Wind power plant reactive power rapid tracking optimization method and system

Abstract

The invention discloses a reactive power rapid tracking optimization method and system for a wind power plant in the field of electric power systems. The method comprises the following steps: constructing a power data set comprising a power change rate, the power of the current time period and the actual power value of the next time period by collecting ultra-short-term power data of a wind turbine generator; performing two-dimensional segmentation on the data according to the power size and the change rate, fitting a power predicted value of a next time period based on multiple linear regression analysis, and calculating a confidence interval of the power predicted value; dynamically correcting a voltage margin by combining an active-voltage sensitivity matrix obtained by Jacobian matrix inversion and a power confidence interval; and inputting the actually measured power of each wind turbine generator into the trained SAC decision model to obtain an optimized reactive equipment control strategy, converting into a Markov decision process through an SAC reinforcement learning algorithm to carry out offline training and online rolling optimization to generate a reactor and converter group control strategy, and updating an iteration model. According to the invention, second-level reactive dynamic regulation is realized, the voltage stability is improved, and the network loss is reduced.

Description

Technical Field

[0001] The present invention relates to a method and system for rapid reactive power tracking optimization of a wind farm, belonging to the technical field of power systems. Background Art

[0002] The proposal of the dual-carbon goal indicates that China's new energy industry has entered a period of rapid development. As one of the representative new energies, wind power has also developed rapidly. The large-scale development and utilization of wind power not only have important significance for alleviating environmental pressure and optimizing the energy structure, but also bring significant challenges to the safe and stable operation and economic efficiency of the power grid due to its inherent intermittency and uncertainty. At present, in terms of reactive power optimization, research related to wind farms focuses on the impact of wind farm grid connection, without considering the reactive power flow and optimal voltage distribution of wind farms. Usually, a certain strategy is adopted to determine the reactive power reference values of each wind turbine. With the expansion of the scale of wind farms and the increase in the single-unit capacity of wind turbines, the importance of reactive power and voltage optimization of wind farms has gradually emerged. The output of wind turbines is random due to environmental influence and difficult to accurately predict. The traditional rolling optimization at the 5-minute and 15-minute levels is difficult to cope with the short-term fluctuations of wind power within the period, resulting in the system deviating from the predicted optimal operating state and even causing accidents such as voltage over-limit. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for rapid reactive power tracking optimization of a wind farm, which can solve problems such as the difficulty of reactive power optimization of wind farms caused by the randomness and uncertainty of wind power, and the difficulty of traditional rolling optimization methods in coping with short-term fluctuations.

[0004] To solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0005] In the first aspect, the present invention provides a method for rapid reactive power tracking optimization of a wind farm, which is characterized by including: Obtain the ultra-short-term power time series of each wind turbine in the wind farm, and construct a power data set based on the ultra-short-term power time series, which includes the power change rate, the power at the current time period, and the actual value of the power at the next time period; Perform two-dimensional segmentation processing on the power data set and divide it into multiple groups of data; for each group of data, take the power change rate and the power at the current time period as independent variables, and fit the predicted value of the power at the next time period through multiple linear regression analysis, and calculate the confidence interval; Use the inverse system Jacobian matrix to obtain the active power-voltage sensitivity matrix, and combine the width of the confidence interval of the predicted value of the power at the next time period to calculate the dynamic voltage margin, and generate dynamic voltage boundary constraints based on the dynamic voltage margin; Taking the minimum network loss and voltage violation penalty terms as the objective function, a reactive power optimization model is constructed in combination with the dynamic voltage boundary constraints and transformed into a Markov decision process; The SAC reinforcement learning algorithm is used to perform offline training on the Markov decision process to obtain a trained SAC decision model, and the active / reactive power of each wind turbine based on the real-time monitoring of the wind farm is input into the trained SAC decision model to obtain an optimized reactive power device control strategy.

[0006] Combined with the first aspect, further, the data set is segmented two-dimensionally into multiple groups of data, including: According to the magnitude of the power in the previous period, the power values in the power data set are sorted in ascending order and divided into three groups of data; among them, the first 20% of the data with the smallest power values are the first group, the middle 60% of the data are the second group, and the last 20% of the data with the largest power values are the third group; Each of the three groups of data is sorted in ascending order according to the power change rate and divided into three groups again; among them, the first 20% of the data with the smallest power change rate are the first subgroup; the first 60% of the data with the middle power change rate are the second subgroup; the last 20% of the data with the largest power change rate are the first subgroup; After two segments, the power data set is divided into 9 groups of data.

[0007] Combined with the first aspect, further, the fitting of the predicted value of the power in the next period through multiple linear regression analysis and the calculation of the confidence interval of the predicted value of the power in the next period include: Taking the power change rate and the power in the current period as independent variables and the predicted value of the power in the next period as the dependent variable, a regression equation is established; Constructing a loss function with the mean square error between the predicted value of the power in the next period and the true value, and iteratively updating the regression coefficients in the regression equation through the gradient descent method until convergence; Calculating the standard error of the predicted value of the power in the next period, and determining the Confidence interval.

[0008] Combined with the first aspect, further, the Expression of the confidence interval of the predicted value of the power in the next period is: ; ; where, represents the standard error of the predicted value of the power in the next period; represents the true value of the predicted power in the next period of the i-th sample; denotes the predicted power value of the next time period for the \(i\)-th sample; \(n\) denotes the total number of samples; \(p\) denotes the number of independent variables; denotes the transpose matrix of the independent variable matrix; denotes the independent variable matrix; denotes the critical value of the confidence interval; denotes the confidence level; denotes the confidence interval of the predicted power value of the next time period.

[0009] Combined with the first aspect, further, the construction of the reactive power optimization model includes: Constructing an objective function by minimizing the sum of the system power loss and the voltage violation penalty term; Based on the confidence interval of the predicted power value of the next time period and the active-voltage sensitivity matrix, calculating the dynamic voltage margin; Determining the dynamic voltage boundary constraint according to the voltage margin; Setting the constraint conditions of the reactive power optimization model, including power flow balance constraint, voltage amplitude constraint and reactive power source output constraint conditions; Integrating the objective function, the dynamic voltage boundary constraint and the constraint conditions to construct a wind farm reactive power optimization model.

[0010] Combined with the first aspect, further, the expression of the wind farm reactive power optimization model is: ; where, \(F\) denotes the objective function; denotes the system power loss; denotes the voltage violation penalty term; denotes the penalty coefficient; denotes the active-voltage sensitivity matrix; denotes the critical value of the confidence interval; denotes the standard error of the predicted power value of the next time period; denotes the confidence level; denotes the voltage change amount generated by wind turbine 1 due to power change; denotes the voltage change amount generated by wind turbine 2 due to power change; denotes the voltage change amount generated by wind turbine \(n\) due to power change; denotes the power change value of wind turbine 1; denotes the power change value of wind turbine 2; denotes the power change value of wind turbine \(n\); denotes the power change value of wind turbine \(i\); denotes the adjusted dynamic voltage upper limit of node \(i\); denotes the original voltage upper limit of node \(i\); Denote the voltage margin of node i; Denote the adjusted lower limit of the dynamic voltage of node i; Denote the lower limit of the dynamic voltage of node i; Denote the active power of node i; Denote the voltage amplitude of the i-th node; m represents the total number of system nodes; j represents the node index value of the power grid; Denote the voltage amplitude of the j-th node; Denote the conductance between node i and j; Denote the phase difference between node i and j; Denote the reactive power of node i; Denote the susceptance between node i and j; Denote the active power of wind turbine i; Denote the reactive power of wind turbine i; Denote the apparent power capacity of wind turbine i; Denote the reactive power of reactor bank i; Denote the switching quantity; Denote the unit switching quantity; Denote the maximum switching quantity of reactor bank i.

[0011] Combined with the first aspect, further, transform the reactive power optimization model into a Markov decision process, including: Initialize the parameters of the Actor network and the Critic network, and synchronize them to the parameters of the corresponding Target Critic network. At the same time, create an experience pool D and set its maximum storage capacity; When the current data volume of the experience pool D has not reached its maximum storage capacity, according to the current state , generate an action probability distribution through the Actor network, and sample to generate the current action according to the probability distribution ; After executing the current action, update the current environmental state, and according to the new environmental observation, the next state , and calculate the reward value , and at the same time, form an interaction data tuple consisting of the current state, action, reward value and the next state and store it in the experience pool D; When the experience pool is full, replace the old interaction data tuple with the newly generated interaction data tuple until the current data volume of the experience pool D exceeds its maximum storage capacity; When the current data volume of the experience pool D exceeds its maximum storage capacity, train the SAC decision model, including: Randomly extract a batch of interaction data tuples from the experience pool D; Two Critic networks are used to calculate the action value Q corresponding to the next state respectively. The minimum value of the two is taken as the conservative Q-value estimate, and the target Q-value is calculated by combining the entropy regularization term. Based on the target Q-value, the parameters of the Critic network are updated by minimizing the prediction error. According to the action value Q evaluated by the Critic network, the policy parameters of the Actor network are adjusted, and through the soft update mechanism, the parameters of the Critic network are synchronized to the Target Critic network at a fixed ratio. When the number of training rounds reaches the preset threshold, it is determined that the SAC decision-making model converges, and the trained SAC decision-making model is saved.

[0012] Combined with the first aspect, further, the Markov decision process is divided into four parts, namely the state space, the action space, the state transition function, and the reward function. Among them, the state space includes the active power and reactive power of each wind turbine in the wind farm; the action space includes the reactive power output value of the converter and the reactive power output value of the reactor bank.

[0013] Combined with the first aspect, further, the saving of the trained SAC decision-making model includes: Read the trained SAC decision-making model, and input the active power and reactive power of each wind turbine in the wind farm into the trained SAC decision-making model to obtain the reactive power output value of the converter and the reactive power output value of the reactor bank. According to the reactive power output value of the converter and the reactive power output value of the reactor bank, a reactive power equipment control strategy is generated. At the same time, the active power and reactive power of each wind turbine in the wind farm monitored in real time are archived, and the SAC decision-making model is continuously trained offline, and the SAC decision-making model is iterated regularly.

[0014] In the second aspect, a wind farm reactive power fast tracking and optimization system includes: A data set construction module for obtaining the ultra-short-term power time series of each wind turbine in the wind farm and constructing a power data set including the power change rate, the current period power, and the actual value of the next period power based on the ultra-short-term power time series. A data grouping module for two-dimensional segmentation of the power data set and dividing it into multiple groups of data; among them, for each group of data, with the power change rate and the current period power as independent variables, the predicted value of the next period power is fitted through multiple linear regression analysis, and the confidence interval; A boundary calculation module for using the inverse of the system Jacobian matrix to obtain the active power-voltage sensitivity matrix, and combining the confidence interval of the predicted value of the next period power to calculate the dynamic voltage boundary constraint. A model establishment module, which is used to construct a reactive power optimization model with the minimum network loss and voltage violation penalty term as the objective function, combine the dynamic voltage boundary constraints, and transform it into a Markov decision process; Adopt the SAC reinforcement learning algorithm to perform offline training on the Markov decision process to obtain a trained SAC decision model; A training and optimization module, which is used to perform offline training on the Markov decision process by using the SAC reinforcement learning algorithm to obtain a trained SAC decision model, and input the active / reactive power of each wind turbine based on the real-time monitoring of the wind farm into the trained SAC decision model to obtain an optimized reactive power device control strategy.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The present invention collects the ultra-short-term power time series of each unit in the wind farm and constructs a power data set including the power change rate, the actual value of the power in the current period and the power in the next period. Subsequently, two-dimensional segmentation processing is performed on the data set, which is divided into nine groups according to the power magnitude and change rate. For each group of data, multiple linear regression analysis is used to fit the predicted value of the power in the next period with the power change rate and the current power as independent variables, and the α confidence interval distribution is calculated synchronously, so as to quantify the uncertainty of power fluctuation. On this basis, the active-voltage sensitivity matrix is obtained by inverting the system Jacobian matrix, and the upper and lower limits of the node voltage are dynamically corrected in combination with the confidence interval of the power prediction value to generate voltage boundary constraints adapted to the grid operation state. Further, the power of the unit, the reactive power output of the controllable device and the dynamic voltage constraint monitored in real time are input into the pre-constructed reactive power optimization model, which is transformed into a Markov decision process, and an SAC reinforcement learning algorithm including an Actor network, a Critic network and an experience pool is used for offline training to form a decision model based on the SAC algorithm. Through online rolling optimization, the model updates the control strategy in real time, realizes second-level reactive power regulation, and continuously collects operation data to iteratively update the training set to ensure the adaptability and robustness of the strategy. This method significantly improves the dynamic adjustment ability of the voltage safety margin by integrating the power prediction confidence interval and sensitivity analysis, effectively reduces the system network loss and suppresses the risk of voltage violation, solves the problem of response lag caused by the randomness of wind power in traditional optimization methods, and provides a reliable reactive power control solution for the high-proportion access of new energy to the power grid. Description of the Drawings

[0016] Figure 1 The figure shows a flowchart of a method for quickly tracking and optimizing the reactive power of a wind farm provided by an embodiment of the present invention; Figure 2 The figure shows a flowchart of power randomness processing provided by an embodiment of the present invention; Figure 3The figure shows the flowchart of training the SAC decision model provided by an embodiment of the present invention; Figure 4 The figure shows the flowchart of applying the SAC decision model provided by an embodiment of the present invention. Detailed implementation manners

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other.

[0018] The term "and / or" is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " generally represents an "or" relationship between the associated objects before and after. Embodiment 1

[0019] See Figure 1 and Figure 4 , this embodiment introduces a method for rapid reactive power tracking optimization of a wind farm, including the following steps: Step S1: Obtain the ultra-short-term power time series of each wind turbine in the wind farm, and construct a comprehensive power data set based on the ultra-short-term power time series; This data set not only includes the power change rate and the power at the current time period, but also covers the actual value of the power in the next time period, providing a comprehensive data basis for subsequent analysis and modeling.

[0020] Step S2: Perform two-dimensional segmentation processing on the constructed power data set, and divide it into multiple groups of data; For each group of data, perform multiple linear regression analysis to fit the predicted value of the power in the next time period with the power change rate and the power at the current time period, and simultaneously obtain the confidence interval of the predicted value of the power in the next time period, so as to effectively quantify the uncertainty of power prediction and provide a more accurate basis for reactive power optimization decision-making.

[0021] Step S3: Obtain the active power-voltage sensitivity matrix by inverting the Jacobian matrix of the system, and combine the width of the confidence interval of the predicted value of the power in the next time period to calculate the dynamic voltage margin. Based on this dynamic voltage margin, generate dynamic voltage boundary constraints, that is, the upper and lower limits of the dynamic voltage; This process fully considers the impact of wind power fluctuations on voltage, helps to plan the voltage regulation strategy in advance, and prevent the occurrence of voltage over-limit accidents.

[0022] Step S4: Construct a reactive power optimization model with the minimum network loss and voltage violation penalty terms as the objective function, combined with dynamic voltage boundary constraints, and transform it into a Markov decision process; Step S5: Use the SAC reinforcement learning algorithm to perform offline training on the Markov decision process, obtain a trained SAC decision model, and input the active / reactive power of each wind turbine based on the real-time monitoring of the wind farm into the trained SAC decision model to obtain an optimized reactive power device control strategy.

[0023] Specifically, to further improve the optimization performance and adaptability of the reactive power optimization model, it is transformed into a Markov decision process, and the SAC (Soft Actor-Critic) reinforcement learning algorithm including an Actor network, a Critic network, and an experience pool is used for offline training to obtain a trained SAC decision model, and the control strategy is updated in real time through online rolling optimization.

[0024] The SAC algorithm continuously interacts with the environment to learn the optimal action strategy in different states, thereby realizing the continuous optimization of the reactive power device control strategy of the wind farm. This reinforcement learning-based method can effectively cope with the dynamic changes of the wind farm, improve the real-time performance and accuracy of reactive power optimization, ensure the safe and stable operation of the power grid, reduce network losses at the same time, and improve the economic operation level of the wind farm. Embodiment 2

[0025] (1) Data preparation Obtain the ultra-short-term power time series of each wind turbine in the wind farm, and construct a power dataset including the power change rate , the power at the current time , and the actual value of the power at the next time .

[0026] Assume that the power time series of wind turbine i is , and its power change rate is , and the expression is as follows: (1) where represents the power change rate; represents the power of the wind turbine at the t-th moment; represents the power of the wind turbine at the (t - 1)-th moment.

[0027] Furthermore, construct a power dataset including the power change rate , the power at the current time , and the actual value of the power at the next time , and its expression is as follows: (2) (3) Wherein, represents the power data set of the i-th wind turbine, that is, the reconstructed power time series; represents the data of the i-th wind turbine in period 1; represents the data of the i-th wind turbine in period 2; represents the data of the i-th wind turbine in period n - 1; represents the data of the i-th wind turbine in period t, including the three key parameters in formula (3).

[0028] (2) Power randomness processing The reconstructed power time series is divided into 9 parts according to the magnitude of the power change rate and the power in the current period, including: Step S21: Arrange the power values in the power data set in ascending order according to the power in the previous period, and divide them into three groups of data; among them, the first 20% of the data with the smallest power value is the first group, the middle 60% of the data with the power value is the second group, and the last 20% of the data with the largest power value is the third group; Step S22: Arrange each of the three groups of data in ascending order according to the power change rate, and divide them into three groups again; among them, the first 20% of the data with the smallest power change rate is the first subgroup; the first 60% of the data with the middle power change rate is the second subgroup; the last 20% of the data with the largest power change rate is the first subgroup; After two segments, the power data set is divided into 9 different data sets.

[0029] See Figure 2 , fit the power prediction value of the next period with the power change rate and the power in the current period, and synchronously obtain the distribution of the confidence interval of the power in the next period, including: Step S31: Establish a regression equation with the power change rate and the power in the current period as independent variables and the power prediction value of the next period as the dependent variable; The expression of the regression equation is: (4) (5) Wherein, y represents the dependent variable, that is, the power prediction value of the next period to be fitted; represents the first independent variable, that is, the power change rate; represents the second independent variable, that is, the power in the current period; represents the error term; represents the linear function; represents the first regression coefficient; represents the second regression coefficient.

[0030] By simultaneously solving Equation (4) and Equation (5), we can obtain: (6)

[0031] Step S32: Use the predicted value of the power for the next period and the true value to construct a loss function with the mean square error, and iteratively update the regression coefficients in the regression equation through the gradient descent method until convergence; The expression of the loss function is: (7) where, MSE represents the mean square error; n represents the number of samples; i represents the index value of the sample; represents the true value of the power for the next period of the i-th sample; represents the predicted value of the power for the next period of the i-th sample.

[0032] To make the fitting result more accurate, the gradient descent method is used to update the first regression coefficient and the second regression coefficient , aiming to minimize the loss function MSE. The parameter update formula of the gradient descent method is as follows: (8) where, represents the updated first regression coefficient or the second regression coefficient; the first regression coefficient or the second regression coefficient before update; represents the learning rate; represents the regression coefficient.

[0033] Step S33: Calculate the standard error of the predicted value of the power for the next period, and determine the confidence interval of the predicted value of the power for the next period according to the standard error and the critical value of the preset confidence interval.

[0034] Assuming that the data is normally distributed, calculate the standard error of the predicted value of the power for the next period. The formula is as follows: (9) where, represents the standard error of the predicted value of the power for the next period; n represents the total number of samples; p represents the number of independent variables; represents the transpose matrix of the independent variable matrix; represents the independent variable matrix.

[0035] Based on the preset confidence interval critical value and the standard error, the expression for calculating the confidence interval of the predicted value of the power for the next period is: (10) Among them, represents the confidence interval critical value; represents the confidence level; represents the confidence interval of the predicted value of the power in the next time period.

[0036] Furthermore, the reactive power optimization model constructs the objective function F by minimizing the sum of the system network loss and the voltage violation penalty term; The expression of this objective function is; (11) (12) (13) Among them, F represents the objective function; represents the system network loss; represents the voltage violation penalty term; represents the penalty coefficient; m represents the total number of system nodes; i represents the index of the node in the power grid; j represents the node index value of the power grid; represents the voltage amplitude of the i-th node; represents the voltage amplitude of the j-th node; represents the conductance between node i and j; represents the phase difference between node i and j; represents the adjusted dynamic voltage upper limit of node i; represents the adjusted dynamic voltage lower limit of node i.

[0037] Then, based on the width of the confidence interval of the predicted value of the power in the next time period and the active-voltage sensitivity matrix, the dynamic voltage margin is calculated; Specifically, the active-voltage sensitivity matrix is obtained by inverting the Jacobian matrix, and its expression is: (14) (15) Among them, represents the voltage margin, that is, the voltage amplitude change vector; represents the active-voltage sensitivity matrix; represents the confidence interval range of the power change; represents the voltage phase angle change vector; represents the reactive-voltage sensitivity matrix; represents the reactive-phase angle sensitivity matrix; represents the active-phase angle sensitivity matrix; It represents the voltage amplitude change vector; N represents the sub - matrix of the Jacobian matrix, characterizing the partial derivative of the nodal active power with respect to the voltage amplitude; H represents the sub - matrix of the Jacobian matrix, characterizing the partial derivative of the nodal active power with respect to the phase angle; L represents the sub - matrix of the Jacobian matrix, characterizing the partial derivative of the nodal reactive power with respect to the voltage amplitude; J represents the sub - matrix of the Jacobian matrix, characterizing the partial derivative of the nodal reactive power with respect to the phase angle.

[0038] When only considering the voltage change caused by the active power change, the following voltage margin is obtained: (16) Within the minute - level time scale, the power change value of the wind turbine is limited. Therefore, can be regarded as unchanged, thus linearizing the process of voltage change caused by active power change. Combining with the confidence interval of the predicted power value at the next time step in formula (10), the dynamic voltage margin is obtained, and its expression is: (17) (18) Among them, represents the voltage change amount generated by the power change of wind turbine 1; represents the voltage change amount generated by the power change of wind turbine 2; represents the voltage change amount generated by the power change of wind turbine n; represents the power change value of wind turbine 1; represents the power change value of wind turbine 2; represents the power change value of wind turbine n; represents the power change value of wind turbine i.

[0039] Adjust the dynamic voltage upper and lower limits according to the dynamic voltage margin, and their upper and lower limit expressions are: (19) (20) Among them, represents the adjusted dynamic voltage upper limit of node i; represents the original voltage upper limit of node i; represents the voltage margin of node i; represents the adjusted dynamic voltage lower limit of node i; represents the dynamic voltage lower limit of node i.

[0040] Set the constraint conditions of the reactive power optimization model, including power flow balance constraints, voltage amplitude constraints, and reactive power source output constraints; The said constraint conditions are as follows: a. Power flow balance constraint (21) (22) Wherein, represents the active power of node i; represents the voltage amplitude of the i-th node; m represents the total number of system nodes; represents the voltage amplitude of the j-th node; represents the conductance between node i and j; represents the phase difference between node i and j; represents the reactive power of node i; represents the susceptance between node i and j.

[0041] b. Voltage amplitude constraint (23) c. Reactive power source output constraint (24) (25) Wherein, represents the active power of wind turbine i; represents the reactive power of wind turbine i; represents the apparent power capacity of wind turbine i; represents the reactive power of reactor bank i; represents the switching quantity; represents the unit switching amount; represents the reactor bank i of the maximum switching quantity.

[0042] Finally, integrate the obtained objective function, dynamic voltage upper and lower limits, and constraint conditions above to construct a reactive power optimization model for the wind farm.

[0043] It should be noted that before solving the reactive power optimization problem using the SAC reinforcement learning algorithm, the constructed reactive power optimization model needs to be transformed into a Markov decision process (MDP), where MDP is divided into four parts, namely the state set, action set, state transition function, and reward function; Specifically, the state set includes the active power and reactive power of each wind turbine in the wind farm, and its expression is: (26) Wherein, represents the state set, that is, the set composed of the environmental parameters obtained at time t; represents the active power of the first wind turbine in the wind farm at time t; denotes the active power of the nth wind turbine in the wind farm during the t period; denotes the reactive power of the first wind turbine in the wind farm during the t period; denotes the reactive power of the nth wind turbine in the wind farm during the t period; action set includes the reactive power output values of controllable devices, which are used for decision-making actions based on the current state set, and its expression is: (27) where, denotes the reactive power output value of the ith reactor group during the t period; denotes the reactive power output value of the nlth reactor group during the t period; denotes the reactive power output value of the first converter during the t period; denotes the reactive power output value of the nWTth converter during the t period; denotes the action set; nl denotes the maximum number of reactor groups; nWT denotes the maximum number of converters.

[0044] state transition function , which characterizes the probability that the agent takes the current action in the current state, and its expression is: (28) where, denotes the set composed of the environmental parameters obtained in the t+1 period; denotes the state transition function reward function , which is the impact of the agent on the environment after the current action, and a reward function is needed to quantify this impact, and its expression is: (29) where, denotes the reward function; denotes the original function value; denotes after the current action function value.

[0045] Specifically, participating in Figure 3 , the constructed reactive power optimization model will be transformed into a Markov decision process, including: Step S71, initialize the parameters of the Actor network and the Critic network and , and synchronize them to the parameters of the corresponding Target Critic network , and make , simultaneously create an experience pool D for storing training data and set its maximum storage capacity; Step S72: When the current data volume in the experience pool D has not reached its maximum storage capacity, during the t-th training, according to the current state , input the training data sampled from the state set into the SAC decision-making model. The Actor network generates an action probability distribution based on the SAC decision-making model and samples a current action according to the probability distribution ; The selected action expression is: (30) where represents the current action; represents the Actor network; represents the current state.

[0046] Step S73: After executing the current action , update the current environmental state, and based on the new environmental observation, obtain the next state , and calculate the reward value , and at the same time store the interaction data tuple composed of the current state, action, reward value, and next state into the experience pool D; The expression of the interaction data tuple is: (31) where represents the reward value; represents the next state.

[0047] When the experience pool is full, replace the old interaction data tuple with the newly generated interaction data tuple, and repeat Step S73 until the current data volume in the experience pool D exceeds its maximum storage capacity; When the current data volume in the experience pool D exceeds its maximum storage capacity, train the SAC decision-making model, including: Step S731: Randomly extract a batch of interaction data tuples from the experience pool D; Step S732: Use two Critic networks to calculate the action value Q corresponding to the next state ; The expression of the action value is: (32) (33) where represents the reward value of the i-th training sample; represents the discount factor; represents at In the state of, different actions are taken through the Actor network , and calculate the values corresponding to these actions minus the regularization term of the expected value; denotes the expected value of choosing the action under the state according to the policy; denotes the action taken in the (i + 1)-th training sample; denotes the environmental state in the (i + 1)-th training sample; denotes the state under which the action is taken; denotes the temperature coefficient; denotes the Q-value output by one of the Critic networks; denotes the Q-value output by the other Critic network.

[0048] It should be noted that and take the minimum term z of the two Critic networks as the conservative Q-value estimate, which can avoid overestimation of the Q-value.

[0049] Furthermore, calculate the target Q-value in combination with the entropy regularization term, and update the parameters of the Critic network by minimizing the prediction error. The formula is as follows: (34) where denotes the loss function of the Critic network; denotes the output of the target Q-value; denotes the i th state in the training sample; denotes the i th action in the training sample.

[0050] Step S733: Adjust the policy parameters of the Actor network according to the action value Q evaluated by the Critic network to maximize the long-term cumulative reward with entropy regularization; where, update the parameters of the Actor network using the policy gradient algorithm and minimize the following objective function: (35) where denotes the objective function of the policy network; denotes the expected value of sampling the state D from the data set ; denotes the state Under this condition, select an action according to the policy expected value; represents the action-value function.

[0051] Furthermore, through the soft update mechanism, synchronize the Critic network parameters to the Target Critic network at a fixed ratio, making the target network parameters slowly approach the current network to avoid oscillations; The expression of the updated Critic network parameters is: (36) where represents the updated Critic network parameters; represents the original Critic network parameters; represents the soft update step size.

[0052] Step S735: When the number of training rounds reaches a preset threshold, determine that the SAC decision model converges, and save the trained SAC decision model.

[0053] Specifically, read the trained SAC decision model, input the active power and reactive power of each wind turbine in the wind farm into the trained SAC decision model to obtain the reactive power control strategy, that is, the reactive power output value of the controllable devices (converters and reactor banks). At the same time, archive the active power and reactive power of each wind turbine in the wind farm monitored in real time, continue to perform offline training on the SAC decision model, and periodically iterate the currently used SAC model. Embodiment 3

[0054] A dataset construction module for obtaining the ultra-short-term power time series of each wind turbine in the wind farm and constructing a power dataset including the power change rate, the current period power, and the actual value of the next period power based on the ultra-short-term power time series; A data grouping module for two-dimensional segmentation of the power dataset into multiple groups of data; among them, for each group of data, with the power change rate and the current period power as independent variables, fit the predicted value of the next period power through multiple linear regression analysis, and calculate the confidence interval; A boundary calculation module for obtaining the active-voltage sensitivity matrix by inverting the system Jacobian matrix, and calculating the dynamic voltage boundary constraint in combination with the confidence interval of the predicted value of the next period power; A model establishment module for constructing a reactive power optimization model with the minimum network loss and voltage violation penalty term as the objective function, in combination with the dynamic voltage boundary constraint, and transforming it into a Markov decision process; The Markov decision process is offline trained using the SAC reinforcement learning algorithm to obtain a trained SAC decision model; A training optimization module is configured to offline train the Markov decision process using the SAC reinforcement learning algorithm to obtain a trained SAC decision model, and input the active / reactive power of each wind turbine based on the real-time monitoring of the wind farm into the trained SAC decision model to obtain an optimized reactive power device control strategy.

[0055] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the claims of the present invention. These all fall within the protection scope of the present invention.

Claims

1. A method for rapid tracking and optimizing reactive power in a wind farm, characterized in that: include: Obtaining an ultra-short-term power time series of each wind turbine in the wind farm, and constructing a power data set including a power change rate, a current period power, and an actual power value of a next period based on the ultra-short-term power time series; The power data set is subjected to two-dimensional segmentation processing to divide it into multiple groups of data; wherein, for each group of data, the power change rate and the power of the current period are used as independent variables, and the power prediction value of the next period is fitted by multiple linear regression analysis, and the power prediction value of the next period is calculated. confidence intervals; The active-voltage sensitivity matrix is obtained by inverting the system Jacobian matrix, and the power prediction value of the next period is combined with the The width of the confidence interval is used to calculate a dynamic voltage margin, and a dynamic voltage boundary constraint is generated based on the dynamic voltage margin. Taking the minimum network loss and voltage over-limit penalty term as the objective function, a reactive power optimization model is constructed in combination with the dynamic voltage boundary constraint and converted into a Markov decision process; The Markov decision process is trained offline using a SAC reinforcement learning algorithm to obtain a trained SAC decision model. The active / reactive power of each wind turbine based on real-time monitoring of the wind farm is input into the trained SAC decision model to obtain an optimized reactive equipment control strategy.

2. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 1, characterized in that: The data set is segmented in two dimensions into multiple groups of data, including: According to the power size of the previous period, the power values in the power data set are arranged in ascending order and divided into three groups of data; among which, the first 20% of the data with the smallest power values are the first group, the middle 60% of the data with the middle power values are the second group, and the last 20% of the data with the largest power values are the third group; Each of the three groups of data is arranged in ascending order of power change rate and further divided into three groups; wherein the first 20% of the data with the smallest power change rate is the first subgroup; the first 60% of the data with the middle power change rate is the second subgroup; and the last 20% of the data with the largest power change rate is the first subgroup; After two segmentations, the power data set is divided into 9 groups of data.

3. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 1, characterized in that: The step of fitting the power prediction value for the next period of time by multivariate linear regression analysis and calculating the confidence interval of the power prediction value for the next period of time includes: Establish a regression equation with the power change rate and the power in the current period as independent variables and the power prediction value in the next period as the dependent variable; A loss function is constructed based on the mean square error between the power prediction value and the true value of the next time period, and the regression coefficient in the regression equation is iteratively updated by a gradient descent method until convergence; Calculate the standard error of the power forecast value for the next period, and determine the power forecast value for the next period based on the standard error and the critical value of the preset confidence interval. Confidence interval.

4. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 3, characterized in that: The power prediction value for the next period The expression for the confidence interval is: ； ； in, Indicates the standard error of the power forecast value for the next period; represents the predicted true value of the power in the next period of the i-th sample; represents the power prediction value of the next period of the i-th sample; n represents the total number of samples; p represents the number of independent variables; represents the transposed matrix of the independent variable matrix; represents the independent variable matrix; represents the critical value of the confidence interval; Indicates the confidence level; Indicates the power forecast value for the next period Confidence interval.

5. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 1, characterized in that: The constructing of the reactive power optimization model includes: The objective function is constructed by minimizing the sum of system network loss and voltage over-limit penalty; Based on the power prediction value of the next period Calculate the dynamic voltage margin based on the confidence interval and the active power-voltage sensitivity matrix; determining a dynamic voltage boundary constraint based on the voltage margin; Setting constraints of the reactive power optimization model, including power flow balance constraints, voltage amplitude constraints, and reactive power source output constraints; The objective function, dynamic voltage boundary constraints and constraint conditions are integrated to construct a reactive power optimization model.

6. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 5, characterized in that: The expression of the reactive power optimization model is: ； Where F represents the objective function; Indicates system network loss; Indicates the voltage over-limit penalty item; represents the penalty coefficient; represents the active power-voltage sensitivity matrix; represents the critical value of the confidence interval; Indicates the standard error of the power forecast value for the next period; Indicates the confidence level; Indicates the voltage change caused by power change of wind turbine 1; Indicates the voltage change caused by power change of motor group 2; Indicates the voltage change of motor group n caused by power change; Indicates the power change value of wind turbine 1; Indicates the power change value of wind turbine 2; Indicates the power change value of wind turbine n; represents the power change value of wind turbine i; represents the adjusted dynamic voltage upper limit of node i; represents the upper limit of the original voltage of node i; represents the voltage margin of node i; represents the lower limit of the dynamic voltage after adjustment at node i; represents the dynamic voltage lower limit of node i; represents the active power of node i; represents the voltage amplitude of the i-th node; m represents the total number of system nodes; j represents the node index value of the power grid; represents the voltage amplitude of the jth node; represents the conductance between nodes i and j; represents the phase difference between nodes i and j; represents the reactive power of node i; represents the susceptance between nodes i and j; represents the active power of wind turbine i; represents the reactive power of wind turbine i; represents the apparent power capacity of wind turbine i; represents the reactive power of reactor group i; Indicates the number of switches; Indicates unit switching capacity; Indicates the maximum switching quantity of reactor group i.

7. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 5, characterized in that: The reactive power optimization model is converted into a Markov decision process, including: Initialize the parameters of the Actor network and the Critic network, and synchronize them with the parameters of the corresponding Target Critic network. At the same time, create the experience pool D and set its maximum storage capacity. When the current data volume of the experience pool D does not reach its maximum storage capacity, according to the current state , generate action probability distribution through Actor network, and generate current action based on the probability distribution sampling After executing the current action, update the current environment state and observe the next state according to the new environment , and calculate the reward value , and the interaction data tuple consisting of the current state, action, reward value and next state Deposit into experience pool D; When the experience pool is full, the old interaction data tuples are replaced by the latest generated interaction data tuples until the current data volume of the experience pool D exceeds its maximum storage capacity; When the current data volume of the experience pool D exceeds its maximum storage capacity, the SAC decision model is trained, including: Randomly extract batch interaction data tuples from the experience pool D; Use two critic networks to calculate the action value Q corresponding to the next state respectively, take the minimum value of the two as the conservative Q value estimate, and calculate the target Q value in combination with the entropy regularization term. Based on the target Q value, update the critic network parameters by minimizing the prediction error; According to the action value Q evaluated by the Critic network, the policy parameters of the Actor network are adjusted, and the Critic network parameters are synchronized to the Target Critic network at a fixed ratio through a soft update mechanism; When the number of training rounds reaches a preset threshold, the SAC decision model is determined to be converged, and the trained SAC decision model is saved.

8. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 7, characterized in that: The Markov decision process is divided into four parts: state space, action space, state transfer function and reward function; The state space includes the active power and reactive power of each wind turbine in the wind farm; the action space includes the reactive output value of the converter and the reactive output value of the reactor group.

9. The method for rapid tracking and optimizing reactive power of a wind farm according to claim 8, characterized in that: Saving the trained SAC decision model includes: Reading the trained SAC decision model, and inputting the active power and reactive power of each wind turbine in the wind farm into the trained SAC decision model to obtain the reactive output value of the converter and the reactive output value of the reactor group; A reactive device control strategy is generated based on the reactive output value of the converter and the reactive output value of the reactor group. At the same time, the active power and reactive power of each wind turbine in the wind farm monitored in real time are archived, the SAC decision model is continuously trained offline, and the SAC decision model is iterated regularly.

10. A wind farm reactive power fast tracking optimization system, characterized in that: include: A data set construction module is used to obtain the ultra-short-term power time series of each wind turbine in the wind farm, and to construct a power data set including the power change rate, the power in the current period, and the actual power value in the next period based on the ultra-short-term power time series; The data grouping module is used to perform two-dimensional segmentation on the power data set and divide it into multiple groups of data; wherein, for each group of data, the power change rate and the power of the current period are used as independent variables, and the power prediction value of the next period is fitted by multiple linear regression analysis, and the power prediction value of the next period is calculated. confidence intervals; The boundary calculation module is used to obtain the active-voltage sensitivity matrix by inverting the system Jacobian matrix and combining it with the power prediction value of the next period. Confidence intervals,calculate dynamic voltage bound constraints; A model building module is used to construct a reactive power optimization model with minimum network loss and voltage over-limit penalty as the objective function, combined with the dynamic voltage boundary constraint, and convert it into a Markov decision process; The Markov decision process is trained offline using a SAC reinforcement learning algorithm to obtain a trained SAC decision model; The training optimization module is used to perform offline training on the Markov decision process using the SAC reinforcement learning algorithm to obtain a trained SAC decision model, and input the active / reactive power of each wind turbine based on real-time monitoring of the wind farm into the trained SAC decision model to obtain an optimized reactive equipment control strategy.

Images