based on the aforementioned reactive current demand: In the formula: This indicates the effective reactive power demand of a wind farm; This represents the reactive power margin factor that is greater than zero. This indicates the PCC fault voltage under conditions of no additional reactive power support. Indicates reactive current demand; Indicates the maximum reactive current; This represents the proportionality coefficient between voltage sag and reactive current. This indicates the preset reactive power support start-up voltage threshold. This indicates the preset deep low voltage threshold.
6. The low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in claim 1, characterized in that, In step S4, the first Active power regulation actions of individual generator groups Transformed according to the linear mapping relationship into the first... Active power instructions for individual machine clusters The active power command range of the linear mapping is jointly determined by the pre-fault active power reference value and the current available active power; the active power adjustment actions of each generator group are mapped to the active power range allowed during the fault period, let the first... The lower and upper limits of active power commands for a single cluster are respectively: and The active power instruction is: In the formula: Indicates the first Active commands generated by individual machine clusters; and These represent the lower and upper limits of active power commands allowed during a fault in the generator group, respectively. Indicates the first The active power regulation actions of each generator group, and The upper and lower limits of the active power command are jointly determined by the pre-fault active power reference value and the current available active power, and satisfy the following conditions: ; Determine the total reactive power target for the wind farm based on the effective reactive power demand and total reactive power correction action: In the formula, This indicates the total reactive power target of the wind farm; Indicates effective reactive power demand; Indicates the total reactive power correction action; This represents the non-negative relative correction range, and ; The total reactive power target is therefore limited to nearby.
7. The low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in claim 6, characterized in that, In step S4, the first The maximum positive reactive power that a single generator group can provide under the current active power command is: In the formula: Indicates the first The maximum positive reactive power that a single generator group can provide under the current active power command; Indicates the first The upper limit of the absolute value of reactive power allowed for a single generator group; Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates the first Active commands for individual machine groups; Based on the forward reactive power of the generator group, the reactive power sensitivity of the PCC voltage, and the reactive power distribution action, an unnormalized reactive power distribution coefficient is constructed: In the formula: Indicates the first Unnormalized reactive power distribution coefficient of individual generator groups; Indicates the first PCC voltage reactive power sensitivity of individual generator groups; , These represent the non-negative weight indices for the non-functionality term and the sensitivity term, respectively. Indicates the intensity of reactive power distribution adjustment; Indicates the first Reactive power distribution actions of individual generator groups; Let the first The set of clusters that still participate in the allocation is The initial round of cumulative reactive power commands for each machine group Initially unassigned reactive power targets Then the first Wheel of Life The normalized reactive power allocation weights for each unsaturated generator group are: No. Round assigned to the first The reactive power increment of each generator group is: Complete the first After the round of allocation, update the cumulative reactive power instructions for each cluster and the reactive power target for the next round that has not yet been allocated: In the formula: Indicates the first Wheel of Life Normalized reactive power allocation weights for individual generator groups; Indicates the first The reactive power targets that have not yet been allocated at the start of the round; , They represent the first Before and after the round of allocation Cumulative reactive power command for each generator group; Indicates the cluster summation index; When the cumulative reactive power command of the generator group reaches When the time comes, the generator group will be removed from the next round of generator group allocation until the remaining reactive power target is less than the preset tolerance or all generator groups reach the upper limit of reactive power capacity. The preset tolerance is a preset small positive number greater than 0.
8. The low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in claim 1, characterized in that, In step S5, the first The active power command of each generator group is limited to between zero and the available active power of that group, resulting in the boundary-limited active power command. ; will the first The reactive power command for each generator group is limited to the absolute value of the allowable reactive power of that generator group, resulting in the reactive power command after boundary limiting. The upper limit of the absolute value of reactive power allowed within the capacity circle is determined based on the active power command after boundary limiting. By limiting the reactive power command after boundary limiting to the reactive power range allowed by the capacity circle, the final reactive power control command is obtained: In the formula: Indicates the first Individual clusters of machines receive active power commands after boundary limiting. The upper limit of the absolute value of reactive power is determined by the capacity circle; Indicates the first The equivalent rated apparent power capacity of each unit group; Indicates reactive power command after boundary limiting; This indicates the final reactive power control command after safety projection; the final active power control command remains at [value missing]. .
9. A low-voltage ride-through control system for wind farms with reactive power guidance and constraint priority playback, used to implement the low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in any one of claims 1-8, characterized in that, include: The module includes a clustering and parameter identification module, a deep reinforcement learning model training module, a fault detection and reactive power demand calculation module, a power command generation and reactive power allocation module, and a power safety projection and command issuance module. The wind turbine clustering and parameter identification module is used to acquire the initial operating information and network topology information of the wind turbines, and to group the wind turbines according to the initial active power, the electrical distance to the grid connection point PCC and the initial terminal voltage, and to determine the operating status, available active power, reactive power and PCC voltage reactive sensitivity of each cluster. The deep reinforcement learning model training module is used to construct the state space, action space and training scenario of low voltage ride-through control in wind farms. It determines the priority of experience samples based on the temporal difference error, constraint violation degree and scenario difficulty of experience samples, and trains the deep reinforcement learning strategy model based on the priority experience playback. The fault detection and reactive power demand calculation module is used to detect whether the wind farm has entered the low voltage ride-through state. When the low voltage ride-through state is detected, the PCC voltage, wind farm power and the operating status of each turbine group are obtained. The effective reactive power demand of the wind farm is determined based on the PCC fault voltage under the condition of no additional reactive power support, and the relevant status is input into the trained deep reinforcement learning strategy model. The power command generation and reactive power allocation module is used to receive the output of the deep reinforcement learning strategy model, generate active power commands for each group and the total reactive power target of the wind farm, determine the reactive power allocation ratio of the group by combining the reactive power capacity of the group, the reactive power sensitivity of the PCC voltage and the reactive power allocation action, and redistribute the remaining reactive power of the group that has reached the upper limit of reactive power capacity. The power safety projection and command issuance module is used to safely project active and reactive power commands based on the power boundary and capacity circle constraints of each generator group, obtain the final active power control command and the final reactive power control command of each generator group, and issue them to the corresponding generator group.
10. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement a low-voltage ride-through control method for wind farms with reactive power guidance and constraint priority playback as described in any one of claims 1-8.
