Steady-state reactive power optimization method for synchronous generators to improve voltage temporary stability based on machine learning

Abstract

The invention relates to a method for optimizing the steady-state reactive power of a synchronous generator for improving the temporary stability of voltage based on machine learning. Specifically, based on a studied power flow mode and an expected fault set, an XGBOOST machine learning method is used to establish a steady-state value of a multidimensional input feature vector according to the studied power flow mode, and a classifier is used to predict whether the system voltage maintains transient stability to an expected fault disturbance. The system transient voltage can maintain stability under a determined fault disturbance as a constraint condition, and the network loss caused by the steady-state reactive power optimization of the generator is minimized as an objective function. A steady-state reactive power optimization model for improving the dynamic reactive power support provided by the synchronous generator for the transient voltage is established, and a particle swarm algorithm is used to call the power flow calculation to solve the optimization model, and the optimal solution of the steady-state reactive power of the generator is determined. The invention can realize the prediction of the stability of the system transient voltage to the expected fault disturbance based on the steady-state power flow calculation, and improve the system transient voltage stability by optimizing the steady-state output reactive power of the generator.

Description

Technical Field

[0001] The present invention relates to a method for optimizing the steady-state reactive output of a synchronous generator, in particular to a method for using machine learning to establish a classifier for predicting the transient voltage stability of a system based on the steady-state value of a multi-dimensional input feature vector of a studied power flow mode, and a generator steady-state output reactive optimization model based on the classifier for improving the dynamic reactive support provided by the synchronous generator for the transient voltage while taking into account the reduction of the influence of reactive regulation on the network loss. Background Art

[0002] Due to the rapid development of DC transmission projects, a large number of DC power lines are connected to the receiving power grid, causing the startup scale of local power units in the power grid to be severely squeezed. When faced with voltage drops caused by system failures, the system cannot provide sufficient reactive power support for voltage recovery, resulting in voltage instability, posing a threat to the safe and stable operation of the power grid.

[0003] At present, the main measures to improve the transient voltage stability of the receiving power grid include installing dynamic reactive compensation devices and emergency load shedding control. Since dynamic reactive compensation devices are expensive, they will increase investment costs, and emergency load shedding control will also cause a certain degree of economic losses to the power grid. Relevant studies have pointed out that considering the reactive output characteristics of generators and capacitors during transient periods, increasing the reactive output ratio of generators in steady state and reducing the reactive output ratio of capacitors in substations can effectively improve the transient voltage stability of the system. However, this method may cause unreasonable reactive distribution during steady-state operation of the system, thereby increasing system network losses and causing economic losses.

[0004] Starting from this, while considering minimizing the problem of increased system network losses caused by the coordination of steady-state reactive output of generators and parallel capacitors, the present invention considers the dynamic characteristics of reactive regulation of generators and capacitor reactors during fault transients, and proposes a classifier modeling method based on the studied power flow mode and expected fault set, using XGBOOST to establish a classifier modeling method based on the steady-state value of a multi-dimensional input feature vector to predict whether the system transient voltage remains stable to the expected fault disturbance. On this basis, with the constraint that the system transient voltage can remain stable under a certain fault disturbance, minimizing the network loss caused by the steady-state reactive optimization of the generator is considered as the objective function, and a steady-state reactive optimization model is established to improve the dynamic reactive support provided by the synchronous generator for the transient voltage. Summary of the invention

[0005] The present invention mainly overcomes the shortcomings of using transient stability calculation to optimize the steady-state reactive output of the generator to improve the transient stability of the grid voltage, which has a complex calculation and analysis process and a large amount of calculation. It provides a classifier that uses machine learning to establish a multi-dimensional input feature quantity based on the steady-state power flow of the power grid and anticipated faults to predict the transient stability of the power grid voltage, and based on the classifier, optimizes the steady-state reactive output of the generator based on the power flow calculation, which can effectively avoid a large number of complex transient stability calculations and analyses in the optimization process.

[0006] Another object of the present invention is to improve the ability of the synchronous generator to provide dynamic reactive support to the power grid during a certain fault disturbance by optimizing the steady-state reactive output of the synchronous generator, so as to improve the transient stability of the system voltage, while reducing the impact of steady-state output reactive regulation on network losses.

[0007] The above technical problems of the present invention are mainly solved by the following technical solutions:

[0008] A method for optimizing the steady-state reactive power of a synchronous generator to improve voltage temporary stability based on machine learning, characterized in that it includes:

[0009] The XGBOOST machine learning method is used to establish a classifier based on the system steady-state multi-dimensional input feature vector and the expected fault type to determine whether the system transient voltage remains stable to the expected fault disturbance.

[0010] Taking the system transient voltage remaining stable under a certain fault disturbance as a constraint condition and considering minimizing the network loss caused by the steady-state reactive power optimization of the generator as the objective function, a steady-state reactive power optimization model is established to improve the dynamic reactive power support provided by the synchronous generator for the transient voltage. The particle swarm algorithm is used to call the power flow calculation to solve the optimization model and determine the optimal steady-state reactive power control of the generator.

[0011] In the above method, the multi-dimensional input feature vector is defined as follows:

[0012] Table 1

[0013]

[0014] In the above method, classifier modeling specifically includes:

[0015] Identify the weak transient stability areas of system voltage and the voltage regulation sensitive generators in the area;

[0016] Based on the transient stability calculation of severe anticipated fault sets in power grid, an XGBOOST classifier is established to predict system voltage transient stability based on multi-dimensional input feature vectors and fault types.

[0017] In the above method, when determining the weak area of ​​system voltage transient stability and the voltage regulation sensitive generator sets in the area, it includes:

[0018] Step 1: Use PSASP and other power system numerical simulation software to establish the mathematical model of power flow and transient stability calculation and analysis for the system's determined operating mode, perform transient stability time-domain simulation calculations for the expected fault set disturbances, and obtain the transient response curve of the bus voltage to the fault disturbance.

[0019] Step 2: Based on the transient voltage response curve of bus i to the expected fault j disturbance, calculate the transient voltage stability margin according to formula (1):

[0020]

[0021] In the formula is the voltage transient stability margin of bus i under fault j disturbance described by multi-binary table, V i N and V i j (t) represent the voltage reference value of bus i and the real-time value under fault j, t k and t k ′ are V i j (t) The voltage below and above the kth critical reference value during the drop and recovery process The moment, K k It means (t k ,t k+1 )∪(t k ' +1 ,t k ′) interval integral weight coefficient, when V i j (t) are less than When the voltage drop and recovery process is located at (t1, t2)∪(t2′, t1′),…,(t k ,t k+1 )∪(t k ' +1 ,t k ′),…,(t n ,t n The integral weight coefficients of the intervals between ′ are K1,…,K k ,…,K n , the weight coefficient of each interval can be based on the voltage transient response to keep the voltage critical transient stability, that is, Determine it by solving formula (2) step by step.

[0022]

[0023] In the formula Indicates that the voltage is lower than The maximum acceptable duration.

[0024] Step 3: According to the transient voltage stability margin obtained based on the multi-binary table Constructing transient stability margin index similarity matrix A is used as the input of the AP clustering algorithm to cluster and partition each bus of the studied power flow mode.

[0025] Step 4: Based on the column vector of A, according to Determine the severity of fault j according to ξ j The sizes are arranged from small to large from the middle to both sides, and the row vector describing the severity of each expected fault in the studied power flow mode is defined, that is, the middle element ξ of the row vector p The value is the smallest and the corresponding fault p disturbance is the most serious.

[0026] Step 5: According to formula (III), the comprehensive transient stability margin ∑η of each bus voltage to different fault disturbances is calculated based on linear weighting: i , and use ∑η i Sorting determines the smaller ∑η i The busbar is taken as the weak busbar of transient voltage stability margin, and then according to the partition where the weak busbar of transient voltage stability margin is located, the weak area of ​​transient voltage stability margin under the studied power flow mode is determined.

[0027]

[0028] Where q represents the number of expected failures, ω j represents the weight of the jth fault, l is composed of the row vector ξ describing the severity of each expected fault j The column position value is determined.

[0029] Step 6: Using the power flow calculation and analysis model established in step 1, increase the output reactive power of the synchronous generator r by △Q r =0.1pu, based on the most serious fault p determined in step 4, transient stability calculation and analysis are performed, and the weak area bus i is recalculated according to formula (1) after the generator r changes its reactive output According to formula (IV), the voltage regulation sensitivity s of the synchronous generator r reactive regulation to the bus with weak voltage transient stability margin is calculated r , and then according to s r The synchronous generator sets with sensitive voltage regulation effect on the bus with weak transient voltage stability margin are determined by sorting.

[0030]

[0031] In the above method, the establishment of XGBOOST classifier includes:

[0032] Step 1: Based on the transient voltage response curve obtained in step 1 of determining the weak area of ​​system voltage transient stability and the voltage-sensitive generator sets in the area, determine the severe anticipated fault set based on the standard that the steady-state value of the transient voltage response cannot be restored to above 0.9pu after the fault disturbance.

[0033] Step 2: Adjust the steady-state output reactive power of the determined voltage-regulated sensitive generator set and the reactive power of the sensitive generator set in the vicinity of the parallel connection and calculate the power flow, and further perform transient stability calculations for each severe anticipated fault disturbance on each convergent power flow;

[0034] Step 3: Taking the steady-state value of transient voltage response restored to above 0.9pu as the standard, determine whether the convergent power flow can maintain transient voltage stability for severe anticipated fault disturbances. Set the label of the convergent power flow with transient voltage stability to 1, otherwise set the label to 0. At the same time, based on the steady-state value of the multi-dimensional input feature vector of the convergent power flow and the severe anticipated fault type, construct a machine learning sample for predicting the system voltage transient stability classifier. On this basis, based on the voltage transient response curves of different convergent power flows to each severe anticipated fault disturbance, establish a sample data set for machine learning.

[0035] Step 4: Divide the sample data set into a training set and a validation set in a ratio of 7:3. The multi-dimensional input feature vector of the training set sample i is used as the XGBOOST classifier decision tree f shown in formula (V). k (x i ) i , calculate the score of the classifier output As a result of predicting whether the system voltage of the studied power flow mode maintains transient stability to the expected fault disturbance, by increasing the number of decision trees in formula (V) until the loss function shown in formula (VI) reaches the minimum value, the iterative training of XGBOOST classifier modeling is stopped, and the classifier for predicting the transient voltage stability of the studied power flow mode corresponding to the training set is obtained.

[0036]

[0037] In the formula, f k (x i ) represents the kth decision tree of XGBOOST, N R is the total number of decision trees when iterative training is stopped, is the error between the model prediction and the true value, Ω(f k ) is a regular term used to avoid overfitting of the model, γ and λ are penalty coefficients of the regular term, T is used to control the number of leaf nodes in the tree model to prevent the tree model from being too complex, and W is used to avoid the leaf node weight vector being too large.

[0038] In the above method, the steady-state reactive power output optimization model is based on the following objective function and constraints:

[0039] minF(Q G )=P loss ,Q G =(Q G1 ,Q G2 …Q Gz )

[0040]

[0041] In the formula, Q G represents the reactive power output combination of the voltage-sensitive generator sets, Gz represents the number of sensitive generator sets, P loss Represents the system network loss, G ij , B ij and θ ij Respectively represent the conductance, susceptance and phase difference between busbars i and j, U i and U j Represent the voltage amplitude of busbars i and j respectively, P Gi , Q Gi and P Li , Q Li They represent the active and reactive power input of the generator of bus i and the active and reactive power demand of the load, respectively. Di represents the injected active power of bus i, Q Ci represents the parallel compensation reactive power of bus i, Q kGmin , Q kGmax They represent the upper and lower limits of reactive power output of voltage-sensitive generators, U imin , U imax They represent the upper and lower limits of the voltage of bus i, N is the number of system buses, h(X) is a classifier based on XGBOOST to predict whether the system voltage maintains transient stability, and X represents the multidimensional input feature vector of the studied power flow mode and the expected fault type.

[0042] In the above method, the steps of solving the optimization model include:

[0043] Step 1: Set the determined steady-state reactive output of the voltage-sensitive generator as the optimization particle, initialize the position and speed of the particle group used to optimize the steady-state reactive output of each voltage-sensitive generator group, and the number of optimization iterations;

[0044] Step 2: According to the optimized particle, the output reactive power of the voltage-sensitive generator set is adjusted, and the reactive power of the parallel compensation of the sensitive generator set in the vicinity is adjusted. If the power flow converges, based on the multi-dimensional input feature vector, the classifier is used to predict whether the system voltage maintains transient stability to the expected fault disturbance. If it is transiently stable, step 3 is executed, and if it is unstable, step 4 is executed;

[0045] Step 3: Calculate the objective function fitness value of the optimization model shown in formula (VII). If the current fitness value is better than the best fitness value in memory, replace it and record the current position as the best position of the particle. At the same time, compare the best position of the particle and the best position of the population. If the best position of the particle is better than the optimal position of the population, replace it, otherwise it remains unchanged.

[0046] Step 4: Update the position and velocity of the optimization particle and return to step 2 for cyclic calculation until the convergence condition of the optimization algorithm is met or the number of iterations reaches the maximum, and determine the optimal command value of the steady-state output reactive power of the voltage-sensitive generator set.

[0047] The present invention can not only predict the stability of the system transient voltage to the expected fault disturbance based on the steady-state power flow calculation, but also improve the system transient voltage stability by optimizing the steady-state output reactive power of the generator, while taking into account the reduction of the adverse effects of the steady-state output reactive power regulation of the generator on the network loss. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the flow chart for optimizing dynamic reactive support of synchronous generators based on XGBOOST classification.

[0049] Figure 2 This is the wiring diagram of the 500kv main grid of Jiangxi Power Grid.

[0050] Figure 3a This is the iterative change diagram of the system network loss (4081 network loss).

[0051] Figure 3b This is the iterative change diagram of the system network loss (4406 network loss).

[0052] Figure 3c This is the iterative change diagram of the system network loss (4438 network loss).

[0053] Figure 4a To control the Nanchang 500a bus voltage diagram before and after (4081 fault).

[0054] Figure 4b To control the front and rear Nanchang 500a bus voltage diagram (4406 fault).

[0055] Figure 4c To control the front and rear Nanchang 500a bus voltage diagram (4438 fault).

[0056] Figure 5a To control the reactive power output diagram of Fuzhou Power Plant before and after (4081 fault).

[0057] Figure 5b To control the reactive power output diagram of Fuzhou Power Plant before and after (4406 fault).

[0058] Figure 5c To control the reactive power output diagram of Fuzhou Power Plant before and after (4438 fault). DETAILED DESCRIPTION

[0059] The technical solution of the present invention is further specifically described below through embodiments and in combination with the accompanying drawings. A method for optimizing dynamic reactive support of synchronous generators based on XGBOOST classification is implemented.

[0060] This application uses XGBOOST to construct a classification model for voltage transient stability, and uses a particle swarm algorithm (PSO) to solve the above-mentioned optimization mathematical model for optimizing the steady-state reactive output of voltage-sensitive generators in weak areas of voltage transient stability.

[0061] The specific implementation steps are as follows:

[0062] 1) Based on the expected fault set, the time domain transient stability simulation calculation is performed in PSASP to obtain the voltage transient stability curve.

[0063] 2) Based on the multi-binary table, the constructed voltage transient stability similarity matrix of different buses under the expected fault set is calculated, and the AP clustering algorithm is used to partition the system.

[0064] 3) Calculate the comprehensive voltage transient stability margin of the system bus and determine the transient voltage weak area of ​​the system. According to the transient voltage stability criterion determined by the Technical Specifications for Power System Safety and Stability Calculation issued by the National Grid Operation and Control Standardization Technical Committee, the multi-binary table critical voltage and integral weight coefficient of the quantified bus voltage transient stability margin shown in Table 2 can be determined.

[0065] Table 2 Multi-binary table critical voltage and integral weight coefficient (V N =1.0pu)

[0066]

[0067] 4) Apply reactive disturbance to the generators in the system, calculate the reactive-voltage transient stability sensitivity index of the generators, and screen out key sensitive generator sets.

[0068] 5) Select the set of anticipated serious faults that the system needs to cope with.

[0069] 6) In PSASP, different reactive power output levels are randomly set for the key sensitive generators selected in 4).

[0070] 7) Perform power flow calculation and time domain transient stability simulation calculation based on the i-th expected severe fault.

[0071] 8) Construct the input feature vector.

[0072] 9) If the expected serious fault set is calculated, execute step 10), otherwise return to 7).

[0073] 10) Generate a sample set.

[0074] 11) Construct a voltage transient stability classification model based on XGBOOST.

[0075] 12) Select the specific fault type p that the system needs to cope with.

[0076] 13) Set the reactive power output of key sensitive generators as particles, and initialize the position and speed of the particle group for reactive power optimization.

[0077] 14) Perform power flow calculation in PSASP. If the power flow converges, proceed to the next step, otherwise proceed to 18).

[0078] 15) The steady-state characteristics obtained by the power flow calculation are combined with the specific fault type p to construct a multi-dimensional input feature vector, and the multi-dimensional feature vector is input into the constructed transient voltage stability classification model to obtain the stability of the system.

[0079] 16) Determine whether the transient stability constraint is satisfied. If so, proceed to the next step. If not, proceed to 18).

[0080] 17) Calculate the objective function (network loss), compare the current fitness and the best fitness in memory, if the current fitness is better than the best fitness, replace it and record the current position as the best position of the particle. Compare the best position of the particle and the best position of the population, if the best position of the particle is better than the optimal position of the population, replace it, otherwise it remains unchanged.

[0081] 18) Update the particle's position and velocity.

[0082] 19) Return to step 14) and perform cyclic calculation until the convergence condition is met or the number of iterations reaches the maximum, and output the optimal solution for the reactive power output of the generator.

[0083] 4 Simulation Example

[0084] This application adopts Figure 2The feasibility of the proposed synchronous generator reactive power control strategy is demonstrated in the Jiangxi power grid shown in the figure. Under the small load level, the total active load of the Jiangxi power grid is 19009MW, the startup scale is 11880MW, and the DC feed-in is 2750MW. The system benchmark capacity is 100MV·A, and the 500kV system benchmark voltage is 525kV. The temporary stability simulation time is 5s. The expected serious fault set of the Jiangxi power grid is shown in Table 3. All faults are three-phase permanent phase faults. The fault occurrence time is t=1.0s and the removal time is t=1.1s. The judgment condition of voltage transient stability is that the 500KV bus voltage of the power grid can be restored to 0.9pu after the fault. The parameters of the particle swarm optimization algorithm are set to the maximum number of iterations of 40 and the number of particles of 30.

[0085] Firstly, based on the expected fault set of Jiangxi power grid (a total of 114 faults), the weak areas of Jiangxi power grid are obtained based on the voltage transient stability margin of multi-binary meters and AP clustering algorithm: Nanchang, Fuzhou, Yuzhang, Yunfeng, Chongren, Mengshan, Yongxiu, Jinxian. Then, according to the reactive power-voltage transient stability sensitivity index, the key sensitive generating units of Jiangxi power grid are Fengcheng 01, Fengcheng 02, Fuzhou Power Plant, Fengcheng Phase II, Gui Phase III, and Huangjinbu.

[0086] In PSASP (power system comprehensive analysis software), different reactive output levels are set for the key sensitive generators in Jiangxi power grid, and based on the actual faults in Jiangxi power grid in Table 3, time domain transient stability simulation calculation is performed, and a total of 1885 groups of transient voltage stability samples are obtained, including 541 stable samples and 1344 unstable samples.

[0087] Table 3 Anticipated severe fault set

[0088]

[0089] The voltage transient stability feature vector is constructed based on Table 1. The training set and the validation set are divided, where the number of training set samples is 1320 groups and the number of test set samples is 565 groups. XGBoost is used to evaluate the classification model of transient voltage stability on the training set, and the model is evaluated on the validation set. The relevant evaluation indicators are shown in Table 4:

[0090] Table 4 Model scores

[0091]

[0092] It can be seen from the evaluation indicators in Table 4 that the voltage transient stability classification model established based on XGBoost has good prediction accuracy and can be extended to practical applications as a constraint condition for reactive power optimization.

[0093] Table 5

[0094]

[0095] Under the three different faults shown in Table 5, the particle swarm algorithm is used to optimize the reactive output of the key sensitive generators in the weak areas of Jiangxi power grid. The iterative changes of the system network loss under the three faults are shown in Figure 3, and the reactive output of the key sensitive generators is shown in Table 6, in Mvar. The network loss comparison under other control strategies that can keep the system transient voltage stable is shown in Table 7. The optimized results of the reactive output of the generator are substituted back into the original system for verification, and the reactive output of the system 500kv bus represented by Nanchang 500a and the key sensitive generator represented by Fuzhou Power Plant are shown in Figures 4 and 5 respectively.

[0096] Table 6 Reactive output of key sensitive generators

[0097]

[0098] Table 7 System network loss

[0099]

[0100] As shown in FIG5, after adopting the method for optimizing the dynamic reactive power support of synchronous generators proposed in the present invention, the generator can generate more reactive power during the transient fault period, thereby providing dynamic support capability for the recovery of the system bus voltage. As shown in FIG4, before adopting the method for optimizing the dynamic reactive power support of synchronous generators proposed in the present invention, the system will experience transient voltage instability under the three faults shown in Table 5, and after adopting the method, the bus voltage of the system can be restored to stability. It is proved that the method for optimizing the dynamic reactive power support of synchronous generators proposed in the present invention can improve the transient voltage stability of the system under different faults.

[0101] From the comparison in Table 7, it can be seen that, compared with other control means that can keep the transient voltage of the system stable, the method for optimizing the dynamic reactive power support of the synchronous generator proposed in the present invention can effectively reduce the network loss of the system.

[0102] From the above results, it can be seen that after adopting the method of optimizing the dynamic reactive support of synchronous generators proposed in this paper, not only can the transient voltage stability of the system be effectively improved for different faults, but also the changes in system network losses can be taken into account, thereby reducing the economic losses of the power system during steady-state operation.

[0103] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for optimizing the steady-state reactive power of synchronous generators to improve voltage temporary stability based on machine learning, characterized in that: include: The XGBOOST machine learning method is used to establish a classifier based on the system steady-state multi-dimensional input feature vector and the expected fault type to determine whether the system transient voltage remains stable to the expected fault disturbance. Taking the system transient voltage remaining stable under a certain fault disturbance as a constraint condition, and considering minimizing the network loss caused by the steady-state reactive power optimization of the generator as the objective function, a steady-state reactive power optimization model is established to improve the dynamic reactive power support provided by the synchronous generator for the transient voltage. The particle swarm algorithm is used to call the power flow calculation to solve the optimization model and determine the optimal control of the steady-state reactive power of the generator. When determining the weak areas of system voltage transient stability and the voltage regulation sensitive generators in the area, it includes: Step 1: Use PSASP power system numerical simulation software to establish a mathematical model for power flow and transient stability calculation and analysis of the system's determined operating mode, perform transient stability time-domain simulation calculations of the expected fault set disturbance, and obtain the transient response curve of the bus voltage to the fault disturbance; Step 2: Based on the transient voltage response curve of bus i to the expected fault j disturbance, calculate the transient voltage stability margin according to formula (1): In the formula is the voltage transient stability margin of bus i under fault j disturbance described by multi-binary table, V i N and V i j (t) represent the voltage reference value of bus i and the real-time value under fault j, t k and t k ′ are V i j (t) The voltage below and above the kth critical reference value during the drop and recovery process The moment K k It means (t k ,t k+1 )∪(t′ k+1 ,t′ k ) interval integral weight coefficient, when V i j (t) are less than V n cr When the voltage drop and recovery process is located at (t1, t2)∪(t2′, t1′),…,(t k ,t k+1 )∪(t k ' +1 ,t k ′),…,(t n ,t n The integral weight coefficients of the intervals between ′ are K1,…,K k ,…,K n , the weight coefficient of each interval can be determined according to the voltage transient response to keep the voltage critical transient stability, that is, Determined by solving formula (2) step by step; In the formula Indicates that the voltage is lower than The maximum acceptable duration, k = 1, 2, ..., n; Step 3: According to the transient voltage stability margin obtained based on the multi-binary table Constructing transient stability margin index similarity matrix j∈[1,q], A is used as the input of AP clustering algorithm to cluster and partition each bus of the studied power flow mode; Step 4: Based on the column vector of A, according to Determine the severity of fault j according to ξ j The sizes are arranged from small to large from the middle to both sides, and the row vector describing the severity of each expected fault in the studied power flow mode is defined, that is, the middle element ξ of the row vector p The value is the smallest, and the corresponding fault p disturbance is the most serious; Step 5: According to formula (III), the comprehensive transient stability margin Ση of each bus voltage to different fault disturbances is calculated based on linear weighting: i , and using Ση i Sort to determine the smaller Ση i The busbar is taken as the weak busbar of transient voltage stability margin, and then the weak area of ​​transient voltage stability margin under the studied power flow mode is determined according to the partition where the weak busbar of transient voltage stability margin is located; Where q represents the number of expected failures, ω j represents the weight of the jth fault, l is composed of the row vector ξ describing the severity of each expected fault j The column position value is determined; Step 6: Using the power flow calculation and analysis model established in step 1, increase the output reactive power of the synchronous generator r by △Q r =0.1pu, based on the most serious fault p determined in step 4, transient stability calculation and analysis are performed, and the weak area bus i is recalculated according to formula (1) after the generator r changes its reactive output According to formula (IV), the voltage regulation sensitivity s of the synchronous generator r reactive regulation to the bus with weak voltage transient stability margin is calculated r , and then according to s r Sequencing determines the synchronous generator sets that have sensitive voltage regulation effects on buses with weak transient voltage stability margin; The establishment of XGBOOST classifier includes: Step 1: Based on the transient voltage response curve obtained in step 1, the severe anticipated fault set is determined based on the standard that the steady-state value of the transient voltage response after the fault disturbance cannot be restored to above 0.9pu; Step 2: Adjust the steady-state output reactive power of the determined voltage-regulated sensitive generator set and the reactive power of the sensitive generator set in the vicinity of the parallel connection and calculate the power flow, and further perform transient stability calculations for each severe anticipated fault disturbance on each convergent power flow; Step 3: Taking the steady-state value of transient voltage response restored to above 0.9pu as the standard, judge whether the convergent power flow can maintain transient voltage stability for severe anticipated fault disturbances, set the label of the convergent power flow with transient voltage stability to 1, otherwise set the label to 0, and at the same time, based on the steady-state value of the multi-dimensional input feature vector of the convergent power flow and the severe anticipated fault type, construct a machine learning sample of the classifier for predicting system voltage transient stability. On this basis, based on the voltage transient response curves of different convergent power flows to each severe anticipated fault disturbance, establish a sample data set for machine learning; Step 4: Divide the sample data set into a training set and a validation set in a ratio of 7:3; use the multi-dimensional input feature vector of the training set sample i as the XGBOOST classifier decision tree f shown in formula (V) k (x i )^ input x i , calculate the score y of the classifier output i As a result of predicting whether the system voltage of the studied power flow mode maintains transient stability to the expected fault disturbance, by increasing the number of decision trees in formula (V) until the loss function shown in formula (VI) reaches the minimum value, the iterative training of XGBOOST classifier modeling is stopped, and the classifier for predicting the transient voltage stability of the studied power flow mode corresponding to the training set is obtained. In the formula, f k (x i ) represents the kth decision tree of XGBOOST, N R is the total number of decision trees when iterative training is stopped, is the error between the model prediction and the true value, Ω(f k ) is a regular term used to avoid overfitting of the model, γ and λ are penalty coefficients of the regular term, T is used to control the number of leaf nodes in the tree model to prevent the tree model from being too complex, and W is used to avoid the leaf node weight vector from being too large; The steady-state reactive power output optimization model is based on the following objective functions and constraints: min F(Q G )=P loss ,Q G =(Q G1 ,Q G2 …Q Gz ) In the formula, Q G represents the reactive power output combination of the voltage-sensitive generator set, Gz represents the number of sensitive generator sets, P loss Represents the system network loss, G ij , B ij and θ ij Respectively represent the conductance, susceptance and phase difference between busbars i and j, U i and U j Represent the voltage amplitude of busbars i and j respectively, P Gi , Q Gi and P Li , Q Li They represent the active and reactive power input of the generator of bus i, and the active and reactive power demand of the load, respectively. Di represents the injected active power of bus i, Q Ci represents the parallel compensation reactive power of bus i, Q kGmin , Q kGmax They represent the upper and lower limits of reactive power output of voltage-sensitive generators, U imin , U imax They represent the upper and lower limits of the voltage of bus i, N is the number of system buses, h(X) is a classifier based on XGBOOST to predict whether the system voltage maintains transient stability, and X represents the multidimensional input feature vector of the studied power flow mode and the expected fault type.

2. According to claim 1, a method for optimizing the steady-state reactive power of a synchronous generator to improve voltage temporary stability based on machine learning is characterized in that: The multi-dimensional input feature vector is based on the following definition: Classification 1, fault characteristics: f0 represents the type of power grid fault; Category 2, steady-state characteristics of weak voltage stabilization areas: f1 represents the bus voltage in weak voltage stabilization areas; f2 represents the reactive power output of voltage regulation sensitive generators; f3 represents the reactive power load in weak voltage stabilization areas; f4 represents the active power output of voltage regulation sensitive generators; f5 represents the active power load in weak voltage stabilization areas; Category III, system steady-state characteristics: f6 represents the total reactive power output of the system generator; f7 represents the total active power output of the system generator; f8 represents the total reactive power demand of the system load; f9 represents the total active power demand of the system load; f 10 Indicates the average value of the system bus voltage.

3. According to the method of claim 1, the method for optimizing the steady-state reactive power of synchronous generators to improve voltage temporary stability based on machine learning is characterized in that: Classifier modeling specifically includes: Identify the weak transient stability areas of system voltage and the voltage regulation sensitive generators in the area; Based on the transient stability calculation of severe anticipated fault sets in power grid, an XGBOOST classifier is established to predict system voltage transient stability based on multi-dimensional input feature vectors and fault types.

4. According to claim 2, a method for optimizing the steady-state reactive power of a synchronous generator to improve voltage temporary stability based on machine learning is characterized in that: The steps to solve the optimization model include: Step 1: Set the determined steady-state reactive output of the voltage-sensitive generator as the optimization particle, initialize the position and speed of the particle group used to optimize the steady-state reactive output of each voltage-sensitive generator group, and the number of optimization iterations; Step 2: According to the optimized particle, the output reactive power of the voltage-sensitive generator set is adjusted, and the reactive power of the parallel compensation of the sensitive generator set in the vicinity is adjusted. If the power flow converges, based on the multi-dimensional input feature vector, the classifier is used to predict whether the system voltage maintains transient stability to the expected fault disturbance. If it is transiently stable, step 3 is executed, and if it is unstable, step 4 is executed; Step 3: Calculate the objective function fitness value of the optimization model shown in formula (VII). If the current fitness value is better than the best fitness value in memory, replace it and record the current position as the best position of the particle. At the same time, compare the best position of the particle and the best position of the population. If the best position of the particle is better than the best position of the population, replace it, otherwise it remains unchanged. Step 4: Update the position and velocity of the optimization particle and return to step 2 for cyclic calculation until the convergence condition of the optimization algorithm is met or the number of iterations reaches the maximum, and determine the optimal command value of the steady-state output reactive power of the voltage-sensitive generator set.

Images